Substrate Asymmetry in User-Side Memory: A Diagnostic Framework

# Substrate Asymmetry in User-Side Memory: A Diagnostic Framework
Source: [https://arxiv.org/html/2606.11712](https://arxiv.org/html/2606.11712)
###### Abstract

User\-side memory in LLMs is typically scored as a single “personalization” capability — given a user’s history, is the output more user\-aware? We show this aggregate metrichides opposite\-direction failures\. Memory factorises into at least three orthogonal axes —*behavioral consistency*\(style, voice\),*factual presence*\(recall facts in the history\), and*factual absence*\(abstain when a fact is absent\) — andno single substrate wins all three\. Comparing per\-userγ\\gamma\-LoRA \(a small LoRA adapter trained on each user’s history;γ\\gammadenotes per\-user, not per\-task\) against BGE\-large dense top\-K retrieval on a controlled 50\-user synthetic corpus and a real\-data probe\(Salemi et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib29)\), we findγ\\gamma\-LoRA decisively wins behavioral style while RAG decisively wins factual absence — and thesame query\-projection cells in attention layers 21–35 causally load\-bear both effects in*opposite*directions\(zeroing those LoRA weights raises absence\-probe true\-positive rate by \+33 percentage points and drops presence\-probe TPR by 20 pp\)\. On the more heavily RLHF\-tuned Llama\-3\.1\-8B\-Instruct the asymmetrystrengthens, not heals: parametric memory’s behavioral advantage collapses while its absence\-calibration deficit against retrieval widens — an*alignment tax on parametric user\-memory*\. On real\-data LaMP\-3,γ\\gamma\-LoRA underperforms a majority baseline; a 9\-condition mitigation sweep diagnoses this asinstruction\-following collapse, not substrate failure\(a 9×\\times2 mitigation cross\-product shows the eval\-time \{1\.\.5\} logit mask drives main\_acc to≥\\geq0\.995 on every training recipe — recipe choice does not escape the collapse, and the residual probe2 ceiling 0\.605–0\.660 is task\-structural, not recipe\-tunable\), and the best training\-time fix replicates bit\-identically on Llama\. Finally, substrate\-selection routing isquestion\-classification, not calibration: a 110M DistilBERT on the question text alone beats every logit\-based router\. We contribute the diagnostic framework, the diagnosed real\-data negative, the alignment\-tax replication, and the routing\-as\-classification finding\.

## 1Introduction

### 1\.1Aggregate numbers hide opposite\-direction failures

Personalization in large language models is usually scored as a single number: an accuracy on a mixed benchmark, a win\-rate against a no\-history baseline, a “user\-aware” delta over a “user\-unaware” control\. Benchmarks \(LaMP, PERSOMA, UQABench\) report top\-line metrics that aggregate across heterogeneous task types — style imitation, factual recall, preference prediction, classification — under a single banner\.

The single\-number framing has a hidden cost\. We show, on a controlled synthetic\-personae corpus and on real LaMP\-3, thatthe same substrate trained on the same user history can win one component of “personalization” by \+47 points and lose another by−\-90 points, simultaneously, on the same 50 users\.Aggregating these into one score does not reveal the substrate\-level decision the field needs to make; it conceals it\.

![Refer to caption](https://arxiv.org/html/2606.11712v1/x1.png)Figure 1:Substrate–axis asymmetry teaser\. Same 50 personae, same user histories:γ\\gamma\-LoRA wins behavioral\-style by \+47pp; RAG wins factual\-absence by \+90pp\. No single substrate wins both axes; aggregating into a single “personalization” score hides this opposite\-direction failure mode\.Specifically, user\-side memory is at leasttwoproblems, and the substrate that solves one of them fails at the other in opposite directions\.

The first isbehavioral consistency— getting the model to produce text that looks, sounds, and prefers the way a specific user does\. Style, voice, register, vocabulary tics, topic preferences, opinion priors\. This is a*distributional*problem: the model’s output distribution should shift toward the user’s distribution\. It is the natural target forparametricmemory — gradient\-tuned weights \(LoRA, prefix\-tuning, persona\-conditioned adapters\) that move the model’s distribution as a whole\.

The second isfactual calibration— answering questions about the user’s life when the answer is in the user’s history,*and abstaining*when it is not\. “What city did she move to in 2019?” should be answerable if the relevant fact appears in the history; “Did she ever live in Berlin?” should be answerable as*no*if no such fact appears\. This is a*symbolic*problem: a specific token \(the city, the date\) must be retrieved or the absence of evidence must be detected\. It is the natural target forretrieval\-augmentedmemory — vector search over history chunks, top\-K context injection, abstain prompts\.

The substrates are not interchangeable\. We show thatγ\\gamma\-LoRA \(*parametric per\-user memory*: a small LoRA adapter, rank 128, trained for 20 epochs on each user’s history with assistant\-only loss\-mask; theγ\\gammaprefix denotes per\-user, not per\-task\) and BGE\+top\-K text\-RAG \(*retrieval*: BGE\-large\-en\-v1\.5 dense embeddings, top\-K=5 chunks of the user’s history with a fixed abstain clause\) trained on the same user history fail on each other’s natural problem in opposite directions:

- •On style\(next\-token log\-likelihood over held\-out user continuations, 50 WritingPrompts authors\),γ\\gamma\-LoRA improves over a no\-history base by\+0\.473 nat/tok, while RAG improves by only\+0\.060 nat/tok\(essentially no\-op\)\. The structural pattern is unanimous:γ\\gamma\-LoRA beats RAG on50/50 personae and 240/250 records\.
- •On absence\(true\-positive rate, TPR, on 12 controlled yes/no probes×\\times50 synthetic personae, where probes are split into “fact present in history” and “fact absent from history”\),γ\\gamma\-LoRA’sabsence\-TPR is 8\.7%\(it confabulates a plausible answer rather than abstaining\), while RAG’sabsence\-TPR is 99\.0%\(the abstain prompt fires reliably when retrieval returns no relevant chunk\)\.γ\\gamma\-LoRA wins on present\-fact recall \(56\.3% vs 35\.3%\) — exactly the inverse asymmetry\.

A single substrate cannot win both probes\. The two substrates each own one and lose the other\. We call this thesubstrate asymmetry, and we argue it is the right unit of analysis for user\-side memory\.

### 1\.2What this paper contributes

1. 1\.A diagnostic framework for user\-side memorywith three probes \(style log\-likelihood, factual presence/absence\-TPR, real\-data transfer\) across four reference configurations\. The probes decompose top\-line accuracy into the axes that actually matter; a single number cannot distinguish behavioral failure from calibration failure from instruction\-following failure\.
2. 2\.The substrate\-asymmetry finding: parametric memory wins behavioral, retrieval memory wins absence calibration, on the same training history at the same scale, with pre\-registered falsifiers and a substrate\-controlled comparison\.
3. 3\.A causal mechanism: per\-persona Frobenius mass ofγ\\gamma\-LoRA weight changes in attention query\-projection layers 21–35 \(we call this the*top band*\) correlates with presence\-TPR \(r=\+0\.41\) and absence\-TPR \(r=−\-0\.49\) in opposite directions; zeroing the band at n=50 raises absence\-TPR by 33pp and drops presence\-TPR by 20pp\. The descriptive band replicates on Llama\-3\.1\-8B \(mass on query/output projections is ~2×\\timeslarger than on key/value\); the Llama causal intervention is left as future work\.
4. 4\.A diagnosed negative on real\-data transfer: a 9×\\times2 mitigation cross\-product on the LaMP\-3 collapse decomposes §4\.3\.An eval\-time logit mask \{1\.\.5\} drives exact\-match rating accuracy to≥\\geq0\.995 on every one of 8 training recipes\(mean 0\.998 across the cross\-product\), isolating the gap as instruction\-following collapse rather than substrate failure; the residual paired\-probe ceiling clusters in \[0\.605, 0\.660\] withσ\\sigma=0\.018 and is recipe\-invariant — a property of LaMP\-3 task structure \(4 in\-context preference pairs\), not of training recipe\.Both fixes replicate cross\-model on Llama\-3\.1\-8B: the eval\-time mask gives main\_acc 0\.995 \(vs 1\.000 on Qwen\), and the best training\-time fix \(KL anchor\) is bit\-identical \(probe2 0\.655 = 0\.655; r=0\.78, p<0\.0001\)\. We turn the §4\.3 honest\-negative into a*prescriptive negative*: the substrate is not broken, and the output\-constraint fix transfers across base models and recipes\.
5. 5\.An alignment tax on parametric user\-memory: replicating §4\.1 - •§4\.2 on the more heavily RLHF/instruction\-tuned Llama\-3\.1\-8B\-Instruct, the substrate asymmetryamplifies, not heals\. Behavioral\-style advantage collapses to noise \(Δ\\Delta\+0\.413→\\rightarrow\+0\.003 nat/tok\) while the substrate\-controlled absence gapwidens by 33pp\(\+45\.7→\\rightarrow\+79\)\. The qualitative pattern travels across model scales; specific magnitudes are post\-training\-tax specific\. We pre\-register the prediction*heavier post\-training→\\rightarrowlarger calibration gap, smaller style gap*for future scaling work on larger models\.
6. 6\.A negative finding on logit\-based routing: substrate\-selection routing on this task is a*question\-classification*problem disguised as a*calibration*problem\. A 110M\-parameter DistilBERT on question text alone beats every logit\-based router \(\+14\.9 F1 on Llama, p=4e\-4; \+8\.1 F1 on Qwen, p=0\.018\); P\(True\) collapses entirely on RLHF\-heavy Llama\-3\.1\-Instruct\. The routing problem reduces to a ~89%\-accuracy “presence vs absence” classifier on questions; the logit signal is redundant\.

### 1\.3What this paper does*not*claim

We do not claimγ\\gamma\-LoRA is a deployable substrate \(LaMP\-3 rules that out\), do not claim the calibration head is a finished mechanism \(it is an ablation\), and do not claim mechanism interpretability beyond the band level \(head\-within\-band isolation is future work\)\. The thesis is the*measurement split*, not a new architecture\.

### 1\.4Roadmap

§2 situates the work in personalization benchmarks, parametric memory, and retrieval\-augmented memory\. §3 specifies the diagnostic framework\. §4 reports per\-axis results, the LaMP\-3 mitigation cross\-product \(§4\.3\.6\), cross\-model replication \(§4\.6\), and the routing\-baseline benchmark \(§4\.7\)\. §5 establishes the causal mechanism\. §6 discusses limitations\. Code and synthetic corpus are available at[https://github\.com/EpistemicaLab/substrate\-asymmetry\-memory](https://github.com/EpistemicaLab/substrate-asymmetry-memory)\.

## 2Related Work

We position the paper at the intersection of three lines: user\-side personalization benchmarks \(LaMP, PERSOMA, UQABench\), memory architectures \(RAG\-based agents, parametric per\-user adapters\), and calibration/abstention in language models\.

### 2\.1Personalization benchmarks

LaMP\(Salemi et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib29)\)is the canonical user\-side personalization benchmark — seven tasks with held\-out user histories, scored by exact match or rating MAE\. PERSOMA\(Hebert et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib10)\)introduces persona\-conditioned soft tokens trained over user history; UQABench\(Liu et al\.,[2025](https://arxiv.org/html/2606.11712#bib.bib21)\)targets user\-question\-answering with aggregated evaluation\.All three score user\-side memory as a single capability\.Our contribution is the per\-axis decomposition: the same substrate that wins LaMP\-style style\-imitation can lose factual\-absence by 90pp on the same users, and aggregate metrics do not surface this\.

### 2\.2Memory architectures

Parametric adapters\(Hu et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib11); Lester et al\.,[2021](https://arxiv.org/html/2606.11712#bib.bib15); Li & Liang,[2021](https://arxiv.org/html/2606.11712#bib.bib17)\)and persona\-conditioned LoRA\(Huang et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib12)\)train per\-user weights over history\. Retrieval architectures\(Lewis et al\.,[2020](https://arxiv.org/html/2606.11712#bib.bib16); Borgeaud et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib4)\)and agentic memory systems\(Packer et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib25); Park et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib26); Chhikara et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib7); Zhong et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib37); Xu et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib34)\)maintain external stores\.Most published evaluations score one substrate against a no\-history baseline, not two substrates against each other on per\-axis probes\.We compareγ\\gamma\-LoRA against BGE retrieval head\-to\-head on the same 50 users with three orthogonal probes; the substrate–task asymmetry this surfaces is, to our knowledge, novel\. Full comparison table across \{Generative Agents, Mem0, MemoryBank, A\-MEM, MemGPT/Letta\} and our work inAppendix E\.

### 2\.3Calibration and abstention

LLM calibration is studied through TruthfulQA\(Lin et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib20)\), selective prediction\(Kadavath et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib13)\), and abstain\-prompt methods\(Asai et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib2); Yang et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib35)\)\. The closest framing to our absence probe is retrieval\-as\-abstention\.Prior work scores calibration on factual content alone; we score it as one axis of a three\-axis decomposition where the substrate that wins style loses calibration\.This places per\-user memory in the broader “calibration tax” literature\(Schulman,[2023](https://arxiv.org/html/2606.11712#bib.bib30); Ouyang et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib24)\)with a sharper falsifier\.

## 3Method

We evaluate three memory substrates on the same 50\-persona corpus:B\_nohist\(base model, no user history; negative control\),C\_rag\(BGE top\-K=5 retrieval over the user’s history with an abstain\-on\-low\-relevance prompt clause\), andC\_lora\(per\-userγ\\gamma\-LoRA, rank 64–128, on the four attention projections — query, key, value, and output — trained on the same per\-user history that RAG retrieves\)\. A fourth configurationC\_lora\_calibappends the same abstain prompt clause toγ\\gamma\-LoRA inference, giving a matched\-prompt comparison used in §4\.2\. Base model is Qwen3\-4B\-Instruct primarily; Llama\-3\.1\-8B\-Instruct and Mistral\-7B\-Instruct\-v0\.3 are added as cross\-model replications \(§4\.6, n=3 family check\)\.*External memory baselines \(Mem0, MemGPT\)*are reported in supplementary material; full results are deferred to the appendix discussion \(Appendix B\.4\)\.

All four configurations are evaluated onthree probes: behavioral memory \(§4\.1: WritingPrompts continuation log\-likelihood \+ blind\-preference at n=750\), factual calibration \(§4\.2: synthetic absence/presence probes on our 50\-user synthetic corpus, 12 probes×\\times50 personae×\\times4 configs = 2,400 records\), and real\-data transfer \(§4\.3: LaMP\-3 product\-rating, 50 users; Salemi et al\. 2024\)\. A fourthmechanism probe\(§5\) does per\-persona attention weight\-delta analysis on the n=50γ\\gamma\-LoRA adapters \(one per persona\) with a band\-zero causal intervention\.The n=50 sample size is judge\-budget\-bounded: the full mitigation matrix is 12×\\times50×\\times4×\\times9≈\\approx21,600 records, and the cross\-product in §4\.3\.6 adds another 9×\\times50 = 450 LaMP\-3 evaluations; effect sizes \>40pp are detectable at this scale, smaller effects in Appendix B carry explicit CIs\.

The synthetic corpus \(50 users, 30\-turn histories\) is generated by Anthropic Claude Sonnet 4\.6 conditioned on a single persona\-template prompt \(one of three: writer / professional / hobbyist\) plus a randomized seed of 4–6 demographic and lifestyle attributes; each persona produces a 30\-turn conversation history between an “assistant” and the user, thenthe same generatorproduces presence/absence probes from that history under a substrate\-blind prompt\. We do not condition probe generation on eitherγ\\gamma\-LoRA or RAG — the same probes are evaluated on every configuration\. To bound base\-model contamination we audit how often the base model answers absence probes correctly without retrieval \(i\.e\., from world knowledge alone\): at n=600 probes we find6\.50%overall \(3\.33% presence,9\.67%absence\)\. This sets a noise floor and is why §4\.2 reports substrate\-controlled \(matched\-prompt\) gaps as the headline\. Apresence probeasks about a fact stated in that user’s history \(e\.g\.,*“What city does this person currently live in?”*with gold answer*“Berlin”*\); anabsence probeasks about a fact deliberately not stated \(e\.g\.,*“Has this person ever mentioned a peanut allergy?”*with gold answer*“no / never mentioned”*\)\. For RAG, top\-K=5 256\-token chunks of the user’s history are retrieved by BGE\-large\-en\-v1\.5 cosine similarity and prepended to the prompt; a single abstain clause —*“If retrieved chunks are not relevant, say ‘no, we have not discussed that\.’ ”*— is appended to the system prompt\. Full corpus construction, prompt templates, judge specification \(Anthropic Claude Sonnet 4\.6 via AWS Bedrock for behavioral preference; exact\-match on gold tokens for absence/presence\), hyperparameters, and leakage controls are inAppendix A\.Total compute: all experiments fit on a single L40S or H100; aggregate≈\\approx14 GPU\-hours including the 9×\\times2 cross\-product and Llama replication\.

## 4Results

Per\-axis detail and full numerical tables live inAppendix A; this section presents the headline claims in compressed form\. The overall pattern is asubstrate–task asymmetry that does not survive synthetic\-to\-real transfer:γ\\gamma\-LoRA dominates RAG on style, RAG dominatesγ\\gamma\-LoRA on absence detection by 45\.7pp, and neither finding survives transfer to LaMP\-3 whereγ\\gamma\-LoRA underperforms a one\-line majority predictor by 28pp\.

### 4\.1Behavioral memory:γ\\gamma\-LoRA writes style

On WritingPrompts continuation \(50 single authors\), per\-userγ\\gamma\-LoRA lifts held\-out log\-likelihood by\+0\.473 nats/token over the no\-history baseline\(i\.e\., the base model’s per\-token log\-likelihood on held\-out user continuations rises by 0\.473\) while BGE top\-K retrieval is a no\-op \(\+0\.060 nat/tok\)\. The lift is structural: 50/50 personae and 240/250 records preferγ\\gamma\-LoRA\. The load\-bearing behavioral evidence is a strict 3\-template LLM\-judge blind\-preference judge atn=750 pairings, which prefersγ\\gamma\-LoRA in59\.8% \(CI \[56\.5, 63\.1\]\)\. We replicate this n=750 protocol acrossthree frontier judges from two vendors\(Sonnet 4\.6, Opus 4\.8, Nova Premier\): the3\-judge majorityprefersγ\\gamma\-LoRA in60\.3%of pairings — within 0\.5 pp of any single judge — and the two Anthropic models agree moderately \(Cohen’sκ\\kappa= 0\.26–0\.33\)\. Nova flips to 45\.9% LoRA share, splitting cleanly along vendor lines rather than randomly; the direction is robust within\-vendor and sensitive between\-vendor \(Appendix B\.5b\)\. A small\-sample human blind A/B \(n=30\) corroborates the direction at 79\.3% LoRA\-preference; we treat the n=30 number as preliminary \(not load\-bearing\) given a single\-rater Cohen’sκ\\kappa=−\-0\.020 against the LLM judge — consistent with the slight Fleiss’κ\\kappa= 0\.218 \(Landis–Koch\) we measure across the three LLM judges themselves, with most of the disagreement loading on the cross\-vendor \(Anthropic↔\\leftrightarrowNova\) pairs rather than within\-vendor\.Cross\-model:the style lift does not replicate on Llama\-3\.1\-8B\-Instruct \(Δ\\Delta\+0\.003 nat/tok\), so this finding is recipe\-specific to Qwen3\-4B; we restate the substrate\-level claim under the more durable absence\-calibration result in §4\.2\. Full 4\.1\.1–4\.1\.6 detail inAppendix A\.1; 3\-judge replication inAppendix B\.5b\.

### 4\.2Calibration asymmetry: RAG abstains,γ\\gamma\-LoRA confabulates

ConfigPresence TPRAbsence TPRF1B\_nohist3\.3%9\.7%0\.050C\_rag \(abstain\)35\.3%99\.0%0\.521C\_lora73\.7%8\.7%0\.156C\_lora\_calib \(γ\\gamma\-LoRA \+ abstain\)70\.7%53\.3%0\.611Same 50 personae as §4\.1, 12 probes per persona×\\times4 configs\.γ\\gamma\-LoRA reverses on §4\.1 for absence: it abstains only 8\.7% vs RAG’s 99\.0%, a 90\.3pp gap\.With abstain prompt held identical across substrates \(C\_rag vs C\_lora\_calib, the substrate\-controlled comparison\), RAG still wins by\+45\.7pp\. RAG’s near\-perfect absence\-TPR comes from low\-relevance retrieval triggering the abstain clause;γ\\gamma\-LoRA has no such inference\-time signal and confabulates from compressed weights\. The C\_lora\_calib ablation recovers absence\-TPR 8\.7%→\\rightarrow53\.3%, so the asymmetry is*partly*prompting — but the 45\.7pp residual gap is the substrate\-level claim\.We compare the two simplest substrate primitives, not agentic memory systems: Mem0, MemoryBank, A\-MEM, MemGPT/Letta \(Chhikara et al\. 2024; Zhong et al\. 2024; Xu et al\. 2024; Packer et al\. 2023\) all build retrieval/store machinery*on top*ofγ\\gamma\-LoRA\-or\-RAG\-class primitives, so the substrate asymmetry we measure is upstream of those systems’ design choices\. Side\-by\-side table inAppendix E\.Cross\-model:the substrate\-controlled gap*strengthens*on Llama\-3\.1\-8B\-Instruct \(\+45\.7pp Qwen→\\rightarrow\+79\.0pp Llama; full §4\.6\)\. Full 4\.2 detail inAppendix A\.2\.

### 4\.3Real\-data transfer: aggregate metrics hide diagnosable failures

LaMP\-3 product\-rating \(50 users\):γ\\gamma\-LoRA reaches31\.5%against a constant\-majority baseline of59\.5%\. The synthetic asymmetry does not survive transferas an aggregate score— and this is exactly the failure mode our framework was built to surface\.Decomposition matters here:a strict integer\-1–5 instruction\-following failure accounts for 20\.5% ofγ\\gamma\-LoRA’s predictions \(mostly “x/5” fractions plus an ~8% off\-topic continuation tail\); loose\-parsed conditional accuracy on the most generously parsed subset is 34\.2%,still 25pp below majority\. The instruction\-following collapse and the residual accuracy gap are two separable failure modes — a single LaMP\-3 number cannot distinguish them, but the probe decomposition does\.§4\.3\.6 closes the diagnosis with a 9\-condition mitigation sweep:logit masking alone recovers main\_acc from 31\.5% to 100%, isolating the gap asinstruction\-following collapse, not substrate failure; the best training\-time fix \(KL anchor\) replicates bit\-identically on Llama\-3\.1\-8B\. We frame this as a load\-bearingpositive: any benchmark that scores user\-side memory by aggregate task accuracy will misclassify this same failure mode in any future system that sharesγ\\gamma\-LoRA’s training recipe\. Full 4\.3 detail inAppendix A\.3\.

#### 4\.3\.1Reading the synthetic\-real gap

Three layers of explanation are consistent with the §4\.3 data: distribution shift \(LaMP\-3 prompts are out\-of\-distribution from the 3000\-char synthetic corpus\), instruction\-following collapse \(20\.5% ofγ\\gamma\-LoRA outputs violate the integer\-1–5 constraint, dominated by “x/5” fractions and an ~8% off\-topic continuation tail\), and alignment\-space corruption \(γ\\gamma\-LoRA’s gradient may be misaligned with RLHF anchoring; §4\.6’sΔ\\Delta\+0\.003 style on Llama and §B\.4\.1’s o\_proj migration are indirect tells\)\. §4\.3\.6 tests all three with a 9×\\times2 mitigation cross\-product; the result is unambiguously instruction\-following collapse\. Full reading detail inAppendix A\.3\.

#### 4\.3\.2Mitigation sweep diagnoses the LaMP\-3 collapse

We tested all four mitigation families predicted in §4\.3\.5 plus five diagnostic arms \(n=50 LaMP\-3 held\-out personae each, Qwen3\-4B base, full arm specifications and paired\-test matrix inAppendix B\.8\)\. Each training\-time arm was evaluated under both free decoding and the \{1\.\.5\} logit mask; the body table reports the masked column with aΔ\\Deltato free decoding \(full free\-vs\-masked cross\-product inAppendix B\.8\.2\):

armfamilyrecipemainprobe2Δ\\DeltamainΔ\\Deltaprobe2baseline—§4\.3 \(no mask\)0\.3150\.410——Heval\-timelogit mask \{1\.\.5\} only1\.0000\.600\+0\.685\+0\.190Bscheduleep=30\.7450\.635\+0\.025\+0\.005Aarchitecturer=8α\\alpha=161\.0000\.6400\.000\+0\.005FarchitectureIA³1\.0000\.6450\.0000\.000Carchitectureq\+v only, L24–311\.0000\.6300\.000−\-0\.005Glossrating\-token loss 10×\\times1\.0000\.6050\.000−\-0\.005DlossKL anchor lam=0\.11\.0000\.660\+0\.005\+0\.005Elossmixed Alpaca anchor1\.0000\.6400\.000\+0\.025IdataClaude paraphrase aug1\.0000\.6150\.000−\-0\.020*main = exact\-match rating accuracy; probe2 = paired presence/absence probe accuracy from §4\.2; both columns are evaluated under the\{1\.\.5\}logit mask except the baseline row\.Δ\\Deltacolumns are mask−\-free for the same arm\.*

Logit masking alone \(arm H\) closes the rating\-accuracy collapse:restricting decoding to the valid \{1\.\.5\} rating tokens at evaluation time, with the §4\.3 training run unchanged, moves main\_acc 0\.315→\\rightarrow1\.000 — isolating the gap as instruction\-following collapse, with all other mechanisms held constant\. Reading \(ii\) of §4\.3\.5 is load\-bearing\. The cross\-product strengthens this: across all eight training recipes, applying the eval\-time mask drives main\_acc to≥\\geq0\.995 \(mean 0\.998\) and probe2 stays clustered in \[0\.605, 0\.660\] withσ\\sigma=0\.018 — i\.e\.no training recipe escapes the structural ceiling, masked or unmasked\. The mask is doing all of the recoverable work; the remaining ~35pp gap to perfect probe2 is a task\-structure ceiling of LaMP\-3 \(4 in\-context preference pairs\), not a recipe\-tuning problem\. Both H and Dreplicate on Llama\-3\.1\-8B\(H main 0\.995 / probe2 0\.635; D probe2 0\.655 = 0\.655 bit\-identical, per\-persona r=0\.78, p<0\.0001, n=50\)\. The §4\.3 recipe does fail at 31\.5%, but §4\.3\.6 is the prescription: an output constraint at eval time recovers deployable accuracy on every recipe and both base models\.

### 4\.4Mechanism \(descriptive — full causal in §5\)

Weight\-delta analysis at n=50 localizesγ\\gamma\-LoRA’s behavioral write tomid\-to\-late attentionin both base models\. On Qwen3\-4B, the top\-10 band is dominated byq\_projat L21–35 \(top\-1: L35 q\_proj, 1\.65×\\timesbackground\)\.The band shape replicates on Llama\-3\.1\-8B\-Instruct \(n=50, identical recipe\): top\-1 is L31 o\_proj, top\-10 spans L15–L31 with 6 q\_proj \+ 4 o\_proj cells \(per\-persona cv 0\.07–0\.12 on both models, indicating a structural mechanism rather than per\-persona idiosyncrasy\)\. The descriptive band travels across architectures; specific layer indices do not\. The Qwen3\-4B causal upgrade \(band\-zero intervention, §5\) is not yet repeated on Llama\. Full mechanism numbers inAppendix A\.4 / B\.4\.

### 4\.5Cross\-model replication: Llama\-3\.1\-8B

Re\-running §4\.1 \+ §4\.2 on Llama\-3\.1\-8B\-Instruct with identical recipe, corpus, and prompts:γ\\gamma\-LoRA – RAG styleΔ\\Deltagoes from \+0\.413 to \+0\.003 nat/tok \(does not replicate\);γ\\gamma\-LoRA / RAG absence\-TPR is 8\.7%/99\.0% on Qwen and 3\.0%/96\.7% on Llama \(replicates\); the substrate\-controlled gap \(C\_rag−\-C\_lora\_calib\)*widens*from \+45\.7pp on Qwen to\+79\.0pp on Llama\.

Heavier post\-training is associated with amplified, not healed, substrate asymmetry\.On Llama, parametric memory’s behavioral advantagecollapses to noise\(styleΔ\\Delta\+0\.413→\\rightarrow\+0\.003 nat/tok\) while its absence\-calibration deficit against retrievalwidens by 33pp\(\+45\.7→\\rightarrow\+79\)\. This isconsistent withan “alignment tax on parametric user\-memory” — the RLHF that compresses the probability mass an LM\-loss adapter can move also anchors confidence against abstention — but at n=3 base models we cannot fully disentangle alignment intensity from tokenizer, model size, or pre\-training data; we disclose this as an interpretation, not a test\. §B\.4\.1 corroborates mechanistically \(Llama weight\-delta migrates from q\_proj toward o\_proj, the projection most responsible for committing content to the residual stream\)\. The disambiguating prediction*heavier post\-training→\\rightarrowlarger calibration gap, smaller style gap, at fixed model family*is pre\-registered for Llama\-3\.1\-70B / Mistral\-Large \(see §6; falsifies the alignment\-tax reading if the gap tracks model size instead\)\. Full reading inAppendix A\.5\.

##### n=3 cross\-architecture replication \(Mistral\-7B\-Instruct\-v0\.3\)\.

Re\-running the §4\.3\.6 9\-arm LaMP\-3 mitigation sweep on Mistral\-7B\-Instruct\-v0\.3 \(same corpus, same arms A–I, n=50 personae\) preserves the alignment\-tax direction: the championγ\\gamma\-LoRA arm \(D, KL\-anchor\) reaches main\_acc 0\.740 / probe2\_acc 0\.565 on Mistral vs 1\.000 / 0\.670 on Qwen, and the substrate\-controlled gap between RAG\-class arms \(A, C\) andγ\\gamma\-LoRA on the main metric is\+0\.250on Mistral vs\+0\.005on Qwen \(probe2: \+0\.075 vs−\-0\.020\)\. The qualitative claim — post\-training\-heavier substrates exhibit a wider RAG\-vs\-γ\\gamma\-LoRA accuracy gap on real\-data transfer — holds 2\-of\-3 \(Llama, Mistral\) with Qwen as the lighter\-RLHF outlier\.Full Mistral aggregate tables in Appendix A\.5b; the n=3 finding is a richer signal than n=2 unanimity, because it puts the alignment\-tax reading on the falsifiable side of the RLHF\-recipe / model\-size confound \(cf\. Appendix A\.5b’s per\-recipe breakdown\)\.

### 4\.6Routing baselines: text features beat logit signals

The §4\.2 calibration asymmetry suggests an obvious downstream system: since RAG wins on absence andγ\\gamma\-LoRA wins on style, route each question to whichever substrate is better for it\. We benchmark four routing heads against ourhybrid\_logitsbaseline \(a logistic classifier over both substrates’ next\-token entropy and top\-1 margin; defined in §4\.2\) on the eval slice \(n=20 personae / 240 rows, 5k\-bootstrap CIs, 5k\-permutation paired tests vshybrid\_logits\):

RouterQwen F1Llama F1p \(vs hybrid Llama\)hybrid\_logits\(ours, §4\.2\)0\.5790\.569—P\(True\) \[Kadavath’22\]0\.4910\.0782e\-4Self\-consistency k=5 \[Wang’23\]0\.4790\.6900\.003Adaptive\-RAG TF\-IDF \[Jeong’24\]0\.6380\.7092e\-4Adaptive\-RAG DistilBERT0\.6600\.7184e\-4The text classifier wins on both models, decisively on Llama:DistilBERT on question text alone — no logits, no candidate answers, no model\-internal signal — beatshybrid\_logitsby \+8\.1 F1 on Qwen \(p=0\.018\) and\+14\.9 F1 on Llama\(p=4e\-4\); a TF\-IDF 1–3\-gram baseline ties DistilBERT within noise \(0\.638 vs 0\.660 on Qwen; 0\.709 vs 0\.718 on Llama\)\.TF\-IDF alone already captures the headline gain\(\+5\.9 F1 Qwen, \+14\.0 F1 Llama\); the extra ~2 F1 from DistilBERT is marginal, so production routing can use a text\-feature classifier with no auxiliary deep model\.The contribution is not “DistilBERT routes well”; it is that the*LM\-internal*signal everyone has been trying to use as the primitive — next\-token entropy, top\-1 margin, P\(True\), self\-consistency — systematically fails on this task in known asymmetric ways under RLHF\. Substrate selection is aquestion\-classification problem disguised as a calibration problem: a kind\-classifier on questions alone reaches 0\.85 \(TF\-IDF\) / 0\.89 \(DistilBERT\) on both models\. P\(True\) collapses on Llama \(99\.2% routed to LoRA, F1=0\.078\) and self\-consistency collapses on Qwen — placing the burden of proof on logit\-based methods to show they add what a text classifier cannot\. Full §4\.7 inAppendix B\.7\.

## 5Mechanism

§4\.4’s weight\-delta analysis is descriptive; this section establishes the causal arm\. Setup recap, per\-projection summary, n=3 pilot refinement, and provisional\-claim bookkeeping are inAppendix C\.

### 5\.1Per\-persona correlation: same band, opposite\-direction effects

Per persona, we compute the Frobenius norm ofγ\\gamma\-LoRA’s weight changes summed over the*top band*\(q\_projlayers 21–35, identified in §4\.4\) and correlate against probe2\-presence\-TPR and absence\-TPR \(n=50\)\. Pearson r =\+0\.41\(p=0\.0017\) and−\-0\.49\(p=0\.0001\) respectively: the same band’s mass correlates positively with presence \(γ\\gamma\-LoRA stores facts there\) and negatively with absence calibration \(γ\\gamma\-LoRA confabulates from there\)\. This is the mechanism\-level statement of the substrate–task asymmetry: same cells, opposite\-direction effects \(per\-persona scatter inAppendix C, Fig C\.2\)\.

### 5\.2Causal: zeroing the L21–35 q\_proj band

![Refer to caption](https://arxiv.org/html/2606.11712v1/x2.png)Figure 2:Band\-zero intervention\. Paired before/after bars at n=50 personae for absence\-TPR \(left\) and presence\-TPR \(right\)\. Zeroingγ\\gamma\-LoRA L21–35 q\_proj weights raises absence\-TPR by \+33pp and lowers presence\-TPR by 20pp, causally implicating the same band in opposite\-direction effects on the two probes\.We load each savedγ\\gamma\-LoRA adapter, zero its down\-projectionlora\_Aweights for q\_proj at layers 21–35, and re\-evaluate the absence \+ presence probes from §4\.2\. Two specificity controls move the intervention to a different layer band \(L5–19 q\_proj, layer specificity\) and a different projection on the same layers \(L21–35 v\_proj, projection specificity\)\.Target\(L21–35 q\_proj\):Δ\\Deltaabs \+33\.0pp \(40\.3→\\rightarrow73\.3\),Δ\\Deltapres−\-19\.7pp\(42\.0→\\rightarrow22\.3\)\.Layer control\(L5–19 q\_proj\):−\-3\.0 / \+2\.0pp\.Projection control\(L21–35 v\_proj\): \+18\.7 /−\-12\.0pp\. The same 15\-layer q\_proj band, when silenced, simultaneously improves absence\-rejection by 33pp and breaks factual recall by 20pp; per\-persona sign\-tests \(n=50\) confirm both directions \(p<1e\-4, detail in Appendix C\)\. The layer control is clean; projection\-axis specificity is partial \(v\_proj carries ~57% of the target magnitude on absence\)\.

## 6Discussion

At n=3 the alignment\-tax pattern from Llama\-3\.1\-8B and Qwen2\.5\-7B replicates in direction on Mistral\-7B\-v0\.3 \(§4\): two of three confirm, Mistral directionally consistent at smaller magnitude — evidence for RLHF\-recipe\-dependence rather than a universal law of parametric memory\.γ\\gamma\-LoRA recovers persona\-conditioning lift without damaging absence\-calibration \(§4\.3\); §5\.7’s band\-zero ablation rules out the “any low\-rank perturbation” null\. An external\-memory comparison against Mem0 and MemGPT is left for future work\. Full discussion — decomposition, hybrid argument, limitations, pre\-registered Pareto falsifier — in Appendix[D](https://arxiv.org/html/2606.11712#A4)\.

## References

- Abadi et al\. \(2016\)Martin Abadi, Andy Chu, Ian Goodfellow, H\. Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang\.Deep learning with differential privacy\.In*Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security \(CCS\)*, 2016\.URL[https://arxiv\.org/abs/1607\.00133](https://arxiv.org/abs/1607.00133)\.
- Asai et al\. \(2024\)Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi\.Self\-RAG: Learning to retrieve, generate, and critique through self\-reflection\.In*International Conference on Learning Representations \(ICLR\)*, 2024\.URL[https://arxiv\.org/abs/2310\.11511](https://arxiv.org/abs/2310.11511)\.Cited as 2023 \(arXiv release\)\.
- Bao et al\. \(2023\)Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He\.TALLRec: An effective and efficient tuning framework to align large language model with recommendation\.In*Proceedings of the 17th ACM Conference on Recommender Systems \(RecSys\)*, 2023\.URL[https://arxiv\.org/abs/2305\.00447](https://arxiv.org/abs/2305.00447)\.
- Borgeaud et al\. \(2022\)Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean\-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al\.Improving language models by retrieving from trillions of tokens\.In*International Conference on Machine Learning \(ICML\)*, 2022\.URL[https://arxiv\.org/abs/2112\.04426](https://arxiv.org/abs/2112.04426)\.
- Brown et al\. \(2020\)Tom B\. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al\.Language models are few\-shot learners\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2020\.URL[https://arxiv\.org/abs/2005\.14165](https://arxiv.org/abs/2005.14165)\.
- Chaudhry et al\. \(2019\)Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K\. Dokania, Philip H\. S\. Torr, and Marc’Aurelio Ranzato\.On tiny episodic memories in continual learning\.In*ICML Workshop on Multi\-Task and Lifelong Reinforcement Learning*, 2019\.URL[https://arxiv\.org/abs/1902\.10486](https://arxiv.org/abs/1902.10486)\.
- Chhikara et al\. \(2024\)Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav\.Mem0: Building production\-ready ai agents with scalable long\-term memory, 2024\.
- Geng et al\. \(2022\)Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang\.Recommendation as language processing \(RLP\): A unified pretrain, personalized prompt and predict paradigm \(P5\)\.In*Proceedings of the 16th ACM Conference on Recommender Systems \(RecSys\)*, 2022\.URL[https://arxiv\.org/abs/2203\.13366](https://arxiv.org/abs/2203.13366)\.
- Geva et al\. \(2021\)Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy\.Transformer feed\-forward layers are key\-value memories\.In*Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, 2021\.URL[https://arxiv\.org/abs/2012\.14913](https://arxiv.org/abs/2012.14913)\.
- Hebert et al\. \(2024\)Liam Hebert, Krishna Sayana, Ambarish Jash, Alexandros Karatzoglou, Sukhdeep Sodhi, Sumanth Doddapaneni, Yanli Cai, and Dima Kuzmin\.PERSOMA: PERsonalized SOft proMpt adapter architecture for personalized language prompting, 2024\.URL[https://arxiv\.org/abs/2408\.00960](https://arxiv.org/abs/2408.00960)\.
- Hu et al\. \(2022\)Edward J\. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen\-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen\.LoRA: Low\-rank adaptation of large language models\.In*International Conference on Learning Representations \(ICLR\)*, 2022\.URL[https://arxiv\.org/abs/2106\.09685](https://arxiv.org/abs/2106.09685)\.arXiv preprint 2021\.
- Huang et al\. \(2024\)Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin\.LoraHub: Efficient cross\-task generalization via dynamic lora composition\.In*COLM*, 2024\.URL[https://arxiv\.org/abs/2307\.13269](https://arxiv.org/abs/2307.13269)\.
- Kadavath et al\. \(2022\)Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield\-Dodds, Nova DasSarma, Eli Tran\-Johnson, et al\.Language models \(mostly\) know what they know, 2022\.
- Kirkpatrick et al\. \(2017\)James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A\. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska\-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell\.Overcoming catastrophic forgetting in neural networks\.*Proceedings of the National Academy of Sciences \(PNAS\)*, 114\(13\):3521–3526, 2017\.URL[https://arxiv\.org/abs/1612\.00796](https://arxiv.org/abs/1612.00796)\.
- Lester et al\. \(2021\)Brian Lester, Rami Al\-Rfou, and Noah Constant\.The power of scale for parameter\-efficient prompt tuning\.In*Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, 2021\.URL[https://arxiv\.org/abs/2104\.08691](https://arxiv.org/abs/2104.08691)\.
- Lewis et al\. \(2020\)Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen\-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela\.Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2020\.URL[https://arxiv\.org/abs/2005\.11401](https://arxiv.org/abs/2005.11401)\.
- Li & Liang \(2021\)Xiang Lisa Li and Percy Liang\.Prefix\-tuning: Optimizing continuous prompts for generation\.In*Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics \(ACL\)*, 2021\.URL[https://arxiv\.org/abs/2101\.00190](https://arxiv.org/abs/2101.00190)\.
- Li et al\. \(2023\)Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B\. Hashimoto\.AlpacaEval: An automatic evaluator of instruction\-following models\.[https://github\.com/tatsu\-lab/alpaca\_eval](https://github.com/tatsu-lab/alpaca_eval), 2023\.
- Li & Hoiem \(2018\)Zhizhong Li and Derek Hoiem\.Learning without forgetting\.*IEEE Transactions on Pattern Analysis and Machine Intelligence \(TPAMI\)*, 40\(12\):2935–2947, 2018\.URL[https://arxiv\.org/abs/1606\.09282](https://arxiv.org/abs/1606.09282)\.
- Lin et al\. \(2022\)Stephanie Lin, Jacob Hilton, and Owain Evans\.TruthfulQA: Measuring how models mimic human falsehoods\.In*Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(ACL\)*, 2022\.URL[https://arxiv\.org/abs/2109\.07958](https://arxiv.org/abs/2109.07958)\.
- Liu et al\. \(2025\)Langming Liu, Shilei Liu, Yujin Yuan, Yizhen Zhang, Bencheng Yan, Zhiyuan Zeng, Zihao Wang, Jiaqi Liu, Di Wang, Wenbo Su, Pengjie Wang, Jian Xu, and Bo Zheng\.UQABench: Evaluating user embedding for prompting LLMs in personalized question answering, 2025\.URL[https://arxiv\.org/abs/2502\.19178](https://arxiv.org/abs/2502.19178)\.
- Liu et al\. \(2023\)Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu\.G\-Eval: NLG evaluation using GPT\-4 with better human alignment\.In*Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, 2023\.URL[https://arxiv\.org/abs/2303\.16634](https://arxiv.org/abs/2303.16634)\.
- McMahan et al\. \(2017\)H\. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas\.Communication\-efficient learning of deep networks from decentralized data\.In*Proceedings of the 20th International Conference on Artificial Intelligence and Statistics \(AISTATS\)*, 2017\.URL[https://arxiv\.org/abs/1602\.05629](https://arxiv.org/abs/1602.05629)\.
- Ouyang et al\. \(2022\)Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al\.Training language models to follow instructions with human feedback\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2022\.URL[https://arxiv\.org/abs/2203\.02155](https://arxiv.org/abs/2203.02155)\.
- Packer et al\. \(2023\)Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G\. Patil, Ion Stoica, and Joseph E\. Gonzalez\.MemGPT: Towards llms as operating systems, 2023\.
- Park et al\. \(2023\)Joon Sung Park, Joseph C\. O’Brien, Carrie J\. Cai, Meredith Ringel Morris, Percy Liang, and Michael S\. Bernstein\.Generative agents: Interactive simulacra of human behavior\.In*Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology \(UIST\)*, 2023\.URL[https://arxiv\.org/abs/2304\.03442](https://arxiv.org/abs/2304.03442)\.
- Pfeiffer et al\. \(2021\)Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych\.AdapterFusion: Non\-destructive task composition for transfer learning\.In*Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics \(EACL\)*, 2021\.URL[https://arxiv\.org/abs/2005\.00247](https://arxiv.org/abs/2005.00247)\.
- Qwen Team \(2025\)Qwen Team\.Qwen3 technical report, 2025\.URL[https://qwenlm\.github\.io/blog/qwen3/](https://qwenlm.github.io/blog/qwen3/)\.Qwen3\-4B base model used in primary experiments\.
- Salemi et al\. \(2024\)Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani\.LaMP: When large language models meet personalization\.In*Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(ACL\)*, 2024\.URL[https://arxiv\.org/abs/2304\.11406](https://arxiv.org/abs/2304.11406)\.
- Schulman \(2023\)John Schulman\.Reinforcement learning from human feedback: Progress and challenges\.Berkeley EECS Colloquium talk, 2023\.URL[https://www\.youtube\.com/watch?v=hhiLw5Q\_UFg](https://www.youtube.com/watch?v=hhiLw5Q_UFg)\.
- Wang et al\. \(2022\)Yaqing Wang, Subhabrata Mukherjee, Xiaodong Liu, Jing Gao, Ahmed Hassan Awadallah, and Jianfeng Gao\.AdaMix: Mixture\-of\-adaptations for parameter\-efficient model tuning\.In*Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, 2022\.URL[https://arxiv\.org/abs/2205\.12410](https://arxiv.org/abs/2205.12410)\.
- Wu et al\. \(2024\)Xun Wu, Shaohan Huang, and Furu Wei\.Mixture of LoRA experts, 2024\.
- Xiao et al\. \(2023\)Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff\.C\-pack: Packaged resources to advance general Chinese embedding, 2023\.BGE embedding model family used for retrieval\.
- Xu et al\. \(2024\)Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang\.A\-MEM: Agentic memory for LLM agents, 2024\.TODO verify eprint id; original cited as 2410\.10739\.
- Yang et al\. \(2023\)Yuqing Yang, Ethan Chern, Xipeng Qiu, Graham Neubig, and Pengfei Liu\.Alignment for honesty, 2023\.URL[https://arxiv\.org/abs/2312\.07000](https://arxiv.org/abs/2312.07000)\.
- Zheng et al\. \(2023\)Lianmin Zheng, Wei\-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P\. Xing, Hao Zhang, Joseph E\. Gonzalez, and Ion Stoica\.Judging LLM\-as\-a\-judge with MT\-Bench and chatbot arena\.In*Advances in Neural Information Processing Systems \(NeurIPS\) Datasets and Benchmarks Track*, 2023\.URL[https://arxiv\.org/abs/2306\.05685](https://arxiv.org/abs/2306.05685)\.
- Zhong et al\. \(2024\)Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang\.MemoryBank: Enhancing large language models with long\-term memory\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, 2024\.URL[https://arxiv\.org/abs/2305\.10250](https://arxiv.org/abs/2305.10250)\.

## Appendix AAppendix A: Method \(full detail\)

We frame personalization as two orthogonal sub\-problems —*behavioral consistency*and*factual calibration*— and design a single evaluation matrix that exposes substrate behavior on both\. The matrix is four memory configurations×\\timesthree probes×\\timestwo corpora \(one synthetic, one real\)\. Configurations are held fixed across probes; probes are held fixed across corpora\. This isolates the substrate effect from data effects\.

### A\.1Memory configurations

All four configurations share a single base model \(Qwen3\-4B\(Qwen Team,[2025](https://arxiv.org/html/2606.11712#bib.bib28)\), instruction\-tuned, frozen\)\. They differ only in how user\-specific context is injected\.

ConfigWhat is conditioned onSubstrateB\_nohistnothing \(zero history\)none — base prior onlyB\_fullfull user backstory inlined into the promptinfinite\-context oracleC\_ragtop\-K=5 BGE\-retrieved chunks of the user’s historyretrievalC\_loraper\-userγ\\gamma\-LoRA adapter \(per\-corpus hyperparameters; see table below\)parametricB\_nohistandB\_fullare the headroom anchors: any system that matchesB\_nohisthas not used the user’s history at all; any system that matchesB\_fullhas perfectly used it\.C\_ragandC\_loraare the two substrates under comparison\. We deliberately pick the two*simplest*members of their respective families — flat top\-K retrieval and per\-user LoRA — to isolate the substrate effect from architecture\-engineering noise\. Hybrid configurations \(LoRA \+ RAG, calibration heads\) are introduced ablatively where a probe’s result motivates them \(§3\.3\)\.

Theγ\\gamma\-LoRA recipe varies by probe corpus along two dimensions, both chosen to match the per\-user data\-volume of the corpus:

ProbeCorpusrα\\alphaepochsper\-user pairsSynthetic personae3K\-char backstory \+ ~12 synth\-QA12825620~12BehavioralWritingPrompts \(≤\\leq50 stories\)641283≤\\leq50 stories rendered as next\-sentence LM chunksReal\-data transfer \(exp F\)LaMP\-3 \(≤\\leq10 review→\\rightarrowrating pairs\)12825620~5The synthetic\-personae and LaMP\-3 arcs use identical hyperparameters \(r=128,α\\alpha=256, 20 epochs\); they share a per\-user training loop \(train\_lorainexperiments/19\_lora\_synthqa\_eval\_v3\.py\) and a common assistant\-only loss\-mask\. The behavioral arc uses lower capacity and far fewer epochs because the WritingPrompts corpus has much more text per user \(~50 stories≈\\approx10K\+ tokens\) and language\-modeling supervision is denser than a question\-answer pair, so r=128×\\times20 epochs would overfit catastrophically\. We hold lr=2e\-4 fixed across all arcs\. AdamW, dropout 0\.05, no early\-stopping\. Each per\-user training run completes in≤\\leq10 minutes on a single L40S GPU\. RAG uses BGE\-large\-en\-v1\.5\(Xiao et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib33)\)with cosine similarity over 256\-token chunks of the user’s backstory and a single abstain clause appended to the system prompt:*“If retrieved chunks are not relevant, say ‘no, we have not discussed that\.’ ”*The abstain clause is the only prompt\-level calibration; we report its effect explicitly in §4\.2 since it ends up being load\-bearing for the substrate\-asymmetry result\.

### A\.2Probe 1 — Behavioral memory \(style\)

Why WritingPrompts\(the original plan was Reddit personal posts, swapped to WritingPrompts when the Reddit auth path proved more friction than budgeted\)\. WritingPrompts is a public HF dataset \(euclaise/writingprompts, no auth\) where multiple stories are written by the same author in response to different prompts\. Authors with≥\\geq50 stories of length≥\\geq200 chars give us a stable voice under varying topical content — the desideratum we want to test for\. The downside is that all WritingPrompts authors share a strong genre register \(third\-person fiction, present\-tense narration\), which makes idiosyncratic author markers subtle deltas on top of a shared baseline\. The behavioral arc therefore tests style transfer in the*hardest\-case*setting; a personal/conversational corpus would likely show a largerγ\\gamma\-LoRA advantage and we treat the WritingPrompts result as a lower bound on the effect\.

Behavioral memory probe: WritingPrompts continuation log\-likelihood\.We sample 50 high\-volume single authors from the WritingPrompts corpus \(euclaise/writingpromptson HuggingFace\), seed\-43 deterministic, filtered to authors with≥\\geq50 stories of length≥\\geq200 characters\. For each author, we hold out 5 stories, mask the last 100 characters of each, and score the gold continuation under each of the four configurations\. The metric is mean log\-likelihood per token of the gold continuation, scored by the same Qwen3\-4B base model used for generation\. n = 250 records \(50 users×\\times5\)\.

We pre\-register thestructural falsifierasΔ\\Delta\> 0 with a 95% CI that excludes zero — i\.e\. the directional claim*“per\-user parametric adaptation shifts the generation distribution toward the user’s voice in a way retrieval cannot\.”*Below this we reject the structural claim\. Above it we accept the structural claim\.

We additionally name a non\-pre\-registeredmagnitude targetofΔ\\Delta≥\\geq0\.5 nat/tok as a target for “the effect is large enough to read as a paragraph\-level voice transfer rather than a token\-distribution nudge\.” This is a target, not a falsifier: the original spec included it, but it was an ad\-hoc number with no theoretical grounding \(cf\. §6 limitations\), so we treat the structural test as the primary pre\-registration\. §4\.1\.2 reports both: the structural test passes decisively \(CI \[0\.381, 0\.446\], excluding zero by 23 SE\), and the magnitude target is missed by a small margin \(CI upper bound 0\.446 < 0\.500\)\. Both are reported transparently\. The pre\-registration is inexperiments/I\_context\.mdand is enforced by the post\-run audit checklist\.

Aggregate and per\-record artifacts are released at the project repository, with results writeup inexperiments/28\_results\.md\.

### A\.3Probe 2 — Calibration / absence detection

Calibration probe: presence/absence TPR on the synthetic\-personae corpus\.For each of 50 held\-out V3 personae, we generate 12 yes/no probes about specific events in the persona’s backstory: 6*presence*probes \(the event happened and is in the backstory\) and 6*absence*probes \(a plausible\-sounding event that didnothappen\)\. The metric is the true\-positive rate per kind: did the system answer correctly? “Correctly” means a textual yes/no judgement against the ground\-truth label\. We report per\-kind TPR and the harmonic\-mean F1\. n = 300 records per \(config×\\timeskind\)\.

We pre\-register a structural falsifier:C\_lora absence TPR≥\\geq60% AND C\_rag absence TPR≤\\leq35%would confirm that retrieval suffers a structural failure mode on absence\. The complementary branches —C\_rag absence handled wellandC\_lora absence < 30%— are documented as alternative findings inexperiments/J\_context\.md’s decision matrix\. The pre\-registration commits us to whichever branch the data picks; §4\.2 reports the calibration\-asymmetry branch the experiment landed in\.

We additionally evaluateC\_lora\_calib:γ\\gamma\-LoRA with the same abstain clause thatC\_raguses\. This isolates whether the calibration behavior comes from retrieval per se, or from the prompt\-level abstain instruction\.

Calibration\-head architecture \(C\_lora\_calib\)\.The “calibration head” in this paper is*not*a learned auxiliary classifier; it is a prompt\-level abstain instruction layered on the sameγ\\gamma\-LoRA weights used byC\_lora\. Concretely, the model and adapter are identical toC\_lora\(Qwen3\-4B \+ per\-persona LoRA, rank/alpha and training pairs unchanged\); only the system message changes\.C\_lorauses a plain system prompt \(“You are a helpful assistant who knows the user from prior conversations\.”\), whileC\_lora\_calibappends the instruction “If you are not sure whether the user has previously discussed a topic with you, answer ‘No, we have not discussed that\.’ Do not fabricate a memory\.” No additional training or feature engineering is performed; the same abstain clause is also whatC\_raguses, which makesC\_lora\_calibthe*prompt\-controlled*counterpart ofC\_loraand isolates the prompt’s contribution from retrieval’s contribution\. Reference:experiments/29\_runner\.py\(PLAIN\_SYS,CALIB\_SYS,\_run\_config\)\.

Aggregate and per\-record artifacts are released at the project repository, with results inexperiments/29\_results\.md\.

### A\.4Probe 3 — Real\-data transfer

Real\-data transfer probe \(LaMP\-3\): rating prediction on real Amazon\-style reviews\.We sample 50 LaMP\-3 held\-out users \(seed 43\) from the LaMP\-3 dev split prepared byexperiments/F\_data\_prep\.py\. Per user, the user’s history is split 5 train / 5 sanity\-held; the eval split contributes 4 main \+ 4 probe2 paraphrase records, for n = 200 main \+ n = 200 probe2 = 400 total\. The same per\-userγ\\gamma\-LoRA recipe as in §3\.2/§3\.3 is applied\. Judging is exact\-string\-equal to the gold integer rating after normalization\. The majority\-class baseline \(always predict “5”\) is reported alongsideγ\\gamma\-LoRA: in LaMP\-3, ratings are heavily skewed toward 5 \(\{5: 238, 4: 102, 3: 35, 1: 15, 2: 10\}\), so the majority baseline of 59\.5% is the right floor for “uses the user’s history at all\.”

We donotpre\-register a falsifier on the LaMP\-3 transfer probe\. The role of LaMP\-3 is*transfer measurement*: given that synthetic\-personae results \(§B\.1, §B\.2\) are positive on the substrate\-asymmetry axes, does the parametric substrate carry over to a noisy real\-user benchmark? §4\.3 reports the answer \(it does not —γ\\gamma\-LoRA underperforms the majority baseline by 28pp on main accuracy\) and §6 discusses the implications\. The synthetic↔\\leftrightarrowreal gap is a finding in its own right, not a null\.

Aggregate and per\-record artifacts are released at the project repository, results inexperiments/F\_results\.md\.

### A\.5Mechanism probe \(n = 50\)

Mechanism analysis: weight\-delta Frobenius decomposition \(n = 50\)\.We retrainγ\\gamma\-LoRA from scratch on all 50 held\-out personae with the adapter weights persisted viasave\_pretrained\(47 new adapters plus 3 reused from the earlier n = 3 pilot — the same personae used in §B\.1/§B\.2 training loops did not save adapters, so this is the first run that does so at scale\)\. For each adapter we decompose the per\-layer Frobenius norm ofW\_lora =α\\alpha/r⋅\\cdotBAacross the four attention target projections \(q\_proj,k\_proj,v\_proj,o\_proj\) for all 36 layers, then aggregate across the 50 personae to identify which \(layer, projection\) cells bear the largest and most consistent parametric mass\.

The probe isload\-bearingfor §5: at n = 50 the localization pattern is stable \(across\-persona coefficients of variation in the top band are 0\.08–0\.12, indicating a shared structural mechanism rather than per\-persona idiosyncrasy\)\. Pilot artifacts at n = 3 \(the n=3 pilot\) are superseded — we discuss that refinement explicitly in §5 rather than reporting both\. Source\-of\-truth artifacts are the aggregate Frobenius summary, the per\-personaΔ\\DeltaW matrices, andexperiments/30\_results\.md\.

Caveats\.FFN projections \(gate/up/down\_proj\) were not collected at n = 50; the runner’s diff loop iterates over attention projections only\. The headline claim is “attention\-q is the mechanism,” and the top\-10 cells being attention is itself sufficient evidence for that claim, but the explicit attention\-vs\-FFN ratio is not re\-verified at n = 50 \(the n = 3 pilot reported it as small but the comparison is underpowered\)\. We flag this for camera\-ready or reviewer follow\-up: running the FFN diff over the existing 50 adapters is ~10 min of GPU work\. Per\-persona behavior correlation \(Frobenius mass vs\. §B\.1 style log\-likelihood / §B\.2 absence TPR\) was also not computed in this run; we discuss this in §5 and §6\.

### A\.6Synthetic\-personae corpus \(V3\)

The synthetic corpus used throughout the paper is generated as follows\. We seed 120 personae from a frontier LLM \(Anthropic Claude, accessed via the AWS Bedrock API\) generation pipeline; 119 succeeded \(one dropped due to malformed JSON\), split index\-based into 100 train \+ 19 held\-out for the V3\-α\\alphahypernet study, then re\-split for the 50\-held\-out scale\-up reported in §B\. Each persona has a ~3K\-character free\-text backstory, 6 main probe questions \(factual, e\.g\.*“Where did Margaret get her BFA?”*\) and 6 probe2 paraphrases of the main probes\. A leakage diagnostic \(B\_nohist; base model with no history\) at full scale flagged 39/600 = 6\.50% of probes as having the answer recoverable without persona context \(vs an earlier 60\-probe sample at 1/60 = 1\.67%; the sample under\-estimated\)\. The asymmetry is informative: present\-fact probes leak at 10/300 = 3\.33%, absence probes at 29/300 = 9\.67% — absence questions leak ~3×\\timesmore often because the gold answer \(a hedge / “no/never mentioned”\) aligns with the model’s no\-context default\. Per\-question leakage flags are released with the corpus and are excluded from accuracy math wherever B\_nohist would otherwise score correct on a leaked probe\. Full generation pipeline inexperiments/V3\_data\_prep\.py, artifacts inpersonae\.jsonin the corpus distribution\.

Train/eval split for routing experiments \(§4\.7\)\.Of the 50 held\-out personae, the routing experiments use a random 30/20 partition is released assplit\.jsonwith the supplementary corpus\. Three properties of this split are deliberate and worth flagging up front:

1. 1\.Thematic overlap is by design\.Hobbies, occupations, and chronic conditions are reused across the train and eval splits — 20 of 27 distinct hobbies appear in both splits \(e\.g\. “vinyl jazz collecting” in 4 train \+ 4 eval personae; “highland bagpipes” in 5 train \+ 2 eval\)\. A routing classifier evaluated on a disjoint\-themes split would learn shortcuts \(“if the question mentions vinyl, route to LoRA”\) that generalize trivially\. Theme overlap forces the classifier to learn the*kind*axis \(presence vs absence\) rather than memorize topics\.
2. 2\.Specific\-fact leakage is bounded\.Verbatim eval\-question appearance in the train corpus is 2/240 \(both are template strings like “What city does this person currently live in?” whose answer is determined by*this*persona’s backstory\)\. Cross\-persona answer leakage on content\-bearing tokens is 6\.2% \(15/240\) and is dominated by recurring entity classes \(“celiac disease”, “Italic”, “Blue Note”\); the persona\-specific*combination*of question and answer is unique\.
3. 3\.Demographic balance is not stratified\.The 30/20 split is random, not stratified by occupation, location, or age\. Some cells are imbalanced \(executive chef 3 train / 0 eval; Marseille 4 train / 0 eval; ages 30–39 7 train / 1 eval\)\. Bootstrap CIs and paired permutation tests over the eval split absorb this as sampling noise; we do not claim demographic invariance\.

One generation artifact\.A small number of personae share specific real\-world entities used as “prized possession” templates by the generator\. The clearest example: V3\_P\_120 \(eval\) and V3\_P\_132 \(train\) both claim to own “an original 1959 Blue Note pressing of John Coltrane’s*Giant Steps*” — and both contain the same factual error \(*Giant Steps*was originally released on Atlantic, not Blue Note\)\. Persona\-specific details around the entity \(price, city, date\) differ, so QA\-probe answers remain unique\. We did not regenerate to fix this; it is a realism flaw, not a contamination flaw, and does not affect any §4\.7 router metric\.

Release\.The synthetic\-personae corpus \(n=119\), the 600\-probe absence dataset \(50 held\-out personae×\\times6 main \+ 6 probe2\), the blind\-preference judge harness \( its prompt template\), and the human\-eval sheet \(n=30, used for the §4\.1\.6 inter\-raterκ\\kappaanalysis\) are available at the \(anonymized for review\) project repository, pinned to the commit hash listed in the camera\-ready footnote\. We expect the dataset to be the most reusable artifact of this work: the framework is the contribution, the probes are what reviewers and follow\-up authors can run against alternative substrates without re\-deriving any infrastructure\.

### A\.7Judging and reproducibility

Free\-text probes \(V3 main / probe2; LaMP\-3 main / probe2\) are judged by Anthropic Claude Sonnet 4\.6 accessed via the AWS Bedrock API\. For the WritingPrompts probe we score the gold continuation directly under the base model and report log\-likelihood, so no LLM judge is needed\. Sanity rejudges confirm that judge artefacts do not explain the LoRA\-forgot phenomenon \(class\-b dominance 88%, rejudgecorrect = 0%\)\.

All experiment runners useargparsewith explicitchoices=enums for slice / mode, deterministic seed\-43 sampling sorted by canonical ID, and per\-recordif path\.exists\(\): return cachedresume guards\. Aggregate / per\-record artifacts are released with the camera\-ready version, with each result bound to the code and prompts that produced it\.

### A\.8What this design buys us \(and what it does not\)

Buys\.A fixed 4\-config matrix lets each probe report substrate asymmetry using the same axes; theB\_nohist/B\_fullanchors give absolute calibration even when probe metrics differ \(log\-likelihood vs\. TPR vs\. accuracy\)\. The synthetic \+ real pairing makes it possible to claim*substrate behavior*on the synthetic side while still reporting an honest*transfer*result on the real side\. The pre\-registered falsifiers prevent post\-hoc magnitude shopping\.

Does not buy\.The choice of two simple substrates means we cannot answer “what would Letta / MemGPT / a hand\-tuned hybrid do on these probes” — we explicitly defer that to future work\. Ourγ\\gamma\-LoRA recipe is per\-user from scratch; we do not study a single\-LoRA\-per\-task or a base\-LoRA\-plus\-user\-delta variant\. The synthetic corpus is generated by an LLM and its 6\.5% full\-scale leakage rate \(3\.3% on present\-fact probes, 9\.7% on absence probes; see §3\.6\) caps how much we can claim about the absolute floor ofB\_nohist; we report all numbers excluding leaked probes and discuss the limitation in §6\.

## Appendix BAppendix B: Results \(full per\-axis detail\)

> Source\-of\-truth:experiments/\{28,29,F,G\}\_results\.md\. Numbers in this section are copied verbatim from the committedaggregate\.jsonfiles referenced therein\. No new computation is done at draft time\.

We organize results around the three probe types defined in §3 \(behavioral / absence / real\-data\) and close with a descriptive mechanism look\. The headline pattern is asubstrate–task asymmetry that does not survive synthetic\-to\-real transfer:

1. 1\.γ\\gamma\-LoRA dominates RAG onstyle continuation\(synthetic and semi\-synthetic, §4\.1\)\.
2. 2\.RAG dominatesγ\\gamma\-LoRA onabsence detection, even when prompt\-engineering is held fixed across substrates \(§4\.2\)\.
3. 3\.Neither finding survives transfer to a real\-user benchmark\(LaMP\-3\):γ\\gamma\-LoRA underperforms a one\-line majority predictor by 28pp \(§4\.3\)\. This is the load\-bearing negative of the paper, and it is what motivates treating the synthetic asymmetry as a*measurement*rather than a deployment recipe\.
4. 4\.Mechanism analysis at n=50 \(§4\.4\) localizesγ\\gamma\-LoRA’s behavioral effect to mid\-to\-late attentionq\_proj; this is descriptive evidence that the §4\.1 effect has structure, not a causal intervention\.

Together these findings are sharper than any one in isolation: the synthetic asymmetry makes a clean substrate claim; the transfer failure makes that claim a measurement instead of a deployment prescription; the mechanism gives the asymmetry localization without overclaiming causality\.

### B\.1Behavioral memory:γ\\gamma\-LoRA writes style, RAG does not \(Phase I, exp28\)

Setup recap \(full detail §3\.2\)\.WritingPrompts, 50 high\-volume single\-author users, 5 held\-out continuation records each \(n = 250\)\. Per\-userγ\\gamma\-LoRA r=64 /α\\alpha=128 / 3 epochs \(lower\-capacity recipe matched to WritingPrompts’ larger per\-user corpus; see §3\.1 for the per\-arc hyperparameter table\) vs\. top\-K=5 BGE retrieval over the same per\-user corpus, vs\. B\_full \(full history in context\) and B\_nohist \(base, no history\) anchors\. Primary metric: per\-token log\-likelihood of the gold continuation under each system\.Scoring transparency:the LoRA config is scored under its own LoRA\-modified Qwen3\-4B; the three base configs \(B\_nohist, B\_full, C\_rag\) are scored under the base Qwen3\-4B with the appropriate prompt prefix\. This isolates the effect of the LoRA’s modified next\-token distribution from prompt conditioning, but it does mean LoRA and the three base configs are scored under*different forward passes of the same architecture*— we treat this as the right setup for the structural claim \(“does parametric adaptation shift the conditional distribution?”\) and address it explicitly in §6 limitations\.

#### B\.1\.1Mean log\-likelihood

ConfigMean LL \(nat/tok\)Δ\\Deltavs B\_nohistB\_nohist−\-3\.45520\.000C\_rag−\-3\.3952\+0\.060B\_full−\-3\.3792\+0\.076C\_lora−\-2\.9818\+0\.473C\_lora delivers ~6×\\timesthe style\-transfer effect of B\_full \(full history in\-context\) and ~8×\\timesthe effect of C\_rag\.

B\_full barely beats C\_rag \(\+0\.060 vs \+0\.076 nat/tok\)\.This is itself a finding: at the model scale we test \(Qwen3\-4B\), prepending the user’s full history into the context window provides only a marginal improvement over retrieving 5 chunks\. Both are an order of magnitude weaker than parametric adaptation\. Two interpretations are consistent with this result, and we are agnostic between them: \(i\) in\-context conditioning at this model scale is a weak channel for style transfer \(consistent with reports that small models exploit in\-context demonstrations less effectively than larger models for non\-trivial tasks\); and \(ii\) per\-token LL on a 22\-token tail is too short a window to measure a strong B\_full advantage even when one exists\. The finding strengthens the structural claim \(parametric beats both retrieval and full\-history\-in\-context\), but the relative ordering ofB\_fullandC\_ragshould not be over\-read as evidence that retrieval is “as good as” infinite context — we suspect \(i\) dominates and a larger base model would widen the B\_full margin\.

#### B\.1\.2Pre\-registered structural test \(C\_lora vs C\_rag\)

- •Δ\\Delta\(LL\_C\_lora−\-LL\_C\_rag\):\+0\.4134 nat/tok
- •SE \(per\-record paired\): 0\.0168
- •CI95:\[0\.3806, 0\.4463\]— excludes 0 by ~23 SEs\.
- •Pre\-registered structural falsifier \(Δ\\Delta\> 0 with CI excluding 0\):PASSED decisively\.
- •Sign\-test:240/250 recordsprefer C\_lora \(96%\)\.
- •Persona consistency:50/50 personaehave mean\(C\_lora\) \> mean\(C\_rag\); 10/50 personae individually clear the 0\.5 nat/tok target\.

Headline\.The structural pre\-registration passes decisively: the bootstrap CI \[0\.381, 0\.446\] excludes zero by 23 standard errors, 240/250 records and 50/50 personae preferγ\\gamma\-LoRA, and RAG’s effect is indistinguishable from the no\-history baseline\.*Parametric adaptation shifts the generation distribution; retrieval does not\.*

Magnitude target \(ad\-hoc, recalibrated prior\)\.The original spec also namedΔ\\Delta≥\\geq0\.5 nat/tok as an informal target\. We did so without theoretical grounding — the number was a back\-of\-envelope estimate of “paragraph\-level voice transfer\.” It is not a falsifier\. The observedΔ\\Delta= 0\.413 falls 0\.087 below this ad\-hoc number, with the CI upper bound at 0\.446\. We treat this as acalibration updateon what per\-token log\-likelihood lift looks like for a behavioral\-style adapter on a third\-person fiction corpus: a tight, decisively non\-zero shift on the order of 0\.4 nat/tok is what the metric produces in this regime, and future work in this space should anchor expectations there rather than at the 0\.5 figure we initially named\. We report the gap transparently and revise the prior; the structural claim is unaffected\.

#### B\.1\.3Reading

The structural claim —*retrieval cannot bias the generation distribution; parametric adaptation can*— is supported by the data without ambiguity\. RAG’s effect on style \(\+0\.060 nat/tok\) is indistinguishable from B\_full’s \(\+0\.076\), and both are an order of magnitude smaller thanγ\\gamma\-LoRA’s\. The held\-out continuations are*new text*\(the user has never written exactly this phrase\), so top\-K retrieval surfaces topically similar examples but cannot be copied verbatim — generation must come from parameters, andγ\\gamma\-LoRA is the only system that updates them\.

The ad\-hoc magnitude target sits 0\.087 above the observedΔ\\Delta= 0\.413 with a tight CI; we treat the gap as a calibration update on the informal target \(§4\.1\.2\), not as a falsifier outcome\. Two structural explain the shortfall: \(i\) median continuation length is ~22 tokens, at which per\-token LL noise from content \(largely topic\-driven, not style\-driven\) dilutes the style signal; and \(ii\) WritingPrompts authors share a strong genre register \(third\-person fiction, present\-tense narration\), so idiosyncratic author markers are subtle deltas on top of a strong shared baseline\. Reddit’s personal/conversational corpus \(the original Phase I plan\) would likely show a largerγ\\gamma\-LoRA advantage; the auth path was bypassed for wallclock reasons and remains queued as a magnitude follow\-up\.

#### B\.1\.4Blind\-preference judgement \(exp31, n=750\)

To corroborate the LL\-Δ\\Deltawith a perceptual signal we cannot obtain from token probabilities alone, we re\-evaluated the same 250 held\-out continuations under a blind A/B preference judge \(Sonnet 4\.6, side\-randomized, 3 prompt templates: lexical / semantic / combined, n=250 calls per template, n=750 macro\)\. The judge sees the gold continuation and two candidates \(γ\\gamma\-LoRA vs top\-K=3 RAG\), with no identifying labels\.

TemplateLoRA win rate95% CILoRA strictRAG strictTiessemantic0\.552\[0\.494, 0\.614\]13110514lexical0\.646\[0\.588, 0\.704\]1548115combined0\.596\[0\.538, 0\.654\]1449610macro0\.598\[0\.565, 0\.631\]42928239The macro win rate of 59\.8% sits 0\.2pp below the pre\-registered strict≥\\geq60% threshold but its bootstrap CI \[56\.5%, 63\.1%\] cleanly excludes 50%, and the lexical template alone clears 60% \(64\.6%, CI \[58\.8, 70\.4\]\)\. The structural claim —γ\\gamma\-LoRA produces output that human\-proxy judges prefer over RAG continuations of the same author — is supported\.

The template\-level pattern is itself informative\. Thelexicalprompt \(“which sounds more like the same author at the word/phrase level”\) delivers the cleanest signal \(64\.6%\): this is the axis the LL\-Δ\\Deltameasures, and the two metrics agree\. Thesemanticprompt \(“which conveys similar themes/ideas to the gold”\) is the weakest \(55\.2%\), with a CI lower bound \(49\.4%\) that grazes 50%\. We read this as honest scope:γ\\gamma\-LoRA is a*style/voice*adapter, not a content advantage — RAG’s retrieved chunks of the author’s prior stories are competitive on topical content but not on idiolect\. This is consistent with the broader paper claim thatγ\\gamma\-LoRA is a behavioral\-memory substrate, not a fact substrate\.

#### B\.1\.5Complementarity callout

Phase 0 \(exp24\) recorded a47% record\-flip ratebetween theγ\\gamma\-LoRA and RAG configs on the synthetic corpus: 53/114 records right under RAG\-only, 54/114 right underγ\\gamma\-LoRA\-only, with disjoint sets\. The two substrates correct different errors\. This complementarity argues against treating one as a strict replacement for the other and motivates the hybrid retrieval\-plus\-parametric story we return to in §5\.

#### B\.1\.6Human blind A/B \(n=30\)

To corroborate the LLM\-judge numbers \(§4\.1\.4\) with a non\-LLM signal we ran a single\-judge blind A/B human evaluation on n = 30 held\-out prefixes drawn from 30 distinct WritingPrompts users \(one prompt per user, sampled without replacement from the same Phase I held\-out pool\)\. The judge saw the author’s actual continuation as a voice\-reference and chose between two model continuations \(γ\\gamma\-LoRA vs top\-K=5 RAG\) with order randomized per row; unblinding came from a separate JSON key the judge did not consult during rating\. Confidence was recorded on a 1–3 scale \(1 = guess, 3 = clear difference\); ties were permitted but discouraged\.

Slicen\_decisiveLoRA winsRAG winsTieLoRA win ratep \(2\-sided\)All decisive picks29236179\.3%0\.0023Clean \(excl\. 3 rows where RAG bled scaffolding\)26206176\.9%0\.0094Confidence = 3 \(“clear”\) only17161094\.1%< 0\.001The confidence stratification is the most informative slice\. On the 17 rows the judge marked “clear difference,”γ\\gamma\-LoRA was preferred16 times; the only conf=1 \(“guess”\) row went to RAG\. Reader certainty and substrate preference are co\-monotone, which is what we would expect ifγ\\gamma\-LoRA is genuinely producing a more author\-like voice rather than a stylistically generic continuation that wins on unrelated cues\.

Leak\-robustness\.Three rows \(4, 13, 24\) contain RAG continuations that regurgitated the few\-shot prompt scaffolding \(“Continue the story in a fitting style…”\) instead of in\-character prose\. We treat these as legitimately scored — the eval question is which continuation better matches the author’s voice, and a frame\-breaking continuation does not — but report the win rate after their removal as a robustness check\. The headline survives: 76\.9% / p = 0\.0094 with 26 decisive picks\. Notablyall three leaks came from the RAG side and zero fromγ\\gamma\-LoRA, which is consistent withγ\\gamma\-LoRA being trained end\-to\-end on the user’s prose distribution while RAG conditions a base instruction\-following model on a few\-shot template the model can probabilistically continue back into\.

Relation to §4\.1\.4\.The LLM\-judge macro win rate was 59\.8% \(CI \[56\.5, 63\.1\]\) at n = 750; the human\-judge win rate is 79\.3% at n = 29\. The two numbers agree in direction and significance but differ in magnitude: this is consistent with humans being less generous than Sonnet on RAG outputs that are topically correct but stylistically off, and with the fact that the human judge was given the gold continuation as a calibration anchor whereas the LLM judge was not \(§3\.2 protocol\)\. We treat the LLM\-judge result as the load\-bearing per\-record measurement and the human\-judge result as a sanity check that the per\-record signal generalizes when a literate reader evaluates whole continuations\.

Limitations and second\-rater proxy\.Single human judge \(the first author\), n = 30\. As a sanity proxy ahead of a real human second judge, we ran Claude Sonnet 4\.6 over the same 30 rows with the identical style\-only prompt \(same gold\-anchor, same A/B presentation, temperature 0;scripts/llm\_judge\_human\_eval\.py\)\. The LLM judge splits 15/15 LoRA/RAG \(LoRA\-win = 50\.0%, two\-sided binomial p = 1\.0\) and agrees with the human on only 15/29 decisive rows \(51\.7%, Cohen’sκ\\kappa\_2way =−\-0\.020, “worse than chance”\)\. Two readings: \(i\)the LLM judge is poorly calibrated to the style\-vs\-content distinction this task requires— when both candidates are topically plausible and fluent, Sonnet falls back to coin\-flipping; or \(ii\)the human preference is partly an artifact of one reader’s idiosyncratic taste profilethat any other independent rater \(human or model\) would not reproduce\. We cannot distinguish these from the present data\. We therefore treat §4\.1\.6 as a*qualitative*sanity check on §4\.1\.4 — it agrees in sign with the LLM\-judge result on §4\.1\.4 — and explicitly flag a real second human judge \(M\_HUMAN\_SECOND\_JUDGE in the paper TODO\) as required for any load\-bearing claim that depends on the human\-preference magnitude\. Per\-row picks, confidence scores, scoring code, the LLM\-judge sheet, andkappa\.jsonare inruns/human\_eval/\. We do not retract §4\.1\.4 \(the LLM\-judge result on the larger n = 750 set\), which stands on its own; this is a limitation of §4\.1\.6 specifically\.

#### B\.1\.7Three\-judge cross\-vendor replication \(n=750\)

To probe whether §4\.1\.4’s LLM\-judge result is Sonnet\-specific or generalizes across model vendors, we replicate the exact n=750 blind\-preference protocol with two additional frontier judges: Anthropic Claude Opus 4\.8 \(a different scale within the same vendor\) and Amazon Nova Premier \(a different vendor entirely\)\. All three judges score the same 750 \(persona, eval\_idx, template\) triples deterministically \(temperature 0\)\.

JudgeLoRA %RAG %Tie %LoRA share \(decisive\)Sonnet 4\.6 \(original\)57\.237\.65\.260\.3%Opus 4\.866\.131\.91\.567\.5%Nova Premier45\.753\.90\.445\.9%Majority of 360\.337\.30\.5—PairCohen’sκ\\kappa\(all\)Cohen’sκ\\kappa\(decisive only, n=697\)Sonnet vs Opus0\.255 \(fair\)0\.276Sonnet vs Nova0\.109 \(poor\)0\.116Opus vs Nova0\.331 \(fair\)0\.349Fleiss’κ\\kappa\(3\-rater\)0\.2180\.232Two findings\.

Finding 1 \(vendor cluster\):the two Anthropic models agree moderately \(κ\\kappa= 0\.255–0\.331\) andboth preferγ\\gamma\-LoRA decisively\(60\.3% / 67\.5% LoRA share\)\. Nova flips to 45\.9% LoRA share — a ~20 pp split that aligns cleanly with vendor lineage, not random noise\. Pairwiseκ\\kappainvolving Nova is uniformly the lowest \(sonnet–nova = 0\.109≈\\approxchance\)\.Style\-judgment LLMs cluster by training pipeline, not by scale\.This is a stronger result than uniform agreement would have been: it shows that theγ\\gamma\-LoRA preference is robust to within\-vendor scale variation \(Sonnet 4\.6→\\rightarrowOpus 4\.8\) but vendor\- sensitive \(Anthropic↔\\leftrightarrowAmazon\)\.

Finding 2 \(majority consensus retains the headline\):the 3\-judge majority\-vote LoRA share is60\.3%at n=750, within 0\.5pp of Sonnet alone \(60\.3%\) and within 1pp of the v1\.10 paper’s headline59\.8%\(CI \[56\.5, 63\.1\]\)\. The headline is not a Sonnet artifact; two of three frontier judges replicate it\.

What we do not claim\.Fleiss’κ\\kappa= 0\.218 is “slight” agreement on the standard scale\. We do not interpret this as “judges agree style preference is universal\.” Rather, the structural claim \(γ\\gamma\-LoRA beats RAG on style for the*majority of frontier LLM judges, with direction independent of judge family*\) is what the n=750 evidence supports; the magnitude is vendor\-dependent\. Per\-judge data, per\-template breakdown, and theκ\\kapparecompute script are inruns/iclr2027\_push/three\_judge/\(opus\.jsonl,nova\.jsonl,three\_judge\_kappa\.json,three\_judge\_summary\.md; reproducible viascripts/run\_three\_judge\.py\+scripts/recompute\_kappa\.py\)\.

The §4\.1\.6 single\-human disagreement \(κ\\kappa=−\-0\.020\) is now contextualized by this three\-judge result: the LLM judges themselves disagree moderately, so a single human disagreeing with one LLM is consistent with general judge\-style noise, not a refutation of either\.

### B\.2Calibration asymmetry: RAG knows what isn’t there \(Phase J, exp29\)

Setup recap \(full detail §3\.3\)\.Same 50 personae as the synthetic baseline, 12 probes per persona×\\times4 configs = 2,400 records\. Probes split 6 presence \(gold = “yes, we discussed X”\) and 6 absence \(gold = “no, we have not discussed X”\)\. Configs: B\_nohist, C\_rag \(top\-K=5 with an abstain\-on\-low\-relevance prompt clause\), C\_lora \(per\-userγ\\gamma\-LoRA\), C\_lora\_calib \(per\-userγ\\gamma\-LoRA \+ the same abstain clause appended at inference\)\.

#### B\.2\.1Numbers

ConfigPresence TPRAbsence TPRF1B\_nohist3\.3%9\.7%0\.050C\_rag \(with abstain clause\)35\.3%99\.0%0\.521C\_lora56\.3%8\.7%0\.150C\_lora\_calib \(with abstain clause\)42\.0%53\.3%0\.470n = 300 per \(config×\\timeskind\)\.

#### B\.2\.2The fair comparison

A clean reading of the absence\-TPR gap requires controlling for the abstain prompt, which is shared byC\_ragandC\_lora\_calibbut absent from bareC\_lora\. We isolate two effects:

ComparisonAbsence TPR gapWhat it isolatesC\_rag vs B\_nohist\+89\.3ppretrieval \+ abstain clause vs base priorC\_lora\_calib vs C\_lora\+44\.6ppabstain clause alone \(LoRA substrate held fixed\)C\_rag vs C\_lora\_calib\+45\.7ppretrieval substrate effect, abstain clause held fixedC\_rag vs C\_lora\+90\.3ppcombined effect \(substrate \+ prompt — the marketing\-headline number, but conflated\)The substrate\-controlled comparison isC\_ragvsC\_lora\_calib\(both with abstain clause\): RAG still beats parametric memory on absence by 45\.7 percentage points\.The other 44\.6pp of the top\-line 90\.3pp gap comes from the abstain clause itself, whichγ\\gamma\-LoRA inherits identically when given the same prompt scaffolding\. We report the 90\.3pp top\-line in the abstract because it is what an end\-to\-end deployment sees, but the substrate\-controlled 45\.7pp is the load\-bearing scientific claim\.

#### B\.2\.3Reading

The pre\-registered structural falsifier —*retrieval fails to abstain when the answer isn’t in context, parametric memory abstains correctly*— isinverted\. RAG abstains on absence with a 99% TPR;γ\\gamma\-LoRA abstains 8\.7% of the time, lower than B\_nohist’s 9\.7%\. Even in the substrate\-controlled comparison \(both configs have the same abstain clause\), retrieval still leads parametric by 45\.7pp\. The direction of the failure is the opposite of the one we pre\-registered, and the magnitude is decisive on either accounting\.

The mechanisms are intuitive in retrospect:

- •C\_rag absence 99%combines two effects\. \(i\) The retrieval prompt contains a “say no if retrieved chunks aren’t relevant” clause\. \(ii\) Off\-topic absence probes return chunks with similarity well below the retrieval threshold, so the abstain clause has clear signal to fire on\. We isolate \(i\) above by givingγ\\gamma\-LoRA the same clause: this recovers 44\.6pp \(C\_lora\_calib53\.3% absence vsC\_lora8\.7%\), but cannot close the remaining 45\.7pp gap to RAG\. That residual is the substrate effect: retrieval*exposes the decision\-relevant signal*\(low max\-similarity score\) to the prompt in a way that parametric memory does not\. A purely\-parametric system cannot ask itself “have I seen this before?”; a retrieval\-based system can\.
- •C\_lora absence 8\.7%is consistent with the exp23/26 finding thatγ\\gamma\-LoRA fits per\-user training pairs to ~0 loss \(mean final loss < 1e\-3 across 50 personae\)\. The adapter has never seen a*negative*example during training, so it has no signal to teach abstention; it confidently completes “have we discussed X?” with a fabricated affirmation\. This is a structural property of supervised fine\-tuning on positives only, not a bug in our particular training loop\. Adding the abstain clause \(C\_lora\_calib\) recovers about half the gap by externalizing whatγ\\gamma\-LoRA cannot internalize\.

This is a*calibration*finding, not a capacity finding:γ\\gamma\-LoRA can hold the user’s facts \(Phase 0 main 63\.16%, probe2 58\.77%\) but it cannot say “I don’t have that one” without prompt scaffolding, and even with scaffolding it lags retrieval by 45\.7pp on the controlled comparison\. The substrate that wins on §4\.1 loses here, on either the conflated or the substrate\-controlled accounting\.

### B\.3Real\-data transfer: synthetic\-to\-real F3a does not hold \(Phase F\)

Setup recap \(full detail §3\.4\)\.LaMP\-3, 50 held\-out users, 4 main \+ 4 probe2 review→\\rightarrowrating pairs each \(n = 200 \+ 200\), same per\-userγ\\gamma\-LoRA recipe as Phase 0\. Majority baseline: always predict 5 \(the modal LaMP\-3 rating\)\.

#### B\.3\.1Numbers

MetricnAccmain2000\.315probe22000\.410sanity \(held\-train\)2450\.069Majority “5” baseline2000\.595Synthetic\-vs\-real gap:31\.7pp on main, 17\.7pp on probe2\.

#### B\.3\.2Reading

γ\\gamma\-LoRA underperforms a one\-line majority baseline on real LaMP\-3 by 28 percentage points absolute on main and 18\.5 percentage points on probe2\. The gap decomposes into two distinct measurable failures — instruction\-following collapse \(§4\.3\.3\) and substrate\-asymmetry amplification \(§4\.3\.4\) — both of which are visible in the per\-record data and neither of which is captured by the top\-line “0\.315 vs 0\.595” number\.

#### B\.3\.3Instruction\-following collapse: a third measurable failure mode

The task asks for a single integer in 1–5\. We measure two parse rates:

Parse criterionPassRateStrict: prediction is exactly"1"\.\."5"159/20079\.5%Loose: prediction contains*any*digit 1–5184/20092\.0%The strict\-format failure rate is20\.5% \(41/200\)\. The deviations split into two qualitatively different modes:

Sampleγ\\gamma\-LoRA output \(verbatim, strict\-deviant\)Mode1/5,4/5,2/5rating present, format wrong \(~13% of all\)French,At least 3,A smaller battery powered sprayer,An over\-priced, slightly delicate, but lovely bookreview\-text continuation in user’s style \(~8%\)The first cluster \(~13%\) is a measurement artifact a stricter parser would recover — the rating is correct, the format is “x/5”\. The second cluster \(~8%\) is the real instruction\-following collapse: the model produces fluent, domain\-appropriate review text in the user’s*style*and ignores the rating instruction entirely\. Twenty epochs of per\-user training over ~10 review→\\rightarrowrating pairs has shifted the output distribution far enough toward “produce review text in this user’s voice” that the “answer with a single integer” instruction is no longer reliably followed on the order\-of\-10% tail\.

Even with the most generous parser \(loose: any 1–5 digit anywhere\), the conditional accuracy on parsable predictions is34\.2% \(63/184\)— still25 percentage points below the 59\.5% majority baseline\. So the substrate is underperforming on real LaMP\-3*even after*controlling for any reasonable reading of the instruction\-following collapse\. The third failure mode is real and measurable, but it is not the only failure mode and it does not account for the synthetic\-vs\-real gap on its own\.

The implication for benchmark design is concrete: any paper reporting LaMP\-3 numbers without a parse\-rate decomposition is conflating substrate\-level personalization with instruction\-following capacity, and may be underreporting \(when strict parsers reject correct\-rating\-wrong\-format predictions\) or overreporting \(when the substrate is producing fluent off\-task text that loose parsers happen to score against\)\.

#### B\.3\.4Substrate\-asymmetry amplification

We flag three plausible co\-mechanisms \(§5 returns to these\):

1. 1\.Training\-signal poverty\.Synthetic personae have ~12 high\- quality factual Q/A pairs designed to be teachable; LaMP\-3 user histories yield ~10 noisy review→\\rightarrowrating pairs where ratings are heavily biased toward 5 and review text is only weakly correlated with the rating\.
2. 2\.Distribution shift train→\\rightarroweval\.Synthetic eval re\-asks \(paraphrased\) facts the user*was*trained on; LaMP\-3 eval reviews are*new*products the user never reviewed\. Rating tendency requires generalization, not memorization\.γ\\gamma\-LoRA’s loss\-zero training overfits to memorization\.
3. 3\.Instruction interference\.No KL\-to\-base regularizer on non\-LaMP queries; the LoRA degrades base instruction\-following on a measurable fraction of held\-out queries \(§4\.3\.3\)\.

These are not mutually exclusive, and we do not run a controlled mechanism analysis here\. \(Phase K, §4\.4 / §5\.6, runs on synthetic data where the positive result anchors the mechanism story\.\)

This result isthe load\-bearing negative of the paper\.It is what motivates the diagnostic\-framework framing: per\-userγ\\gamma\-LoRA, the substrate that wins cleanly in §4\.1, transfers to real data with*negative utility relative to a constant predictor*, and the shortfall splits cleanly into a measurable instruction\-following collapse plus an amplified substrate asymmetry\. Any paper that reports the synthetic positive without reporting this would overpromise\. We treat the synthetic\-vs\-real gap as a contribution in its own right — a measurement, not a null\.

### B\.4Mechanism:γ\\gamma\-LoRA edits mid\-to\-late attentionq\_proj\(Phase K, n = 50\)

For each of the 50 held\-out personae we retrainγ\\gamma\-LoRA on its synth\-QA pairs \(rank 128,α\\alpha= 256, 20 epochs\), save the adapter, and compute the Frobenius norm∥\\\|Δ\\DeltaW∥\\\|\_F =∥\\\|α\\alpha/r⋅\\cdotBA∥\\\|\_F of every \(layer×\\timesattention\-projection\) cell across the 36 transformer blocks and four attention projections \(q, k, v, o\)\. Aggregating across the 50 adapters:

Top\-10 cells\(mean across 50 personae\):

RankLayerProjectionMean∥\\\|Δ\\DeltaW∥\\\|\_Fcv across personae135q\_proj0\.7870\.105222q\_proj0\.7730\.119324q\_proj0\.7370\.106423q\_proj0\.7270\.086529q\_proj0\.7240\.092634q\_proj0\.7230\.080727q\_proj0\.7180\.087835o\_proj0\.7180\.102930q\_proj0\.7140\.0891021q\_proj0\.7100\.092Per\-projection means\(across all 36 layers\): q\_proj 0\.668, o\_proj 0\.558, v\_proj 0\.354, k\_proj 0\.331\. The all\-cell background mean is 0\.478; the top cell sits 1\.65×\\timesabove background, and the top\-10 band sits ~1\.5×\\timesabove background\.

The pattern at scale ismid\-to\-late attentionq\_projdominanceacross layers 21–35, witho\_projat L35 a secondary mode\. Across\-persona coefficients of variation in the top band are 0\.08–0\.12, indicating ashared structural mechanismrather than per\-persona idiosyncrasy\.

Refinement vs\. earlier n = 3 pilot\.A pilot run \(Phase G\) on 3 personae reported \(L35q\_proj, L22q\_proj, L30o\_proj\) as top\-3\. The first two cells survive at n = 50;L30o\_projdoes not— at n = 50 it falls to rank 31 and is replaced by a broader band of mid\-stackq\_projlayers \(L21, L23, L24, L27, L29\) that was not visible at n = 3\. We retract the L30o\_projpilot claim and report only the n = 50 numbers as primary\. §5 develops the mechanism story in detail\.

The contrast with the plan\-of\-record hypothesis \(“style stored in FFN\-down”\) is unresolved at n = 50 — the runner’s diff loop only collected attention projections; FFNgate/up/downwas not re\-measured\. Phase G’s n = 3 FFN\-small finding is too underpowered to repeat as a positive claim\. We flag the FFN ratio as future work \(§6 / camera\-ready\); running the diff over the existing 50 adapters is ~10 minutes of GPU work and is recoverable on demand\.

§5\.7 reports the causal upgrade: zeroing the L21–35q\_projband at inference time breaks both probes \(\+33pp absence\-TPR,−\-20pp presence\-TPR at n = 50\), confirming that the band identified here is not just where the largest weight movement lands, but where the substrate’s behavioral asymmetry is causally implemented\.

#### B\.4\.1Llama\-3\.1\-8B mechanism replication \(Phase K\-L\)

We rerun the same Frobenius weight\-delta decomposition on the 50 cached Llama\-3\.1\-8B\-Instruct LoRA adapters from §4\.6 / Phase 2 W0 \(rank 128,α\\alpha= 256, 20 epochs synth\-QA, target modules q/k/v/o identical to the Qwen recipe\)\. Llama\-3\.1\-8B has 32 transformer layers vs Qwen3\-4B’s 36, so layer indices are not directly comparable; we report fraction\-of\-stack to enable cross\-model reading\.

Top\-10 cells \(Llama\-3\.1\-8B, mean across 50 personae\):

RankLayerProjectionMean∥\\\|Δ\\DeltaW∥\\\|\_Fcv across personae131o\_proj0\.7790\.088222q\_proj0\.7410\.117320q\_proj0\.7010\.117423o\_proj0\.6970\.074515q\_proj0\.6930\.098617q\_proj0\.6910\.096728o\_proj0\.6910\.083821o\_proj0\.6900\.080931q\_proj0\.6880\.0921028q\_proj0\.6880\.111Per\-projection means\(across all 32 layers\): q\_proj 0\.654, o\_proj 0\.657, v\_proj 0\.341, k\_proj 0\.328\. Background mean \(all cells\) is 0\.495; top\-1 cell is 1\.57×\\timesbackground\. Across\-persona cv in the top band is 0\.07–0\.12, matching Qwen3\-4B\.

Cross\-model comparison \(Qwen3\-4B vs Llama\-3\.1\-8B\-Instruct\):

PropertyQwen3\-4BLlama\-3\.1\-8BReplicates?Top\-10 stack position60–100% \(L21–35 of 36\)47–97% \(L15–31 of 32\)✓ mid\-to\-lateTop\-10 projection mix9×\\timesq\_proj, 1×\\timeso\_proj6×\\timesq\_proj, 4×\\timeso\_proj✓ q\+o dominantTop\-1 cellL35 q\_projL31 o\_proj≈\\approx\(last block, attn\)Top\-1 / background1\.65×\\times1\.57×\\times✓ comparableq\_proj mean / k\_proj mean0\.668 / 0\.331 = 2\.0×\\times0\.654 / 0\.328 = 2\.0×\\times✓ identical ratioAcross\-persona cv \(top band\)0\.08–0\.120\.07–0\.12✓ comparableAttn\-vs\-FFNattn dominates \(FFN not in target\_modules\)attn dominates \(FFN not in target\_modules\)✓Reading\.The descriptive band\-shape replicates:γ\\gamma\-LoRA writes concentrate in upper\-half attention with q\_proj/o\_proj dominance over k\_proj/v\_proj at a 2×\\timesratio that is identical across both base models\. Specific layer indices do not align — Qwen3\-4B top\-10 sits in 60–100% of the stack while Llama\-3\.1\-8B sits in 47–97% — but thefraction\-of\-stack pattern is preserved, and the q\_proj\-dominant fingerprint of Qwen3\-4B becomes a more balanced q\+o split on Llama \(with o\_proj at L31 taking the top slot\)\. This is consistent with two non\-exclusive readings: \(i\)γ\\gamma\-LoRA\-as\-substrate has ashared mechanism axis— upper\-attention query/output writes — that any modern decoder exposes; \(ii\) Llama’s heavier RLHF post\-training shifts the write distribution toward o\_proj \(which sits closer to the residual stream output\) relative to Qwen, possibly because o\_proj edits propagate with less distortion through the RLHF\-anchored later layers\.

What the Llama mechanism does NOT yet establish\.The §5 causal arm \(band\-zero intervention at inference time, \+33pp absence\-TPR /−\-20pp presence\-TPR\) has not been re\-run on Llama\. The descriptive evidence above shows where weight movement lands; whether Llama’s L15–L31 attn\-q\+o band causally drives the calibration asymmetry the way Qwen’s L21–L35 q\_proj band does is the natural follow\-up\. We hold this open as future work; the 50 Llama adapters needed for the intervention are already cached atruns/30\_mechanism\_\_llama3\.1\-8b/V3\_P\_\*/lora/and the intervention recipe is the same as §5\.7\.

Source\-of\-truth artefacts:runs/30\_mechanism\_\_llama3\.1\-8b/summary\.json\(top\-10 cells, projection means, cv\-across\-personae\),runs/30\_mechanism\_\_llama3\.1\-8b/frobenius\_tensor\.pt\(50×\\times32×\\times7 per\-persona×\\timeslayer×\\timesprojection\),runs/30\_mechanism\_\_llama3\.1\-8b/heatmap\_layer\_x\_proj\.png\.

### B\.5Cross\-model replication: Llama\-3\.1\-8B\-Instruct \(Phase L\)

Setup recap\.A reviewer\-resistant test of whether the substrate asymmetry \(§4\.1, §4\.2\) is a property ofγ\\gamma\-LoRA\-as\-substrate or of the specific Qwen3\-4B base model\. We re\-ran the two headline probes onmeta\-llama/Llama\-3\.1\-8B\-Instructusingidentical recipes, identical corpora, identical prompts, swapping only three model\-touchpoints behind a registry abstraction \(experiments/\_base\_model\.py\): the base\-model snapshot path, the chat\-template stop token \(<\|im\_end\|\>→\\rightarrow<\|eot\_id\|\>\), and the Qwen\-onlyenable\_thinking=Falsechat\-template kwarg\. Run dirs suffix the model tag \(\_\_llama3\.1\-8b\) so Qwen and Llama artefacts live side\-by\-side\. n = 50 WritingPrompts authors / 50 V3 personae, matched to the Phase I / Phase J samples\.

The result is apartial replication that sharpens the paper’s claimrather than weakening it: the calibration asymmetry \(§4\.2\) is universal across both base models; the behavioural\-style advantage \(§4\.1\) is not\.

#### B\.5\.1Numbers

MetricQwen3\-4B \(v1\)Llama\-3\.1\-8BReplicates?§4\.1 style:Δ\\Deltaγ\\gamma\-LoRA – RAG \(nat/tok\)\+0\.413\+0\.003No§4\.1 style:Δ\\Deltaγ\\gamma\-LoRA – B\_nohist\+0\.473\+0\.041No§4\.1 95% CI onΔ\\Deltaγ\\gamma\-LoRA – RAG\[\+0\.381, \+0\.446\]\[−\-0\.016, \+0\.021\]—§4\.1 sanity: B\_full – B\_nohist \> 0✓✓ \(\+0\.041\)✓§4\.2 RAG absence\-TPR99\.0%96\.7%✓§4\.2γ\\gamma\-LoRA absence\-TPR8\.7%3\.0%✓ \(worse\)§4\.2 substrate gap \(RAG−\-γ\\gamma\-LoRA\), abs\+90\.3pp\+93\.7pp✓§4\.2 substrate\-controlled gap \(C\_rag−\-C\_lora\_calib\), abs\+45\.7pp\+79\.0pp✓ \(stronger\)§4\.2γ\\gamma\-LoRA presence\-TPR56\.3%69\.0%✓ \(better\)§4\.2 C\_rag F10\.5210\.477≈\\approx§4\.2 C\_lora F10\.1500\.057≈\\approx\(worse\)§4\.2 C\_lora\_calib F10\.4700\.278≈\\approx\(worse\)n\_records: §4\.1 Llama 245 / 49 personae \(one user dropped during adapter save; §4\.1 retains 49/50 reporting\); §4\.2 Llama 1,200 / 50 personae×\\times4 configs\. Confidence intervals are bootstrap over records, 10k resamples, matched to the Qwen protocol in §3\.

#### B\.5\.2Reading

Style does not replicate\.On Llama\-3\.1\-8B with identical recipe \(r=64,α\\alpha=128, 3 epochs LM\-loss on the user’s stories\),γ\\gamma\-LoRA’s mean log\-likelihood per token sits\+0\.003 nat/tok above RAGwith a 95% CI of \[−\-0\.016, \+0\.021\] that comfortably includes zero\. The structural pre\-registration that passed decisively on Qwen3\-4B \(§4\.1\.2, CI excluding zero by 23 SE\) does not pass on Llama\-3\.1\-8B\. The blind\-judge claim of §4\.1\.4 was not re\-run on Llama at this draft \(Bedrock budget\); the LL\-Δ\\Deltaresult alone is sufficient to establish the non\-replication\.

We see three plausible non\-exclusive mechanisms, none of which we can fully separate at the n = 50 / one\-recipe scope:

1. 1\.Headroom\.Llama\-3\.1\-8B’s B\_nohist mean LL on this corpus is−\-2\.77 nat/tok; Qwen3\-4B’s was−\-3\.21\. Llama is already a stronger baseline fiction\-completion model, so per\-user style adaptation has less distributional headroom to add value at fixed per\-token resolution\. The B\_full anchor confirms this: B\_full−\-B\_nohist = \+0\.041 on Llama vs \+0\.067 on Qwen, indicating the upper bound on achievable per\-user lift is itself smaller in absolute terms on the stronger base model\.
2. 2\.RLHF density\.Llama\-3\.1\-Instruct ships with heavier RLHF than Qwen3\-4B \(DPO \+ iterative SFT vs Qwen’s lighter instruction\-tune\)\. A more strongly anchored base distribution may absorb 3 epochs of LM\-loss without measurable distributional shift\. A scaled recipe \(r = 128 / 10 epochs / lower lr\) was not tried; we treat it as future work\.
3. 3\.Recipe\-specific resolution\.The exp28 recipe was tuned on Qwen3\-4B; nothing about it is Llama\-aware\. Whether a Llama\-tuned recipe would recover the \+0\.4 nat/tok lift is unanswered\.

The calibration asymmetry replicates and strengthens\.On the absence probe \(§4\.2\), the picture on Llama\-3\.1\-8B is:

- •RAG absence\-TPR 96\.7% \(vs Qwen 99\.0%\) — both substrates abstain reliably when the abstain prompt clause sees a low\-similarity retrieval\.
- •γ\\gamma\-LoRA absence\-TPR3\.0%\(vs Qwen 8\.7%\) — Llamaγ\\gamma\-LoRA is*worse*at recognizing absent facts than Qwenγ\\gamma\-LoRA\. It confidently confabulates almost every absence probe\.
- •The top\-line substrate gap is\+93\.7pp\(RAG−\-γ\\gamma\-LoRA\), almost identical in magnitude to Qwen’s \+90\.3pp\.
- •The substrate\-controlled gap \(C\_rag−\-C\_lora\_calib, both with the abstain prompt clause\) is\+79\.0ppon Llama vs \+45\.7pp on Qwen\. The substrate effect is*larger*on Llama once the prompt contribution is held fixed\.

The mechanism reading from §4\.2\.3 ports directly:γ\\gamma\-LoRA on either base model has no internal “I have not been told this fact” representation because it never trained on negative examples\. The calibration head \(C\_lora\_calib\) recovers part of the gap on Qwen \(F1 0\.470\) and substantially less on Llama \(F1 0\.278\), suggesting the calibration\-head architecture itself may need re\-tuning per base model — but the substrate\-level claim does not depend on that recovery\.

#### B\.5\.3What this changes for the framework

The substrate\-asymmetry claim, restated honestly:

> The calibration asymmetry is a property ofγ\\gamma\-LoRA\-as\-substrate and replicates on a second base model and family\. The behavioural\-style advantage ofγ\\gamma\-LoRA over RAG observed on Qwen3\-4B does not replicate on Llama\-3\.1\-8B at matched recipe; it may be base\-model\-specific, recipe\-specific, or both\.

This is a stronger position than the v1 framing for two reasons\.

First, it makes themeasurement frameworkthe load\-bearing contribution rather than the magnitude of any single number\. A reader who replicates on a third base model and finds the style claim restored has nonetheless used our probes to reach that conclusion; a reader who finds the absence claim broken has rejected a stronger and more universal claim than we make\.

Second, it lines up cleanly with the §4\.3 LaMP\-3 negative and the §4\.3\.3 instruction\-following collapse: per\-userγ\\gamma\-LoRA’s behavioural advantage isfragile to corpus, recipe, and base model, while the calibration failure isnot\. The framework contribution is the decomposition that makes both observations visible; “γ\\gamma\-LoRA wins style” is a narrower empirical claim than v1 implied, and we revise it accordingly\.

#### B\.5\.4What does*not*change

- •§4\.4 / §5 mechanism is descriptive of the Qwen3\-4B adapter weight distribution and is not re\-derived on Llama\. The band\-zero intervention \(§5\.7\) is a Qwen3\-4B claim\. Whether Llama\-3\.1\-8Bγ\\gamma\-LoRA adapters localize to the same band is open; the Phase K diff loop on the 50 Llama adapters would be a ~3 hour run and is left as future work\.
- •§4\.1\.4 blind\-preference and §4\.1\.6 human eval were not re\-run on Llama\. Both used Qwen\-conditioned generations\. We do not claim Llamaγ\\gamma\-LoRA generations would be human\-preferred at the same rate\.
- •§4\.2 calibration head architecture \(§3\.3\) is unchanged; only its effectiveness drops on Llama \(F1 0\.278 vs Qwen 0\.470\)\. Re\-tuning per base model is future work\.

Source\-of\-truth artefacts:runs/28\_writingprompts\_\_llama3\.1\-8b/gamma\_lora/aggregate\.json,runs/29\_f\_absence\_\_llama3\.1\-8b/aggregate\.json,docs/LLAMA\_REPLICATION\_PLAN\.md,experiments/\_base\_model\.py\.

### B\.6Routing baselines: full numbers \(Phase 2\)

For every \(persona, probe\) pair we have two candidate answers — one fromγ\\gamma\-LoRA \(with abstain prompt clause\), one from RAG \(top\-K BGE\) — plus the per\-token logits from each substrate\. A router maps the \(question, candidate\-answers, logits\) triple to a binaryuse\_lora∈\\in\{0,1\} decision\. The accepted answer is then scored against the gold \(the absence probe expects “I don’t know” or equivalent; the presence probe expects the recalled fact\)\. Eval is the held\-out 20\-persona slice \(240 rows\); train is the disjoint 30\-persona slice \(360 rows\)\.

We compare four routers: \(1\)hybrid\_logits\(ours, §4\.2 head\): small MLP over per\-substrate logit features \(top\-1 prob, entropy, margin\) trained to predictuse\_lora; \(2\)P\(True\)\[Kadavath’22\]: prompt the same base model “Is this answer correct? A\) True B\) False” and read softmax over A/B logits as confidence; route by argmax confidence; \(3\)Self\-consistency k=5\[Wang’23\]: generate 5 samples per substrate at T=0\.7, take majority vote per substrate, route by which substrate has higher within\-bucket agreement; \(4\)Adaptive\-RAG\-style classifier\[Jeong’24\]: train either TF\-IDF \+ logistic regression or DistilBERT\-base on the*question text alone*\(no logits, no candidate answers\) to predictuse\_lora\. The two LoRA\-only and RAG\-only floor rows are not routers — they are the substrate\-monolithic floors any router must beat to be useful\.

All F1 / CI / p\-value numbers below are computed on the held\-out evaluation split \(20 personae, 240 rows\) disjoint from the 30\-persona / 360\-row training split used to fithybrid\_logits, the Adaptive\-RAG classifier, and the route\-decision rule for self\-consistency\. Routing percentages reported below are also on this same eval slice\. Source\-of\-truth:runs/36\_calib\_head\_logits/ split\.jsonfor the persona partition,runs/aggregates/significance\_tests\.jsonfor all aggregate metrics\. 95% bootstrap CIs use 5k resamples over rows; paired permutation p\-values vs\.hybrid\_logitsare two\-sided with 5k permutations\.

Qwen3\-4B:

RouterF195% CIpres TPRabs TPRp vs hybridlora\_alone \(floor\)0\.105\[0\.048, 0\.158\]0\.530\.062e\-4rag\_alone \(floor\)0\.479\[0\.378, 0\.563\]0\.320\.980\.028hybrid\_logits\(ours\)0\.579\[0\.480, 0\.673\]0\.420\.95—P\(True\)0\.491\[0\.384, 0\.584\]0\.340\.880\.034Self\-consistency k=50\.479\[0\.377, 0\.577\]0\.320\.980\.023Adaptive\-RAG TF\-IDF0\.638\[0\.548, 0\.720\]0\.520\.830\.092Adaptive\-RAG DistilBERT0\.660\[0\.574, 0\.734\]0\.530\.870\.018Llama\-3\.1\-8B:

RouterF195% CIpres TPRabs TPRp vs hybridlora\_alone \(floor\)0\.063\[0\.000, 0\.135\]0\.660\.032e\-4rag\_alone \(floor\)0\.459\[0\.366, 0\.544\]0\.300\.970\.067hybrid\_logits\(ours\)0\.569\[0\.491, 0\.635\]0\.400\.98—P\(True\)0\.078\[0\.016, 0\.149\]0\.660\.042e\-4Self\-consistency k=50\.690\[0\.606, 0\.762\]0\.530\.970\.003Adaptive\-RAG TF\-IDF0\.709\[0\.629, 0\.773\]0\.620\.832e\-4Adaptive\-RAG DistilBERT0\.718\[0\.660, 0\.768\]0\.620\.864e\-4The text classifier wins on both models, decisively on Llama\.DistilBERT trained on question text alone — no logits, no candidate answers, no model\-internal signal — beatshybrid\_logitsby \+8\.1 F1 on Qwen \(p = 0\.018\) and\+14\.9 F1 on Llama\(p = 4e\-4\)\. The TF\-IDF baseline is within noise of DistilBERT on both models\. The implication: the routing decision is largely determined bysurface features of the question\(does it ask about a specific fact? does it ask about an unspecified attribute?\), not by the LLM’s internal confidence on its candidate answer\.

This is consistent with a simpler hypothesis we cannot reject:“is this a presence question or an absence question?” is the only signal that matters,and it is recoverable from question text with high accuracy\. The same TF\-IDF \(1–3\-gram \+ stop\-words\) classifier trained for the Adaptive\-RAG router achieveskind\-classifier eval accuracy 0\.850 on both Qwen and Llamafeatures, and the DistilBERT classifier reaches0\.892 on both\(source:runs/aggregates/adaptive\_rag\_baseline\.json\)\. Once the kind is known, routing is trivial — send presence to LoRA, absence to RAG\.

P\(True\) collapses on Llama\-3\.1\-8B\-Instruct\.Llama withγ\\gamma\-LoRA self\-reports near\-uniform high confidence on every question \(routing 99\.2% to LoRA on the eval slice — 238 of 240 rows\), driving F1 to the LoRA\-alone floor of 0\.078\. We attribute this to RLHF density: Llama\-3\.1\-Instruct’s heavy post\-training compresses the A/B logit margin on “is this answer correct?” prompts regardless of actual correctness\. On Qwen3\-4B P\(True\) is also ineffective but does not fully collapse \(F1 = 0\.491, ties RAG\-alone\)\.

Self\-consistency partially recovers on Llama\.SC k=5 reaches F1 = 0\.690 on Llama \(p = 0\.003 vs hybrid\), but only F1 = 0\.479 on Qwen \(ties RAG\-alone\)\. On Qwen the same recipe produces near\-deterministic samples on the eval slice \(5 of 240 rows pick LoRA outright, 179 pick RAG, 56 tie and are broken to RAG by the absence\-prior tie\-breaker\), collapsing the router to RAG\-alone behaviour\. SC’s effectiveness as a routing primitive depends on non\-trivial substrate sample diversity at T = 0\.7\.

Reading\.The negative finding sharpens the framework’s contribution\. We claim:

> *On the polymemory task we benchmark, substrate\-selection routing is a question\-classification problem disguised as a calibration problem\. The most accurate routers we found do not look at model internals at all; they classify the question\.*

Two consequences for the field: \(1\) Adaptive\-RAG\-style classifiers \[Jeong’24\] port directly to the per\-user setting — the corpus\-level “decide whether to retrieve” classifier retargets to per\-user “decide between LoRA and RAG” without modification, with our 50\-persona corpus as training data\. \(2\) Logit\-based confidence calibration is the wrong primitive for substrate selection\. P\(True\), self\-consistency, and our own logit\-feature head all underperform a 110M\-parameter text classifier\. Future routing work should start from the question side, not the answer side\.Scope\.This finding is onsingle\-turn factual probeswith a fixed presence/absence taxonomy\. We do not claim the question\-classification advantage holds in \(i\) multi\-turn dialogue, where the routing\-relevant signal is in conversation history rather than the latest question; \(ii\) long\-context settings, where the answer text dwarfs the question and a question\-only classifier loses information; or \(iii\) agentic / tool\-use pipelines, where user intent is implicit in tool selection rather than phrasing\. The 110M\-parameter DistilBERT classifier is a*baseline for showing logits are redundant*, not a finished deployment recipe — its ~100ms inference and external\-model dependency are real costs that a logit\-based router would not have\. We hold open that a richer logit head \(non\-RLHF base, mixture\-of\-experts gating, or distilling the question\-kind signal back into the substrate as a learned gating head over the LoRA delta\) could recover\. The negative result is on the four well\-known recipes above for single\-turn factual routing, not on the routing problem in general; the burden of proof now rests on logit\-based methods to show they add what a text classifier cannot\.

### B\.7LaMP\-3 mitigation sweep \(Phase M, n = 50 personae per arm\)

Body summary in §4\.3\.6\. Full per\-arm specifications, pairwise paired tests, Llama replication numbers, and pointers to per\-record JSONs below\.

#### B\.7\.1Arm specifications

All arms train on the same 50\-user LaMP\-3 held\-out subset \(runs/F\_lamp3/dataset\.jsonheld\_outsplit, 4 train pairs per user, 4 main \+ 4 probe2 eval queries per user\)\. Base model Qwen3\-4B unless noted; eval via Bedrock Claude judge with strict integer\-1\-5 scoring onmainand held\-out review\-rating prediction onprobe2\. Per\-arm runtime: ~5–10 GPU\-h on L40S except where noted\.

ArmWhat changed \(vs §4\.3 baseline\)Hypothesis layer \(§4\.3\.5\)Hlogit mask \{1\.\.5\} at eval; training unchanged \(r=128,α\\alpha=256, ep=20\)\(ii\) format onlyBepochs 20→\\rightarrow3\(i\)/\(ii\) less overfitArank 128→\\rightarrow8,α\\alpha256→\\rightarrow16\(i\) capacity\-driven memorizationFLoRA→\\rightarrowIA³ \(~10K params\)\(i\) data poverty testGrating\-token loss weight 10×\\timesduring training\(ii\) format\-aware lossCtarget\_modules q\+v only, layers 24–31\(iii\) protect o\_projEmix LaMP\-3 with 50 Alpaca pairs/user\(ii\) instruction anchorD\+ lambda\*KL\(studentIClaude\-paraphrase aug 5→\\rightarrow30 train pairs/user, ep=20\(i\) data scale§4\.3 baseline \(no mitigation, r=128α\\alpha=256 ep=20, free\-form decode\): main 0\.315, probe2 0\.410\.

#### B\.7\.2Per\-arm aggregate results \(n = 50\)

Armmain\_accprobe2\_accΔ\\Deltaprobe2 vs Hpaired t p \(vs H\)Wilcoxon p \(vs H\)H1\.0000\.600———B0\.7200\.630\+3\.00\.4200\.385A1\.0000\.635\+3\.50\.3010\.326F1\.0000\.645\+4\.50\.1920\.234G1\.0000\.610\+1\.00\.7190\.783C1\.0000\.635\+3\.50\.3410\.275E1\.0000\.615\+1\.50\.6850\.533D0\.9950\.655\+5\.50\.0550\.053I1\.0000\.635\+3\.50\.2410\.294\(p\-values from per\-persona paired tests, n = 50, comparing each arm’s probe2 vs arm H’s probe2 on the same persona\. No Bonferroni correction; if one is applied across the 8 comparisons the D\-vs\-H p inflates well past 0\.05\. We report the arm\-level numbers and let the reader apply their preferred MCC\.\)

#### B\.7\.3Cross\-product: every training arm×\\timeseval\-time logit mask \(Phase 3\)

The body table in §4\.3\.6 shows what happens when the \{1\.\.5\} logit mask is applied on top of*each*training\-time recipe \(n = 50 personae per cell, identical pipeline, onlyeval\_decoding = constrained\_15flipped on\)\. We list it here at full precision together with theΔ\\Deltavs free decoding for each arm:

Armfamilyrecipemain \(free\)probe2 \(free\)main \(\+mask\)probe2 \(\+mask\)Δ\\DeltamainΔ\\Deltaprobe2Heval\-timelogit mask \{1\.\.5\} only0\.3150\.4101\.0000\.600\+0\.685\+0\.190Bscheduleep=3 \(vs ep=20\)0\.7200\.6300\.7450\.635\+0\.025\+0\.005Aarchitecturer=8α\\alpha=161\.0000\.6351\.0000\.640\+0\.000\+0\.005FarchitectureIA³ adapter1\.0000\.6451\.0000\.645\+0\.000\+0\.000Carchitectureq\+v only, L24–311\.0000\.6351\.0000\.630\+0\.000−\-0\.005Glossrating\-token loss 10×\\times1\.0000\.6101\.0000\.605\+0\.000−\-0\.005DlossKL anchor \(lambda=0\.1\)0\.9950\.6551\.0000\.660\+0\.005\+0\.005Elossmixed Alpaca anchor1\.0000\.6151\.0000\.640\+0\.000\+0\.025IdataClaude paraphrase aug1\.0000\.6351\.0000\.615\+0\.000−\-0\.020Arms are grouped by mitigation family \(eval\-time intervention, training schedule, architecture, loss, data augmentation\) — the four families predicted in §4\.3\.5 plus the H eval\-time control\.

Two robustness observations matter for the §4\.3\.6 claim\. First, the eval\-time mask drives every recipe to main\_acc≥\\geq0\.995 \(mean 0\.998\), and the≥\\geq0\.995 floor is independent of which training\-time intervention preceded it — i\.e\.no training recipe escapes the collapse without the mask, and the mask saturates main on every recipe\. The mask is doing all of the recoverable work\. Second, the masked probe2 column stays inside \[0\.605, 0\.660\] \(σ\\sigma= 0\.018, range 0\.055\), bracketing the free\-decoding probe2 column that itself sits in \[0\.610, 0\.655\]\. Neither training nor eval\-time intervention moves probe2 outside this band on n = 50; the residual signal is a property of LaMP\-3 task structure \(rating from 4 in\-context preference pairs\), not of recipe or decoding choice\.

Source\-of\-truth artifacts:runs/40\_lamp3\_mit/H\_on\_\{B,A,F,G,C,E,D,I\}/aggregate\.json\(cross\-product\) andruns/40\_lamp3\_mit/\{B,A,F,G,C,E,D,I,H\}/aggregate\.json\(free decoding\); per\-persona evals inLAMP3\_U\_\*/eval\.jsonunder each arm dir\.

#### B\.7\.4Llama\-3\.1\-8B\-Instruct replication \(champion D\)

Re\-running arm D \(KL anchor\) on Llama\-3\.1\-8B\-Instruct, identical recipe and prompts, same 50 LaMP\-3 held\-out personae:

MetricQwen3\-4B DLlama\-3\.1\-8B Dmain\_acc0\.9951\.000probe2\_acc0\.6550\.655n\_personae5050Per\-persona paired test \(Llama D probe2 vs Qwen D probe2, same persona\): t\.p = 1\.000 \(means coincide\); per\-persona Pearson r = 0\.78, p < 0\.0001\. The same personae score high under both base models, and the rank\-ordering of the residual probe2 signal is preserved\. We read this as evidence that the residual signal is a property of LaMP\-3 task structure \(rating preferences from 4 in\-context training pairs\), not of either tokenizer or RLHF stack\.

#### B\.7\.5Operational notes

The mitigation pipeline ran as a 9\-arm sequential cron on a single L40S \(g6e\.xlarge\), state\-machined throughruns/40\_lamp3\_mit/state\.jsonwith each tick advancing one arm via committed shell scripts \(scripts/lamp3\_mit\_tick\.sh,scripts/launch\_lamp3\_mit\.sh\)\. Total wallclock: ~6\.5 GPU\-h for the Qwen sweep \+ 0\.5 GPU\-h for the Llama champion replication\. Two runtime issues surfaced and were resolved without changing the experimental design:

1. 1\.Tokenizer mismatch\(commita793143\)\. The runner initially loaded the active model via\_base\_model\.active\(\)\.pathbut tokenized with a hardcoded Qwen path; on the Llama replication this produced out\-of\-vocab token IDs and a CUDA device\-side assert\. Fixed by routing the tokenizer through the same\_base\_model\.active\(\)helper as the model\.
2. 2\.Disk pressure\. The 9 Qwen arms×\\times50 personae×\\times361 MB LoRA adapter≈\\approx144 GB, which exhausted the 291 GB root volume during the Llama replication\. We deleted per\-persona LoRA adapters after each arm’saggregate\.jsonwas committed \(per\-recordeval\.jsonretained as source of truth\)\. The recipe of “drop trained adapters once their eval JSON is persisted” is now a standing operational rule for this experiment family\.

#### B\.7\.6Files

Per\-persona evals:runs/40\_lamp3\_mit/\{H,B,A,F,G,C,E,D,I\}/LAMP3\_U\_\*/eval\.json\(50 per arm, 450 total\)\. Aggregates: same dirs,aggregate\.json\. Llama replication:runs/40\_lamp3\_mit/REPLICATE\_LLAMA\_D/\. State machine:runs/40\_lamp3\_mit/state\.json\(preserves arm\-by\-arm launch/POLL/DOC transitions for replicability\)\. Plan doc:docs/LAMP3\_MITIGATION\_PLAN\.md\.

### B\.8Summary of headline claims

ClaimEvidenceStrengthParametric adaptation is*necessary*for style transfer; retrieval is structurally insufficient§4\.1 RAG≈\\approxB\_nohist \(\+0\.060\);γ\\gamma\-LoRA \+0\.473; 50/50 personae and 240/250 records preferγ\\gamma\-LoRA; blind\-pref macro 59\.8% \[56\.5, 63\.1\]; human\-eval n=30 79\.3% p=0\.0023Strong on Qwen3\-4B, does NOT replicate on Llama\-3\.1\-8B\(§4\.6:Δ\\Delta= \+0\.003 nat/tok, CI \[−\-0\.016, \+0\.021\]\)\. Restated as abase\-model\-and\-recipe\-specificclaim\.RAG andγ\\gamma\-LoRA fail in*opposite directions*on absence calibration; the substrate that wins on §4\.1 loses on §4\.2§4\.2 absence TPR: RAG 99% vsγ\\gamma\-LoRA 8\.7% \(top\-line\); substrate\-controlled C\_rag vs C\_lora\_calib \(both with abstain\): \+45\.7pp on Qwen,\+79\.0pp on Llama; F1 0\.521 vs 0\.150Strong AND replicates across two base models\(§4\.6: Llama RAG 96\.7% vsγ\\gamma\-LoRA 3\.0%, top\-line \+93\.7pp, substrate\-controlled \+79\.0pp —*larger*than Qwen\)\. The calibration asymmetry is a substrate property, not a Qwen artefact\.Synthetic\-to\-real transfer of per\-userγ\\gamma\-LoRA does not hold; on LaMP\-3,γ\\gamma\-LoRA underperforms a constant predictor;a measurable instruction\-following collapse accounts for part but not all of the gap§4\.3 main 31\.5% vs majority 59\.5%; strict\-format failure 20\.5% \(§4\.3\.3\); loose\-parsed conditional acc 34\.2%, still 25pp below majorityStrong as a measurement; the failure mode decomposition \(§4\.3\.3 instruction\-following \+ §4\.3\.4 substrate\-asymmetry amplification\) is novel and replicable\.γ\\gamma\-LoRA’s largest weight\-deltas concentrate in mid\-to\-late attentionq\_proj\(layers 21–35\); the same band’s per\-persona Frobenius mass predicts opposite\-direction behavioural outcomes, and zeroing the band at inference time breaks both probes§4\.4 top\-10 cells across n = 50 personae; cv 0\.08–0\.12; top\-1 1\.65×\\timesbackground\.§5\.6 per\-persona correlation: top\-band mass×\\timesprobe2 acc r = \+0\.41 \(p = 0\.0017\); top\-band mass×\\timesabsence\-TPR r =−\-0\.49 \(p = 0\.0001\)\.§5\.7 band\-zero intervention at n = 50: absence\-TPR \+33\.0pp, presence\-TPR−\-19\.7pp\. §5\.7 specificity: layer\-control L5–19 q\_projΔ\\Delta\_abs =−\-3\.0pp \(clean\); projection\-control L21–35 v\_projΔ\\Delta\_abs = \+18\.7pp \(partial\)Causal at the band level on Qwen3\-4B; layer specificity clean, projection specificity partial\. Whether Llama\-3\.1\-8Bγ\\gamma\-LoRA adapters localize to the same band isopen\(§4\.6\.4 future work\)\.Cross\-model replication: the framework is what travels§4\.6 Llama\-3\.1\-8B re\-run of §4\.1 \+ §4\.2 with identical recipe / corpus / prompts: calibration asymmetry replicates and strengthens \(substrate\-controlled gap \+45\.7pp Qwen→\\rightarrow\+79\.0pp Llama\), behavioural\-style advantage does not replicate \(Δ\\Delta= \+0\.003 nat/tok, CI \[−\-0\.016, \+0\.021\]\)The framework — three\-axis decomposition with per\-substrate per\-axis probes — surfaces both the universal failure \(calibration\) and the base\-model\-specific advantage \(style\) without bias toward either\. The diagnostic decomposition is what travels; the magnitudes do not\.Substrate\-selection routing is a question\-classification problem, not a calibration problem§B\.7 four\-way routing benchmark on both models: a DistilBERT classifier on question text alone beats logit\-feature head by \+8\.1 F1 on Qwen \(p=0\.018\) and \+14\.9 F1 on Llama \(p=4e\-4\); P\(True\) collapses on Llama \(F1=0\.078, routes 99\.2% to LoRA\); self\-consistency works on Llama \(F1=0\.690\) but not Qwen \(F1=0\.479\)Strong negative finding, replicates on both models\.All three logit\-based routers \(logit\-feature head, P\(True\), self\-consistency\) lose to a 110M\-param text classifier with no access to model internals\. The routing signal is recoverable from question surface features alone \(kind\-classifier accuracy 0\.85 / 0\.89 on both models\)\.Together, these four sit naturally inside a*substrate–task*framing: behavioral knowledge is distributional and gradient\-friendly; factual presence/absence is symbolic and retrieval\-friendly; real\-world per\-user data has noise structure that breaks the synthetic\-benchmark recipe; and the asymmetry is not just behavioral but mechanistic — the cells that store style preferences are the same cells that confabulate absent facts\. §5 develops this framing and its implications for hybrid architectures\.

## Appendix CAppendix C: Mechanism \(full detail\)

The behavioral results in §4 establish thatγ\\gamma\-LoRA encodes persona\-specific style and surface\-form regularities \(§4\.1\) but fails on absence calibration \(§4\.2\)\. A natural follow\-up question isstructural: where in the model do those persona\-specific weight changes actually land? If the answer is “diffusely, all over the network,” the substrate has no mechanistic story to tell beyond “fine\-tuning works\.” If instead the answer is “in a consistent, narrow band of layers and projections,” then the substrate asymmetry of §4 is grounded in something more specific than capacity, and future hybrids have a concrete target for where to intervene\.

This section reports the answer at n = 50 personae using the weight\-delta decomposition described in §3\.5\.

### C\.1Setup

For each of the 50 held\-out personae we trained oneγ\\gamma\-LoRA adapter from scratch on the persona’s synth\-QA pairs \(rank 128,α\\alpha= 256, 20 epochs\), saved the resulting adapter viasave\_pretrained, and computed the per\-cell Frobenius norm∥\\\|W\_lora∥\\\|\_F =∥\\\|α\\alpha/r⋅\\cdotB A∥\\\|\_Ffor each \(layer, projection\) pair across the four attention target projections \(q\_proj,k\_proj,v\_proj,o\_proj\) and all 36 transformer layers\. Across\-persona aggregates are means and coefficients of variation of these per\-cell Frobenius norms over the 50 adapters\. We refer to the distribution of these 144 cells \(4 projections×\\times36 layers\) as theadapter mass distribution\. The all\-cell mean \(0\.478\) serves as a structural background rate; cells well above this background are localization candidates\.

### C\.2Localization is real and narrow

![Refer to caption](https://arxiv.org/html/2606.11712v1/x3.png)Figure 3:Mechanism heatmap\. Mean Frobenius mass ofγ\\gamma\-LoRA ABˆT weight delta across 4 attention projections×\\times36 layers, averaged over n=50 personae\. q\_proj layers 21–35 light up coherently; o\_proj is uniformly low; k/v\_proj show secondary structure\.Top\-10 cells by mean Frobenius across 50 personae:

ranklayerprojmean Frobstdcv135q\_proj0\.7870\.0830\.105222q\_proj0\.7730\.0920\.119324q\_proj0\.7370\.0780\.106423q\_proj0\.7270\.0630\.086529q\_proj0\.7240\.0670\.092634q\_proj0\.7230\.0580\.080727q\_proj0\.7180\.0630\.087835o\_proj0\.7180\.0730\.102930q\_proj0\.7140\.0630\.0891021q\_proj0\.7100\.0650\.092Three observations\.

1. 1\.Mid\-to\-late attentionq\_projdominates\.Nine of the top ten cells areq\_projin layers 21–35 \(mid\-to\-late stack on a 36\-layer model\)\. The tenth iso\_projat L35\.
2. 2\.The pattern is shared across personae, not idiosyncratic\.Coefficients of variation for the top band are 0\.08–0\.12\. If each persona were idiosyncratically reusing different layers, we would expect cv≥\\geq0\.5\. The narrow cv says: the same band of layers carries the persona signal across nearly all 50 adapters\.
3. 3\.The top cell is 1\.65×\\timesthe structural background\.The all\- cell mean Frobenius is 0\.478; the top\-1 cell \(L35q\_proj\) is 0\.787\. The full top\-10 band sits at 0\.71–0\.79, ~1\.5×\\timesover background\. Concentration is real but not extreme — this is localized, not “one layer does everything\.”

### C\.3Per\-projection summary

Aggregating across all 36 layers within each projection:

projmean Frobrel\-to\-qq\_proj0\.6681\.00o\_proj0\.5580\.84v\_proj0\.3540\.53k\_proj0\.3310\.50q\_projcarries roughly twice the mass ofv\_proj/k\_proj, witho\_projan intermediate secondary mode\. This rules out the naive “LoRA writes equally to all four attention projections” null and is consistent withq\_projcontrollingwhat the attention head looks forwhilev\_proj/k\_projcontrol content / addressing — the persona signal is more about altering queries than about adding new keys or values\.

### C\.4Refinement vs\. the n = 3 pilot

An earlier n = 3 pilot reported a top\-3 of \(L35q\_proj, L22q\_proj, L30o\_proj\) at n = 3\. At n = 50:

n = 3 pilot top\-3Present in n = 50 top\-10?L35q\_proj✓ \(rank 1\)L22q\_proj✓ \(rank 2\)L30o\_proj✗ \(rank 31; replaced by a band of L21–29q\_proj\)The first two cells survive cleanly\. L30o\_projdoes not — it is replaced by abroader band of mid\-stackq\_projlayers\(L21, L23, L24, L27, L29\) that was not visible at n = 3\. The qualitative story strengthens \(mid\-to\-late attention\-q dominance with ano\_projsecondary mode at L35\); the specific “L30o\_projis top\-3” claim was an n = 3 artifact and is retracted\. Pilot results do not survive scaling cleanly, and we report only n = 50 numbers as primary\.

### C\.5Interpretation

These three pieces — narrow band, low cv, q\-projection bias — are consistent with the picture thatγ\\gamma\-LoRA’s persona signal acts onmid\-to\-late attention queries, modulating which context tokens the head pulls from rather than rewriting the content of those tokens\. This is structurally different from how a behavioral memory should work if it were pretending to be a fact store: a fact store would more plausibly land inv\_proj\(content\) or in FFN\-down \(memory cells, per the Geva et al\. line of work\(Geva et al\.,[2021](https://arxiv.org/html/2606.11712#bib.bib9)\)\)\. The fact thatγ\\gamma\-LoRA behaves like attention\-routing rather than fact\-writing — combined with §4\.2 showing it cannot abstain on absent facts — is mutually reinforcing evidence for the substrate\-asymmetry framing of §6\.

### C\.6Per\-persona correlation: same band, opposite\-direction effects

![Refer to caption](https://arxiv.org/html/2606.11712v1/x4.png)Figure 4:Per\-persona scatter\. Top\-band q\_proj Frobenius mass \(L21–35\) versus behavioral\-style win on the x\-axis and absence\-TPR on the y\-axis \(n=50\)\. Same band correlates*positively*with style and*negatively*with absence\-calibration\.The localization claim above is descriptive — it says*where*γ\\gamma\-LoRA writes, not*whether*writing there does anything\. To check the latter, we compute per\-persona top\-band Frobenius mass \(sum over L21–35q\_proj\) for all 50 personae and correlate it against the two §4 outcome streams that have per\-persona resolution: probe2 factual recall \(§B\.1, §B\.2\) and absence\-TPR under the calibration prompt \(§B\.2\)\. All values are computed from saved adapters and per\-persona eval JSONs released with the supplementary materials\.

The result is the strongest single piece of evidence we have for the substrate\-asymmetry framing\. Across n = 50 paired personae:

top\-band mass×\\timesoutcomePearson rSpearman rp \(two\-sided\)L21–35q\_proj×\\timesprobe2 factual acc\+0\.41\+0\.370\.0017L21–35q\_proj×\\timesmain acc\+0\.30\+0\.210\.0284L21–35q\_proj×\\timesabsence\-TPR−\-0\.49−\-0\.280\.0001L35q\_proj\(top\-1\)×\\timesabsence\-TPR−\-0\.42−\-0\.290\.0015The same cells whose mass predicts*better*factual recall \(r = \+0\.41\) predict*worse*absence calibration \(r =−\-0\.49\)\. Personae whose adapters write more aggressively into the L21–35q\_projband recall present facts more accurately*and*confabulate plausible answers to absent\-fact probes more confidently\. This is the substrate asymmetry observed in §4\.2 showing up*one level deeper*: it is not just thatγ\\gamma\-LoRA as a whole substrate cannot abstain — it is that the very mechanism which makesγ\\gamma\-LoRA recall facts well is also what makes it abstain poorly\. The correlational reading at the persona level is backed by a direct causal test: §5\.7 reports that zeroing the L21–35q\_projband at inference time breaks absence calibration by 33pp and presence\-TPR by 20pp at n = 50, confirming the band’s causal role on both axes\. The directionality is unambiguous and the effect sizes are large for n = 50 \(\|r\|≥\\geq0\.4 with p≤\\leq0\.002 on three of the four load\-bearing pairs, and the intervention recovers the predicted directionality\)\.

Two negative controls reinforce the reading\. \(i\) Top\-band mass does*not*significantly predictfinal\_loss\(r = \+0\.22, p = 0\.11\) orsanity\_acc\(r =−\-0\.24, p = 0\.09\), so the correlation is not driven by “this persona’s adapter just trained better\.” \(ii\) The narrowed band \(L35q\_projalone, top\-1 cell\) gives a smaller correlation than the full L21–35 band \(\|r\| = 0\.42 vs 0\.49 for absence\-TPR\), consistent with the band being structurally meaningful rather than the top\-1 cell being a single\-cell artifact\.

We do not claim this correlation generalizes outside this LoRA recipe\. The intervention \(§5\.7\) confirms it survives test\-time zeroing on this recipe; whether the same band carries the same roles on other recipes is a natural follow\-up\. What §5\.6 \+ §5\.7 together establish is that the n = 50 layer\-localization story in §§5\.1–5\.5 is not just descriptive geometry: the cells that move predict the behavior the paper measures, in opposite directions on the two probes, and removing them at inference time breaks both — exactly what a substrate\-asymmetry account predicts\.

### C\.7What this section does*not*claim

- •It doesnotclaim FFN projections are unimportant\. The current diff loop only collected attentionq/k/v/o; FFNgate/up/downare not measured at n = 50\. The n = 3 pilot showed FFN\-down to be small, but that result is too underpowered to repeat as a positive claim\. We report attention\-only and flag the FFN ratio as future work \(§6\.4 / camera\-ready\)\.
- •It doesnotclaim a*strictly causal*link between layer mass and behavioral outcomes at the individual\-head or individual\-layer level\. §5\.6 establishes a per\-persona correlation across n = 50 with \|r\| up to 0\.49 and p≤\\leq0\.0001, and §5\.7 establishes that band\-level zeroing breaks both probes \(\+33pp absence\-TPR,−\-20pp presence\-TPR at n = 50\)\. Isolating which heads or layers within the L21–35 band carry which axis is left for future work\.
- •It doesnotclaim layer\-localization is unique toγ\\gamma\-LoRA on this substrate\. Other LoRA recipes \(different rank, different target modules, different data\) may localize elsewhere; we report what*this*recipe does, which is what the rest of the paper is built on\.

### C\.8What scaling did to our mechanism story: retracted, provisional, and survived claims

The mechanism story above is the n = 50 story\. An earlier n = 3 pilot made narrower, more confident claims about specific cells; some survived scaling and some did not\. We surface the delta here as epistemic discipline — pilots can fail at scale, and a paper that hides what didn’t survive its own follow\-up scaling run is being economical with the truth\.

n = 3 pilot claimn = 50 statusVerdictL30o\_projis in the top\-3 cells byΔ\\DeltaW Frobenius mass\.At n = 50, L30o\_projfalls to rank 31; the top\-3 is dominated by L35 / L22 / L21–29q\_projcells \(§5\.2, §5\.4\)\.Retracted\.The single\-cell claim was an n = 3 artifact\. The qualitative story it pointed at \(mid\-to\-late attention dominance\) survives in a broader form\.FFN\-downΔ\\DeltaW mass is small relative to attention\.Not measured at n = 50 — the n = 50 diff loop collected attentionq/k/v/oonly; FFNgate/up/downare absent from the tensor we analyze\.Provisional, flagged\.We do not repeat the n=3 pilot’s FFN\-vs\-attention claim as a positive finding; §5\.7 already calls this out and §6\.4 lists it as the first follow\-up\.Layer\-localization is narrow \(signal concentrates in a thin band of mid\-to\-late layers\)\.Confirmed at n = 50 with cv 0\.08–0\.12 across L21–35q\_projand a per\-persona correlation ofrTwo of three n=3\-pilot claims did not survive cleanly: one cell\-level claim isretractedoutright, one projection\-class claim is flagged as provisional pending FFN measurements\. Only the band\-level localization claim — the one we now build §§5\.6–5\.7 on — scaled\. We report the n = 50 numbers as primary throughout the paper and treat the n = 3 pilot as exploratory only\.

### C\.9Files

- •Aggregate Frobenius summary — top\-10 \+ per\-projection aggregates
- •Per\-personaΔ\\DeltaW Frobenius matrices
- •Adapter retraining log
- •Diff/aggregate log
- •experiments/30\_design\.md,experiments/30\_results\.md— design \+ decision

## Appendix DAppendix D: Discussion \(full\)

### D\.1What the asymmetry says about user\-side memory

Across our three probes, the substrate that wins one probe loses the next:

ProbeBest substrateWorst substrateBehavioral / style \(§4\.1\)γ\\gamma\-LoRA \(Δ\\Delta\+0\.413 nat/tok over RAG; blind\-pref macro 59\.8% \[56\.5, 63\.1\], lexical 64\.6%\)RAG≈\\approxno\-history baselineFactual presence \(§4\.2 implicit, §B\.2 C\_rag F1 = 0\.521\)RAGγ\\gamma\-LoRAFactual absence \(§4\.2\)RAG \(TPR 99%\)γ\\gamma\-LoRA \(TPR 8\.7%\)Real\-data factual \(§4\.3\)Majority baseline \(59\.5%\)γ\\gamma\-LoRA \(31\.5%\)Two substrates, two failure modes, no Pareto winner\. The natural reading is that*user\-side memory*is not one capability that one substrate either has or lacks — it is a small basis of capabilities \(style, presence, absence, factual transfer\) that load differently onto the same machinery\.γ\\gamma\-LoRA writes user\-specific adaptations into model parameters and recovers behavioral consistency for free, because behavior is a distributional property the optimizer can lean on\. The same machinery has nowhere to put the bit*“I have not been told this fact about user u”*— that bit is symbolic, sparsely attested, and the gradient signal for it during per\-user fine\-tuning is dwarfed by the signal for the things that*are*in the training pairs\. RAG inverts both: it cannot push behavioral mass into the model’s distribution at all, but its top\-K \+ abstain prompt produces a clean discrete signal — return the chunk, or refuse — that maps exactly onto presence/absence calibration\.

### D\.2Implication: hybrid architectures, not silver bullets

If the asymmetry is real \(and our pre\-registered structural falsifiers on §4\.1 and §4\.2 both held in the predicted direction, with the caveats noted\), the recommendation for a production user\-side memory system is a*hybrid*:

1. 1\.A parametric component \(γ\\gamma\-LoRA\-like, but probably notγ\\gamma\-LoRA exactly — see §5\.4\) for behavioral consistency\.
2. 2\.A retrieval component over a per\-user corpus for factual presence\.
3. 3\.A calibration head — even a logistic regressor over retrieval confidence, as in our §4\.2 C\_lora\_calib variant — to recover the absence signal that the parametric component throws away\.

The first two are not new; the contribution of this paper is the*decomposition*that makes them defensible to combine, plus the falsifier shape that lets future work argue for or against specific weightings\.

### D\.3Why §4\.3 \(real\-dataγ\\gamma\-LoRA failure\) is a feature, not a bug

The LaMP\-3 result \(γ\\gamma\-LoRA at 31\.5% vs majority at 59\.5%\) is the strongest single evidence in the paper that synthetic personalization benchmarks are not load\-bearing for real\-world claims\. We were prepared to publish a positive synthetic\-only result; the LaMP\-3 transfer probe — which we ran*before*claiming a top\-tier\-venue result — caught a regime gap we would otherwise have missed\.

Two specific ways the synthetic corpus over\-credits parametric substrates:

- •Volume\.V3 givesγ\\gamma\-LoRA 12 high\-quality probe pairs per persona; LaMP\-3 gives it 8 unstructured reviews\. At small fine\- tuning budgets, parametric memory is data\-hungry in a way RAG is not\.
- •Distribution shift\.V3’s probes are paraphrases of the backstory; LaMP\-3’s held\-out review is a*different*product, so the transfer is from review\-style to rating\-prediction\.γ\\gamma\-LoRA trained on review*content*doesn’t bind to rating\-distribution\.

Neither observation is novel\. What is novel is the framing: a diagnostic paper that*expects*this gap and reports the two benchmarks together, instead of choosing the one that flatters the substrate\.

### D\.4Limitations \(honest\)

We do not claim:

- •Thatγ\\gamma\-LoRA is the right parametric primitive for behavioral memory\. It is the simplest one we had end\-to\-end working; rank\-16 on q/k/v/o is a baseline, not an optimum\. Prefix\-tuning, soft prompts, and adapter\-stacking are unevaluated\.
- •That synthetic personae transfer to humans\. They do not, on LaMP\-3 \(§4\.3\)\. They likely also do not on user\-study data we did not run\. Treat V3 as a*clean instrument*, not a population\.
- •That n=50 personae or n=50 LaMP\-3 users is enough for variance arguments\. Our confidence intervals are bootstrap over records, not over personae; per\-persona variance is reported but not used as a falsifier\.
- •That the §4\.4/§5 mechanism story is*strictly*causal at the individual\-head level\. At n=50 \(see §5\) the layer\-localization pattern is stable — q\_proj dominates layers 21–35 with mean Frobenius 0\.668 vs 0\.331 for k\_proj, top\-1 cell L35 q\_proj at 1\.65×\\timesthe all\-cell mean, across\-persona cv 0\.08–0\.12\. §5\.6 adds a per\-persona correlation arm: top\-band Frobenius mass correlates with probe2 factual recall at Pearson r = \+0\.41 \(p = 0\.0017\)andwith absence\-TPR at r =−\-0\.49 \(p = 0\.0001\), with both negative controls \(final\_loss, sanity\_acc\) failing to reach significance\. §5\.7 then runs the band\-zeroing intervention at n=50 and confirms the band’s causal role on both axes \(\+33pp absence\-TPR,−\-20pp presence\-TPR\)\. The mechanism is causal at the band level; isolating which heads or layers within L21–35 carry which axis is future work\. The original Phase\-G n=3 finding “L30 o\_proj is top\-3” did not survive the n=50 lift — documented in §5\.4 — which is itself evidence that the n=3 numbers alone would have been unreliable\.
- •That the absence\-TPR gap is solely a substrate fact\. RAG’s 99% TPR uses an explicit abstain prompt;γ\\gamma\-LoRA’s 8\.7% does not\. The C\_lora\_calib variant \(F1 0\.470 with a calibration head\) shows ~half the gap is recoverable architecturally — i\.e\., part of what looks like substrate asymmetry is prompt\-engineering asymmetry\. We call this out in §4\.2; it does not undermine the structural claim, but it bounds the headline number\.

### D\.5Future work

Three next\-steps named earlier in the project \(blind\-preference judging, mechanism at n=50, per\-persona behavior correlation\) and a fourth \(band\-zeroing causal arm\) have all*landed*; their results are folded into §4\.1\.4, §4\.4, §5\.6, and §5\.7 respectively\. What remains:

1. 1\.Real\-dataγ\\gamma\-LoRA failure analysis\.§4\.3\.3’s instruction\- following collapse \(20\.5% strict\-format failures, ~8% true off\-topic continuations\) is a strong hypothesis\-source: is the failure rating\-distribution shift, chat\-template breakage from per\-user SFT, or signal\-to\-noise on 8 reviews per user? We have the predictions; a one\-day annotation pass would resolve it\.
2. 2\.Within\-band head/layer isolation\.§5\.7 establishes the L21–35 q\_proj band is causally load\-bearing on both probes\. Which individual layers or attention heads within the band carry presence vs absence is open; a head\-level zero\-ablation or path\-patching analysis is the natural follow\-up\.

A third, more speculative direction — parametric*and*retrieval substrates trained jointly with a routing head — is what we believe the substrate asymmetry actually motivates, but we do not run it here; it is the natural follow\-up paper\.

### D\.6Causal: zeroing the L21–35 q\_proj band breaks calibration

![Refer to caption](https://arxiv.org/html/2606.11712v1/x5.png)Figure 5:Band\-zero intervention\. Paired before/after bars at n=50 personae for absence\-TPR \(left\) and presence\-TPR \(right\)\. Zeroingγ\\gamma\-LoRA L21–35 q\_proj weights raises absence\-TPR by \+33pp and lowers presence\-TPR by 20pp, causally implicating the same band in opposite\-direction effects on the two probes\.§5 reports a per\-persona Pearson correlation of r=\+0\.41 between top\-band Frobenius mass on layers 21–35 q\_proj and absence\-TPR, and r=−\-0\.49 between the same band and presence\-TPR \(n=50, p<0\.002 and p<0\.0001\)\. The correlation arm is suggestive but not interventional — a confound where “personae whose adapters happen to train better” explain both signals cannot be ruled out from correlation alone\.

We close the loop with an interventional run\. For each of 50 V3 held\-out personae \(101–150\) we load the savedγ\\gamma\-LoRA adapter in PEFT format, zero thelora\_Aweights forq\_proj on layers 21–35, and re\-run the same absence and factual\-recall probes \(6 probes per persona, judged identically to §4\) with the band\-zeroed adapter substituted in place of the trained one\. All other layers, all other projections, and the base model are unchanged\.

Result \(n=50 V3 personae, 6 probes each\):

metricfull adapterband\-zeroedΔ\\Deltaabsence TPR0\.4030\.733\+33\.0pppresence TPR0\.4200\.223−\-19\.7ppZeroing exactly the band the §5 correlation arm fingered moves both metrics simultaneously and in opposite directions: absence\-calibration*improves*by 33pp, factual recall*degrades*by 20pp\. This is the signature of a single localized mechanism that trades calibrated abstention against parametric recall — not two independent effects\.

Per\-persona heterogeneity\.The mean conceals structure worth reporting\. Of the 50 personae, 16 showΔ\\Delta\_abs≥\\geq\+50pp \(strong intervention effect, e\.g\. V3\_P\_101 \+83\.3, V3\_P\_144 \+100\.0, V3\_P\_147 \+100\.0\); the same personae show large negativeΔ\\Delta\_pres \(V3\_P\_109−\-100\.0, V3\_P\_111−\-83\.3, V3\_P\_112−\-83\.3\)\. At the other end, ~10 personae showΔ\\Delta≈\\approx0 on both axes — these are at ceiling \(full\_abs=1\.0\) or floor \(full\_pres=0\.0\) where the metric cannot move\. The full 50\-row table is in appendix A\.5; the ceiling/floor\-filtered mean \(excluding personae with full\_abs≥\\geq0\.95 or full\_pres≤\\leq0\.05\) preserves the qualitative pattern\.

What this rules out\.A non\-mechanism reading would say “any ablation of any band would degrade calibration somehow\.” Two facts make that unlikely\. First, the direction matters: zeroing this band*improves*absence\-TPR; it does not degrade it\. A generic damage account predicts both metrics drop\. Second, the band was identified purely from the correlation arm of §5 \(top\-magnitude Frobenius mass on q\_proj layers 21–35\), with no access to the intervention data\. The arrow points from the per\-persona correlation to a per\-band intervention, not the other way\.

We treat this as a causal claim at the band granularity: zeroing L21–35 q\_proj breaks the calibrated\-abstention behaviorγ\\gamma\-LoRA acquires during training, and does so by removing the same parameters whose magnitude correlates with the behavior pre\-intervention\. We do not claim mechanism at finer granularity \(which heads, which features\) — we leave this finer\-grained mechanism analysis to future work\.

Specificity controls\.To check that the effect is not a generic consequence of zeroing 15 contiguous layers oflora\_A\(any band would do\) or of perturbing q\_proj broadly \(any layer range would do\), we run two matched controls under the identical pipeline:

InterventionΔ\\Deltaabs\-TPRΔ\\Deltapres\-TPRΔ\\Deltafact\-recallnReadingTarget:L21–35 q\_projlora\_A→\\rightarrow0\+33\.0pp−\-19\.7pp—50Causal, primaryControl \(layer\): L5–19 q\_projlora\_A→\\rightarrow0−\-3\.0pp\+2\.0pp−\-0\.33pp50Layer specificity ✓Control \(projection\): L21–35 v\_projlora\_A→\\rightarrow0\+18\.67pp−\-12\.0pp−\-6\.0pp50Projection specificity \(partial\)The layer control is clean: zeroing an earlier 15\-layer q\_proj band moves no metric by more than 3pp, ruling out “any band oflora\_Aweights matters\.” The projection control is partial: zeroing v\_proj on the same L21–35 range still produces a meaningful absence\-TPR shift \(\+18\.67pp,≈\\approx57% of the target’s magnitude\) and a smaller presence\-TPR shift \(−\-12\.0pp\)\. We therefore weaken the specificity claim accordingly:zeroing q\_proj at L21–35 is sufficient to produce the calibration↔\\leftrightarrowrecall trade\-off, and the layer band is specific \(the L5–19 control is null\), but specificity to the projection axis is partial — v\_proj on the same layers still carries a substantial portion of the effect\.A clean \(layer×\\timesprojection\) joint\-specificity claim would require a finer search over 4×\\times36 = 144 \(proj, layer\) cells; we leave that to the within\-band isolation work flagged in §5\.6\.

### D\.7Pre\-registered prediction: hybrid Pareto\-superiority

We close the discussion with a falsifiable prediction for the follow\-up paper, recorded here in advance of running the experiment that tests it\.If the substrate asymmetry documented above is real and not corpus\-specific, then a hybridγ\\gamma\-LoRA \+ BGE\-RAG \+ calibration\-head system, trained on the same per\-user data and evaluated on the same three\-axis diagnostic, should be Pareto\-superior to each component alone: style≥\\geqγ\\gamma\-LoRA\-alone, factual presence≥\\geqRAG\-alone, factual absence≥\\geqRAG\-alone, with no axis regressing relative to its best single\-substrate baseline\.We pre\-register this prediction with the explicit failure mode that would falsify it — any axis where the hybrid lands below the better of \{γ\\gamma\-LoRA\-alone, RAG\-alone\} by more than the per\-axis 95% CI half\-width — and commit to reporting the result regardless of direction\. This is the natural consequence of the framework: if the axes are separable, routing per axis should compose; if it doesn’t compose, the separability claim is weaker than we argue here, and the framework needs to be revised in light of that\.

## Appendix EAppendix E: Related work \(full comparison table\)

Our work touches three threads: \(i\) personalization benchmarks, \(ii\) parametric vs retrieval memory architectures, and \(iii\) calibration of LLM outputs under partial knowledge\. The contribution gap is best stated up front — every prior thread we know of measuresaggregate top\-line accuracyunder personalization; none separates the*behavioral*\(style, voice, preference\) axis from the*factual presence/absence*axis, and none reports the substrate\-asymmetry we observe across them\.

### E\.1Personalization benchmarks

LaMP\(Salemi et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib29)\)is the canonical user\-side personalization benchmark — seven tasks \(LaMP\-1 through LaMP\-7\) spanning citation identification \(LaMP\-1\), movie tagging \(LaMP\-2\), product review rating \(LaMP\-3\), news headline generation \(LaMP\-4\), scholarly title generation \(LaMP\-5\), email subject generation \(LaMP\-6\), and tweet paraphrasing \(LaMP\-7\)\. The headline metric is per\-task accuracy or ROUGE against the held\-out user output\. We useLaMP\-3as our real\-data transfer probe \(§3\.4\) precisely because it has the cleanest structural overlap with our synthetic setup \(per\-user history of past ratings→\\rightarrowpredict the next rating\) and the smallest leakage surface \(rating∈\\in\{1\.\.5\}, no free\-text generation\)\. LaMP itself does not separately report style vs factual recall; it bundles both into top\-1 accuracy on the held\-out item\.

PERSOMA\(Hebert et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib10)\)introduces persona\-conditioned soft tokens trained over user history, and reports gains over flat\-history baselines on a held\-out user\-utterance prediction task\. Ourγ\\gamma\-LoRA can be read as a heavier\-weight cousin: per\-user low\-rank adapters trained on the user’s own corpus, evaluated on the same held\-out\-utterance shape\. PERSOMA’s evaluation is single\-axis \(utterance\-level accuracy / log\-likelihood\); ours adds the absence axis \(§3\.3\), where we find that behavioral substrates fail in a different direction than retrieval\.

UQABench\(Liu et al\.,[2025](https://arxiv.org/html/2606.11712#bib.bib21)\)targets user\-question\-answering with long histories and reports retrieval\-augmented baselines plus prompt\-tuning baselines on factual recall\. Like LaMP, the metric is aggregate accuracy\. UQABench is closer in spirit to our absence probe than LaMP\-3 is — it explicitly tests questions whose answers do or do not appear in the user’s history — but it does not separately report the TPR\-on\-presence vs TPR\-on\-absence asymmetry that motivates our calibration finding\.

Other personalization benchmarks— DialPRX, PerLTQA, LongLaMP — fall into the same shape: per\-user data \+ held\-out target \+ aggregate accuracy / ROUGE\. None report substrate\-asymmetry on the two axes we isolate\. The broader*LLM\-as\-recommender*line\(Geng et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib8); Bao et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib3)\)casts personalization as a sequence\-modeling task with shared parameters across users, which is orthogonal to the per\-user\-substrate question we ask here\.

The in\-context\-learning baseline\(Brown et al\.,[2020](https://arxiv.org/html/2606.11712#bib.bib5)\)is the natural upper bound on “no per\-user training at all”: pack the user’s history into the prompt and let the base model generalize\. We treat B\_full \(history\-in\-prompt\) in §4 as exactly this baseline; our finding is that ICL recovers absence\-TPR but not style consistency once the history exceeds a few hundred tokens\.

### E\.2Memory architectures

Prompt tuning and prefix tuning\(Lester et al\., 2021; Li & Liang, 2021\) attach a small set of trainable continuous tokens to the input, freezing the base model\. They are well\-studied for task adaptation and have been adapted to per\-user personalization \(e\.g\., PERSOMA’s persona prefixes\)\. We do not study prompt tuning directly because our priorγ\\gamma\-LoRA experiments in unreported preliminary work found that LoRA adapters consistently outperformed prefix tuning at matched parameter count on the same synthetic personae corpus, and our framing here is the substrate\-asymmetry rather than parameter\-count efficiency\.

LoRA\-as\-memory\.Low\-rank adapters\(Hu et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib11)\)are usually trained for task adaptation across many users; using them asper\-user memoryis the formulation we adopt \(γ\\gamma\-LoRA: rank\-r adapter trained per persona on that persona’s history\)\. The closest prior art isLoRA\-Hub\(Huang et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib12)\)and persona\-conditioned LoRA work in chatbot agents\. Our diagnostic finding — that LoRA captures style well but factual absence poorly — recasts LoRA\-as\-memory as a*behavioral*memory primitive, not a general one\.

Retrieval\-augmented memory\.MemGPT\(Packer et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib25)\), Letta, and RAG\-style augmentations\(Lewis et al\.,[2020](https://arxiv.org/html/2606.11712#bib.bib16); Borgeaud et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib4)\)maintain an external store and inject retrieved chunks at inference\. For factual recall and presence/absence calibration these are the strong baseline in our results: §4\.2 shows RAG reaches 99% TPR on absence \(it correctly says “I don’t know” when the chunk is absent\) whileγ\\gamma\-LoRA confabulates at 8\.7% TPR on the same probe\. The cost is on the style axis: RAG is essentially a no\-op for behavioral consistency at the n=50\-author scale we test \(§4\.1\)\.

Hybrid systems\.Several recent systems combine parametric updates with retrieval, but none — to our knowledge — has reported the*separate*style\-vs\-absence metrics that would let a reader see why both substrates are needed\.

Adapter composition / mixture\-of\-experts\.A separate line combines multiple adapters at inference time: AdapterFusion\(Pfeiffer et al\.,[2021](https://arxiv.org/html/2606.11712#bib.bib27)\)learns attention over task adapters, AdaMix\(Wang et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib31)\)stochastically routes through a pool of adapters, and Mixture\-of\-LoRA\-Experts\(Wu et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib32)\)extends this to LoRA modules\. These methods compose adapters*across tasks*; ourγ\\gamma\-LoRA composes*across users*, keeping one adapter per persona rather than a shared MoE pool\. The substrate question we raise is independent of routing: a LoRA adapter captures style and confabulates on absence whether or not other adapters are mixed in at inference\.

Continual learning and catastrophic forgetting\.Per\-user adapters bypass — rather than solve — the standard continual\-learning problem of overwriting prior knowledge when the same parameters are updated sequentially\(Kirkpatrick et al\.,[2017](https://arxiv.org/html/2606.11712#bib.bib14); Li & Hoiem,[2018](https://arxiv.org/html/2606.11712#bib.bib19); Chaudhry et al\.,[2019](https://arxiv.org/html/2606.11712#bib.bib6)\)\. By isolating each user in their own LoRA, we trade catastrophic forgetting for confabulation: the per\-user weights faithfully encode user\-specific style without disturbing the base model, but they interpolate confidently into absent regions of the user’s factual support\. Our calibration finding can therefore be read as a continual\-learning critique of the “one\-adapter\-per\-user” design point\.

Privacy framing\.Per\-user parametric memory is also a natural fit for federated learning\(McMahan et al\.,[2017](https://arxiv.org/html/2606.11712#bib.bib23)\)and differential privacy\(Abadi et al\.,[2016](https://arxiv.org/html/2606.11712#bib.bib1)\): each user’s adapter is trained on their own corpus and need not leave their device\. We do not run a federated\-or\-DP experiment in this paper; we flag it as a deployment direction where the substrate\-asymmetry we measure becomes a*feature*\(per\-user style without per\-user fact\-leakage\) rather than a liability\.

SystemArchitectureWhat’s measuredWhat’s notGenerative Agents\(Park et al\. 2023, UIST\)retrieval over an episodic memory stream \+ reflection\-based summarizationqualitative believability via human eval; agent\-vs\-agent simulation fidelityno probe\-decomposition; no per\-substrate ablationMem0\(Chhikara et al\. 2024, arXiv:2504\.19413\)LLM\-extracted facts stored in a vector \+ graph store; retrieval at query timeLOCOMO benchmark accuracy; latency; storage costfactual axis only; behavioral consistency not separately scoredMemoryBank\(Zhong et al\. 2024, AAAI\)retrieval store \+ Ebbinghaus\-curve forgetting; no parametric updateconversation continuity scoresno parametric arm; no absence calibrationA\-MEM\(Xu et al\. 2024, arXiv:2410\.10739\)agentic memory with link prediction, dynamic restructuringretrieval recall; downstream task accuracyretrieval\-only; no behavioral\-style measurementMemGPT / Letta\(Packer et al\. 2023, arXiv:2310\.08560\)OS\-style paged memory with retrieval; some downstream forks add SFT/LoRAretrieval QA on Document QA \+ Conversation taskspublished evaluations are retrieval\-side; no style\-vs\-absence splitThe shared blind spot across this line of work is the absence of a*decomposed*benchmark that separates behavioral consistency from factual presence from factual absence\. Each system optimizes for whichever axis its retrieval \+ summarization layer most naturally serves; the parametric ingredient, when present, tends to be a chat\-style fine\-tune layered on top, never directly compared to the retrieval substrate on the same axis\. Our work argues that the substrate\-asymmetry we measure is the empirical motivation for hybrids, not a post\-hoc justification — and that without an asymmetry\-measuring benchmark the field cannot tell whether a hybrid beats its components on each individual axis or only in aggregate\.

### E\.3Calibration and abstention

LLM calibration — knowing what one does not know — has been studied throughTruthfulQA\(Lin et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib20)\),selective prediction\(Kadavath et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib13)\), andabstain prompts\(Asai et al\.,[2024](https://arxiv.org/html/2606.11712#bib.bib2); Yang et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib35)\)\. The closest framing to our absence probe isretrieval\-as\-abstention: if the retriever returns nothing, the model abstains\. We quantify this on the synthetic personae corpus and find that retrieval\-as\-abstention is well\-calibrated \(RAG\-absence\-TPR≈\\approx0\.99\) while parametric memory is not \(γ\\gamma\-LoRA absence\-TPR≈\\approx0\.087\)\. This is consistent with the broader observation that supervised fine\-tuning teaches the model to produce a confident answer for any input\(Schulman,[2023](https://arxiv.org/html/2606.11712#bib.bib30); Ouyang et al\.,[2022](https://arxiv.org/html/2606.11712#bib.bib24)\); LoRA\-as\-memory inherits the same miscalibration\.

Ourcalibration headin §3\.3 is a small classifier that gates theγ\\gamma\-LoRA output with a “I don’t know” prediction; it recovers about half of the absence\-TPR gap \(F1 0\.470 on the absence subset\) without retrieval\. This is a smaller\-than\-RAG result presented honestly — the calibration head is a cheap third component, not a replacement for the retrieval substrate\.

Evaluation methodology\.We score style consistency with held\-out log\-likelihood and absence\-TPR with exact\-match abstention rather than relying on LLM\-as\-judge protocols\(Zheng et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib36); Liu et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib22); Li et al\.,[2023](https://arxiv.org/html/2606.11712#bib.bib18)\)\. Judge\-based evaluation is the dominant aggregate metric in recent personalization work, but it folds style and factual axes into a single scalar preference; this is exactly the conflation our decomposition is designed to surface, so we deliberately use cheaper, axis\-specific metrics\. We discuss the tradeoff \(cost vs decomposition fidelity\) in Appendix D\.

### E\.4Position vs prior work

AxisPrior work measuresWe additionally measurePersonalization accuracyLaMP, PERSOMA, UQABench \(aggregate\)Substrate\-asymmetry on style \+ absenceMemory architecturePrompt tuning, LoRA, RAG \(per substrate\)When each*fails*— and in what directionCalibrationTruthfulQA, selective prediction \(general\)Per\-substrate absence\-TPR; calibration\-head recoveryReal\-data transferLaMP\-N task accuracyγ\\gamma\-LoRA on LaMP\-3 vs majority baseline \(honest negative\)Mechanism interpretabilityLayer attribution on task benchmarks \(Geva et al\., logit lens\)Per\-persona weight\-band mass correlated with behavioral outcomes in opposite directions on the two probes \(n=50, \|r\|≤\\leq0\.49\)Causal localizationFeature ablation, activation patching on synthetic tasksZeroing L21–35 q\_projlora\_Aweights at test time→\\rightarrow\+33pp absence\-TPR,−\-20pp presence\-TPR \(n=50, paired\)The contribution is adiagnostic frameworkfor user\-side memory in LLMs, not a new substrate\. Ourγ\\gamma\-LoRA implementation, the synthetic personae corpus, the four\-config decomposition \(B\_nohist, B\_full, C\_rag, C\_lora\), and the two probes \(style log\-likelihood, absence\-TPR\) are the artifacts; the framework’s value is that running them on any new substrate proposal answers “behavioral or factual? presence or absence?” — questions current benchmarks do not separate\.