Grokers: Bottom-Up Inductive Comprehension and Write-Time Intelligence over Typed Knowledge Graphs

# Bottom-Up Inductive Comprehension and Write-Time Intelligence over Typed Knowledge Graphs
Source: [https://arxiv.org/html/2606.00050](https://arxiv.org/html/2606.00050)
###### Abstract

We presentGrokers, an architecture for building persistent, structured comprehension of typed knowledge graphs through bottom\-up inductive traversal of dependency subgraphs\. Unlike retrieval\-augmented generation \(RAG\), which pays full comprehension cost at every query,Grokerspushes intelligence to*write time*: autonomous Groker agents analyze nodes in a typed stream graph, extract structured attributes via governed language model \(LM\) calls, and inductively compose that understanding upward through dependency relations—writing enriched typed attributes that serve all future queries at zero additional LM cost\. We prove three formal properties: \(1\) the*Byte\-Identity Theorem*, establishing that context blocks assembled from a transactionally\-maintained denormalization index are byte\-identical across LM turns between semantic changes, enabling KV\-cache hit rates approaching 100%; \(2\) the*Accumulation Monotonicity Theorem*, establishing that the fraction of interactions resolved without LM calls is non\-decreasing in the number of completed interactions under a governed wisdom library growth protocol; and \(3\) the*Dual\-Traversal Ordering Theorem*, establishing that top\-down generation and bottom\-up comprehension are the unique correct traversal orderings for their respective tasks over a dependency DAG, and that their composition closes into a complete generation\-comprehension cycle\. We further present a deterministic alternative to embedding\-based semantic search, with a synonym caching protocol whose LM fallback rate converges to zero for finite\-vocabulary domains\. A reference implementation is provided in the open\-source Qbix / Safebox / Safebots stack, building on the Magarshak Machine SPACER substrate\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]\.

## IIntroduction

The dominant paradigm for grounding language model \(LM\) responses in structured knowledge is*retrieval at query time*: embed the query, find nearest\-neighbor document chunks by vector similarity, inject them into the prompt context\[[10](https://arxiv.org/html/2606.00050#bib.bib3)\]\. This approach has a structural deficiency: it pays the full cost of comprehension on every query, regardless of whether that query is novel or structurally equivalent to thousands of prior queries\. For knowledge domains with recurring interaction patterns—enterprise software, document management, structured task execution—this is architecturally wasteful\.

We presentGrokers, an architecture that inverts this design choice\. Comprehension work is performed*once*, at write time, by autonomous agents that traverse the dependency graph of a typed knowledge substrate bottom\-up—from leaf nodes toward roots—extracting and storing structured typed attributes via governed LM calls\. By query time, the relevant structure is already present as typed attributes on graph nodes\.

The core substrate is the*typed stream graph*: a directed graph of typed nodes with structured fields, typed attributes, and weighted typed edges, realizing the Streams \(SS\) and Relations \(RR\) components of the Magarshak Machine SPACER framework\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]\. A transactionally\-maintained denormalization table \(Streams\_Category\) stores, for each node, the complete pre\-ranked relation neighborhood updated atomically on every relation modification, providing complete context in≈1\\approx\\\!1\\,ms\.

### Contributions

1. 1\.TheGrokerswrite\-time comprehension architecture \(§[V](https://arxiv.org/html/2606.00050#S5)\)\.
2. 2\.The*Byte\-Identity Theorem*with KV\-cache cost analysis \(§[IV](https://arxiv.org/html/2606.00050#S4)\)\.
3. 3\.The*Wisdom Library*with the*Accumulation Monotonicity Theorem*\(§[VI](https://arxiv.org/html/2606.00050#S6)\)\.
4. 4\.The*Dual\-Traversal Ordering Theorem*with incremental expansion corollary \(§[VII](https://arxiv.org/html/2606.00050#S7)\)\.
5. 5\.A deterministic semantic search system with*Synonym Cache Convergence*\(§[IX](https://arxiv.org/html/2606.00050#S9)\)\.

### Positioning

Individual components—knowledge graphs\[[6](https://arxiv.org/html/2606.00050#bib.bib4)\], LM\-based enrichment\[[11](https://arxiv.org/html/2606.00050#bib.bib5)\], KV caching\[[1](https://arxiv.org/html/2606.00050#bib.bib10)\], program libraries\[[4](https://arxiv.org/html/2606.00050#bib.bib16)\]—are each established\. Our contribution is the formal analysis of their composition and the three proved properties above, none of which appear in prior work\.

## IIRelated Work

Retrieval\-augmented generation\.Lewis et al\.\[[10](https://arxiv.org/html/2606.00050#bib.bib3)\]established RAG as the standard for grounding LM responses\. GraphRAG\[[6](https://arxiv.org/html/2606.00050#bib.bib4)\]adds community summarization over a knowledge graph but retains embedding similarity as the retrieval mechanism; neither GraphRAG nor any RAG variant achieves byte\-identity of assembled context or monotone LM\-call elimination\.

Code comprehension agents\.Bottom\-up comprehension is established in software engineering\[[2](https://arxiv.org/html/2606.00050#bib.bib13)\]\. Recent LM tools \(Copilot\[[8](https://arxiv.org/html/2606.00050#bib.bib6)\], Devin\[[5](https://arxiv.org/html/2606.00050#bib.bib7)\], RepoUnderstander\[[11](https://arxiv.org/html/2606.00050#bib.bib5)\]\) operate reactively without persistent enriched structure or formal traversal\-ordering guarantees\.

Memory systems\.MemGPT\[[15](https://arxiv.org/html/2606.00050#bib.bib8)\]and Mem0\[[14](https://arxiv.org/html/2606.00050#bib.bib9)\]retrieve facts by embedding similarity at query time\.Grokersdiffers: enrichment is write\-time, retrieval is a typed graph read, and the accumulation property is proved\.

KV\-cache optimization\.Anthropic prompt caching\[[1](https://arxiv.org/html/2606.00050#bib.bib10)\], SGLang\[[19](https://arxiv.org/html/2606.00050#bib.bib11)\], and LMCache\[[9](https://arxiv.org/html/2606.00050#bib.bib12)\]cache whatever prefix the caller provides\.Grokersis, to our knowledge, the first to*architect*context assembly so that stable prefixes are byte\-identical from first principles\.

Magarshak Machine\.Magarshak\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]introduces the SPACER substrate: append\-only streams, policy governance, five\-phase action execution \(Compute→Require→Execute\\textsc\{Compute\}\\to\\textsc\{Require\}\\to\\textsc\{Execute\}\), and the bidirectional relation index\.Grokersis an application layer on this substrate; its formal properties derive in part from SPACER’s substrate guarantees \(Minimal Cache Invalidation, Embarrassing Parallelism,ε\\varepsilon\-deterministic view capabilities\)\.

Hierarchical generation\.Hierarchical generation\[[7](https://arxiv.org/html/2606.00050#bib.bib17),[17](https://arxiv.org/html/2606.00050#bib.bib18)\]and planning\-then\-writing\[[16](https://arxiv.org/html/2606.00050#bib.bib19)\]do not formalize traversal ordering or relate it to KV\-cache stability\. The Dual\-Traversal Ordering Theorem \(§[VII](https://arxiv.org/html/2606.00050#S7)\) provides this connection\.

## IIIThe Typed Stream Graph

###### Definition 1\(Typed Stream Graph\)\.

A*typed stream graph*G=\(V,E,τ,α,w\)G=\(V,E,\\tau,\\alpha,w\)consists of: node setVVwith type functionτ:V→𝒯\\tau:V\\to\\mathcal\{T\}; directed typed edge setE⊆V×ℛ×VE\\subseteq V\\times\\mathcal\{R\}\\times V; typed attribute functionα:V→\(K→A\)\\alpha:V\\to\(K\\to A\); and weight functionw:E→ℝ≥0w:E\\to\\mathbb\{R\}\_\{\\geq 0\}driven by vote aggregationw​\(v,r,u\)=∑ivotei​\(v,r,u\)w\(v,r,u\)=\\sum\_\{i\}\\mathrm\{vote\}\_\{i\}\(v,r,u\), updated in the same database transaction as vote events \(realizing theRR/Relations component of\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]\)\.

###### Definition 2\(Dependency Subgraph\)\.

The*dependency subgraph*D​\(v\)D\(v\)is the induced subgraph on nodes reachable fromvvby𝖽𝖾𝗉𝖾𝗇𝖽𝗌​\_​𝗈𝗇\\mathsf\{depends\\\_on\}edges\. A nodeu∈D​\(v\)u\\in D\(v\)is a*leaf*if it has no outgoing𝖽𝖾𝗉𝖾𝗇𝖽𝗌​\_​𝗈𝗇\\mathsf\{depends\\\_on\}edges withinD​\(v\)D\(v\)\. We assumeD​\(v\)D\(v\)is a DAG for allv∈Vv\\in V\.

###### Definition 3\(Enriched / Stale Nodes\)\.

Nodevvis*enriched*ifα​\(v\)\\alpha\(v\)contains non\-null values for the core schema fields\{𝖽𝗂𝗀𝖾𝗌𝗍,𝗄𝖾𝗒𝗐𝗈𝗋𝖽𝗌,𝗊𝗎𝖺𝗅𝗂𝗍𝗒𝖲𝖼𝗈𝗋𝖾\}\\\{\\mathsf\{digest\},\\allowbreak\\mathsf\{keywords\},\\allowbreak\\mathsf\{qualityScore\}\\\}\. \(Implementations may include additional fields such asgrokVersionfor cache invalidation; these are outside the formal model\.\) Nodevvis*stale*ifα​\(v\)​\[𝗌𝗍𝖺𝗅𝖾\]=𝚝𝚛𝚞𝚎\\alpha\(v\)\[\\mathsf\{stale\}\]=\\mathtt\{true\}\.

## IVDenormalization Index and the Byte\-Identity Theorem

### IV\-AIndex Structure

TheStreams\_Categorytable maintains for eachv∈Vv\\in V:

N​\(v\)=\{r↦\{wj↦\[pj,sj,tj,ij\]\}j∣\(v,r,uj\)∈E\}N\(v\)=\\bigl\\\{r\\mapsto\\\{w\_\{j\}\\mapsto\[p\_\{j\},s\_\{j\},t\_\{j\},i\_\{j\}\]\\\}\_\{j\}\\mid\(v,r,u\_\{j\}\)\\in E\\bigr\\\}wherepj,sjp\_\{j\},s\_\{j\}are the publisherId/streamName of thejj\-th neighbor at weightwjw\_\{j\}under relation typerr, andtj,ijt\_\{j\},i\_\{j\}are its denormalized title and icon\. The table is updated in the same database transaction as every𝖲𝗍𝗋𝖾𝖺𝗆𝗌::𝗋𝖾𝗅𝖺𝗍𝖾\(\)\\mathsf\{Streams::relate\(\)\}/𝖲𝗍𝗋𝖾𝖺𝗆𝗌::𝗎𝗇𝗋𝖾𝗅𝖺𝗍𝖾\(\)\\mathsf\{Streams::unrelate\(\)\}call—no background jobs, no eventual consistency\.

###### Definition 4\(Context Block\)\.

The*context block*C​\(v\)C\(v\)is the output of the deterministic function𝖻𝗎𝗂𝗅𝖽𝖢𝖺𝖼𝗁𝖾𝖽𝖢𝗈𝗇𝗍𝖾𝗑𝗍​\(v\)\\mathsf\{buildCachedContext\}\(v\): readsα​\(v\)\\alpha\(v\)andN​\(v\)N\(v\); renders them as structured text with relation types ordered by signal strength \(unique types first, non\-unique by maximum weight descending\)\.

### IV\-BThe Byte\-Identity Theorem

###### Theorem 1\(Byte\-Identity\)\.

Lett1<t2t\_\{1\}<t\_\{2\}be timestamps such that no write toα​\(v\)\\alpha\(v\)or any edge incident tovvoccurs in\(t1,t2\)\(t\_\{1\},t\_\{2\}\)\. ThenCt1​\(v\)=Ct2​\(v\)C\_\{t\_\{1\}\}\(v\)=C\_\{t\_\{2\}\}\(v\)as byte strings\.

###### Proof\.

C​\(v\)C\(v\)is a deterministic function of\(α​\(v\),N​\(v\)\)\(\\alpha\(v\),N\(v\)\)\.N​\(v\)N\(v\)is modified only within transactions that modify edges incident tovv; none occur in\(t1,t2\)\(t\_\{1\},t\_\{2\}\)by hypothesis\.α​\(v\)\\alpha\(v\)is modified only by governed writes \(via𝖠𝖼𝗍𝗂𝗈𝗇\.𝗉𝗋𝗈𝗉𝗈𝗌𝖾\\mathsf\{Action\.propose\}in SPACER’sExecutephase\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]\); none occur by hypothesis\. Both inputs are identical att1t\_\{1\}andt2t\_\{2\}; determinism givesCt1​\(v\)=Ct2​\(v\)C\_\{t\_\{1\}\}\(v\)=C\_\{t\_\{2\}\}\(v\)\. ∎

###### Corollary 1\(KV Cache Hit Rate\)\.

Assume semantic changes tovvfollow a Poisson process with mean inter\-arrival timeTc​\(v\)T\_\{c\}\(v\), and turns arrive with mean inter\-turn intervalTtT\_\{t\}, withTt≪Tc​\(v\)T\_\{t\}\\ll T\_\{c\}\(v\)\. The expected fraction of turns for whichC​\(v\)C\(v\)is a cache hit is1−Tt/Tc​\(v\)1\-T\_\{t\}/T\_\{c\}\(v\), which approaches 1 asTc​\(v\)/Tt→∞T\_\{c\}\(v\)/T\_\{t\}\\to\\infty\.

###### Proof\.

Under the Poisson assumption, the probability that at least one semantic change occurs in any given inter\-turn interval\[t,t\+Tt\]\[t,t\+T\_\{t\}\]is1−e−Tt/Tc​\(v\)≈Tt/Tc​\(v\)1\-e^\{\-T\_\{t\}/T\_\{c\}\(v\)\}\\approx T\_\{t\}/T\_\{c\}\(v\)forTt≪Tc​\(v\)T\_\{t\}\\ll T\_\{c\}\(v\)\. By Theorem[1](https://arxiv.org/html/2606.00050#Thmtheorem1), a cache hit occurs in every turn where no change has occurred since the previous turn\. The expected hit rate is therefore1−Tt/Tc​\(v\)→11\-T\_\{t\}/T\_\{c\}\(v\)\\to 1asTc​\(v\)/Tt→∞T\_\{c\}\(v\)/T\_\{t\}\\to\\infty\. ∎

For synchronous or near\-synchronous interaction \(e\.g\., chat\),TcT\_\{c\}is hours\-to\-days andTtT\_\{t\}is seconds, so the hit rate approaches 100%\. For asynchronous workflows whereTt∼TcT\_\{t\}\\sim T\_\{c\}, the hit rate decreases accordingly; the benefit is greatest in systems with rapid turn cadence relative to semantic change rate\. No RAG\-assembled context achieves this: retrieved chunks vary with the query, making context non\-deterministic across turns even when the underlying knowledge has not changed\.

### IV\-CCost Analysis

Letksk\_\{s\}denote the stable prefix token count \(a value determined by the goal system prompt and the denormalization\-assembled neighborhood block; in our reference implementation,ksk\_\{s\}is on the order of 2,000–3,000 tokens\)\. At 10% cache\-read pricing \(Anthropic prompt caching\[[1](https://arxiv.org/html/2606.00050#bib.bib10)\]charges 10% of normal input cost for cache hits and 125% for cache writes, with the write cost amortized over the cache lifetime; we account for read cost only in the steady\-state analysis\):

Grokers:0\.1​ks\+kd​per turn\\displaystyle:\\;0\.1k\_\{s\}\+k\_\{d\}\\text\{ per turn\}RAG:kr\+kd​per turn\(kr≈ks​–​3​ks\)\\displaystyle:\\;k\_\{r\}\+k\_\{d\}\\text\{ per turn\}\\quad\(k\_\{r\}\\approx k\_\{s\}\\text\{\-\-\}3k\_\{s\}\)wherekdk\_\{d\}is dynamic context tokens\. Stable prefix cost reduction: up to10×10\\timesin the limitkd≪ksk\_\{d\}\\ll k\_\{s\}withkr=ksk\_\{r\}=k\_\{s\}; the reduction approaches\(kr\+kd\)/\(0\.1​ks\+kd\)\(k\_\{r\}\+k\_\{d\}\)/\(0\.1k\_\{s\}\+k\_\{d\}\), which increases askrk\_\{r\}grows relative tokdk\_\{d\}\.

## VGroker Agents: Bottom\-Up Comprehension

###### Definition 5\(Valid Processing Order\)\.

A sequencev1,…,vnv\_\{1\},\\ldots,v\_\{n\}is a*valid Groker processing order*for DAGHHif for everyvjv\_\{j\}and everyuuwith\(vj,𝖽𝖾𝗉𝖾𝗇𝖽𝗌​\_​𝗈𝗇,u\)∈E​\(H\)\(v\_\{j\},\\mathsf\{depends\\\_on\},u\)\\in E\(H\),uuappears beforevjv\_\{j\}\(i\.e\., a topological sort ofHHwith leaves first\)\.

###### Theorem 2\(Composability of Groker Enrichment\)\.

Call a Groker invocation on nodevv*individually correct*if, given that all dependencies ofvvare correctly enriched, it produces enrichment forvvthat passes schema validation and maintains fitness≥fmin\\geq f\_\{\\min\}\(see §[XI](https://arxiv.org/html/2606.00050#S11)\)\. If each Groker invocation is individually correct, then processing nodes in any valid order \(Definition[5](https://arxiv.org/html/2606.00050#Thmdefinition5)\) produces correct enrichment for all nodes\.

###### Proof\.

Induction on topological order\.*Base*: leaf nodes have no dependencies; correctness holds vacuously\.*Step*: all prior nodes are correctly enriched by hypothesis; all dependencies ofviv\_\{i\}appear before it in the order, so they are correctly enriched; individual correctness gives correct enrichment forviv\_\{i\}\. ∎

Each Groker requires only its node’s content and pre\-enriched dependency attributes—not global context\. Independent Groker passes on nodes with no shared dependency path write to disjoint attribute namespaces \(α​\(v\)\\alpha\(v\)for each distinctvv\), satisfying the disjoint write\-set condition ofℳ​ℳ\\mathcal\{MM\}’s Embarrassing Parallelism Theorem\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]for attribute enrichment, enabling coordination\-free parallel enrichment scaling linearly with publisher count\. Index updates to the inverted keyword indexIdx\\mathrm\{Idx\}involve shared mutable state and are serialized by the substrate’s transactional index maintenance, separate from and non\-blocking to attribute enrichment parallelism\.

### V\-AStaleness Propagation

###### Proposition 1\(Staleness Completeness\)\.

Letvvbe modified\. Define the reverse dependency closureS​\(v\)=\{u∣∃path​u↝v​via​𝖽𝖾𝗉𝖾𝗇𝖽𝗌​\_​𝗈𝗇\}S\(v\)=\\\{u\\mid\\exists\\text\{ path \}u\\rightsquigarrow v\\text\{ via \}\\mathsf\{depends\\\_on\}\\\}\. After marking allu∈S​\(v\)u\\in S\(v\)stale and reprocessing them in valid topological order, no node is stale\.

###### Proof\.

By Theorem[2](https://arxiv.org/html/2606.00050#Thmtheorem2): reprocessingS​\(v\)S\(v\)in valid order withvvalready updated restores correctness to all affected nodes\. Nodes not inS​\(v\)S\(v\)have no dependency path tovvand are unaffected\. ∎

Propagation cost isO​\(\|S​\(v\)\|\)O\(\|S\(v\)\|\)—the build\-system model applied to semantic comprehension\. This mirrorsℳ​ℳ\\mathcal\{MM\}’s Minimal Cache Invalidation Theorem\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]: only streams reachable from the modified node in the dependency graph require reprocessing\.

## VIThe Wisdom Library and Accumulation Monotonicity

### VI\-AStructure

The*wisdom library*𝒲=\{pi\}\\mathcal\{W\}=\\\{p\_\{i\}\\\}is a finite set of*sandboxed imperative programs*, each with an execution phaseϕ​\(pi\)∈Φ\\phi\(p\_\{i\}\)\\in\\Phi, input/output schema, and*fitness*f​\(pi\)=\|\{x∈𝒳eval:pi​routes and handles​x​correctly\}\|/\|𝒳eval\|∈\[0,1\]f\(p\_\{i\}\)=\|\\\{x\\in\\mathcal\{X\}\_\{\\mathrm\{eval\}\}:p\_\{i\}\\text\{ routes and handles \}x\\text\{ correctly\}\\\}\|/\|\\mathcal\{X\}\_\{\\mathrm\{eval\}\}\|\\in\[0,1\]measured over a rolling evaluation window𝒳eval\\mathcal\{X\}\_\{\\mathrm\{eval\}\}\. Execution contract \(realizing SPACER’sCompute/Require/Execute\\textsc\{Compute\}/\\textsc\{Require\}/\\textsc\{Execute\}separation\): reads from a pre\-loaded immutable input object \(no live DB queries\); writes only via proposal accumulation \(no direct writes, corresponding toRequire\); time bound≤50\\leq 50\\,ms; memory≤64\\leq 64\\,MB; no network access\. Programs areε\\varepsilon\-deterministic view capabilities withε=0\\varepsilon=0\(imperative code, not stochastic models\), enablingℳ​ℳ\\mathcal\{MM\}’s Probabilistic Consensus mechanism\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]to achieve agreement probability exactly 1 across Safebox nodes executing the same program on the same input\.

###### Definition 6\(LM\-Call Elimination Rate\)\.

E​\(𝒲\)=\|\{x∈𝒳:routing​\(x,𝒲\)≠𝖫𝖬\}\|\|𝒳\|E\(\\mathcal\{W\}\)=\\tfrac\{\|\\\{x\\in\\mathcal\{X\}:\\mathrm\{routing\}\(x,\\mathcal\{W\}\)\\neq\\mathsf\{LM\}\\\}\|\}\{\|\\mathcal\{X\}\|\}where𝒳\\mathcal\{X\}is the interaction distribution for goal typegg\.

### VI\-BThree Growth Mechanisms

*Initial authoring*: at goal type creation, an LM generates𝒲\(0\)=\{p1\(0\),…,pk\(0\)\}\\mathcal\{W\}^\{\(0\)\}=\\\{p\_\{1\}^\{\(0\)\},\\ldots,p\_\{k\}^\{\(0\)\}\\\}covering all phases, reviewed before activation\.

*Pattern promotion*: after completed interactionxtx\_\{t\}, a Groker analysis pass identifies uncovered patterns and proposesAtA\_\{t\}:𝒲\(t\+1\)=𝒲\(t\)∪At\\mathcal\{W\}^\{\(t\+1\)\}=\\mathcal\{W\}^\{\(t\)\}\\cup A\_\{t\}\.

*Evolutionary selection*: programs withf​\(pi\)<fminf\(p\_\{i\}\)<f\_\{\\min\}are replaced by higher\-fitness fork variants evaluated in parallel—the program\-level instantiation of Grokers expert specialization\.

###### Theorem 3\(Accumulation Monotonicity\)\.

E​\(𝒲\(t\+1\)\)≥E​\(𝒲\(t\)\)E\(\\mathcal\{W\}^\{\(t\+1\)\}\)\\geq E\(\\mathcal\{W\}^\{\(t\)\}\)for allt≥0t\\geq 0\.

###### Proof\.

*Pattern promotion*:𝒲\(t\+1\)=𝒲\(t\)∪At\\mathcal\{W\}^\{\(t\+1\)\}=\\mathcal\{W\}^\{\(t\)\}\\cup A\_\{t\}\. Here “resolved” means*correctly*resolved: the program routesxxand its output passes schema validation\. The pattern promotion protocol requires all programs inAtA\_\{t\}to pass review \(schema validation and sandbox execution against held\-out examples\) before joining𝒲\\mathcal\{W\}; this gates out programs that would route incorrectly\. Under this guard, for anyxx: if routing\(x,𝒲\(t\)\)≠𝖫𝖬\(x,\\mathcal\{W\}^\{\(t\)\}\)\\neq\\mathsf\{LM\}, existing programs still correctly routexxunder𝒲\(t\+1\)\\mathcal\{W\}^\{\(t\+1\)\}\(new programs cannot remove correct coverage from previously resolved inputs\)\. Ifxxwas unresolved but correctly covered byp∈Atp\\in A\_\{t\}, it is now correctly resolved\. Therefore the correctly\-resolved set is non\-decreasing:E​\(𝒲\(t\+1\)\)≥E​\(𝒲\(t\)\)E\(\\mathcal\{W\}^\{\(t\+1\)\}\)\\geq E\(\\mathcal\{W\}^\{\(t\)\}\)\.

*Evolutionary selection*: We assume the*fitness faithfulness*condition:f​\(p′\)\>f​\(p\)f\(p^\{\\prime\}\)\>f\(p\)implies\|\{x∈𝒳:p′​\(x\)≠𝖫𝖬\}\|≥\|\{x∈𝒳:p​\(x\)≠𝖫𝖬\}\|\|\\\{x\\in\\mathcal\{X\}:p^\{\\prime\}\(x\)\\neq\\mathsf\{LM\}\\\}\|\\geq\|\\\{x\\in\\mathcal\{X\}:p\(x\)\\neq\\mathsf\{LM\}\\\}\|, i\.e\., empirical fitness is a reliable proxy for coverage on the full distribution\. Under this assumption,p′p^\{\\prime\}covers at least the interactions well\-handled bypp\. Fitness\-based replacement therefore cannot decrease coverage, givingE​\(𝒲\(t\+1\)\)≥E​\(𝒲\(t\)\)E\(\\mathcal\{W\}^\{\(t\+1\)\}\)\\geq E\(\\mathcal\{W\}^\{\(t\)\}\)\.

Since both mechanisms are non\-decreasing and the theorem holds under the stated assumption, the result follows\. ∎

###### Corollary 2\(Decreasing Marginal LM Cost\)\.

C𝖫𝖬\(t\)=\(1−E​\(𝒲\(t\)\)\)⋅c𝖫𝖬C\_\{\\mathsf\{LM\}\}^\{\(t\)\}=\(1\-E\(\\mathcal\{W\}^\{\(t\)\}\)\)\\cdot c\_\{\\mathsf\{LM\}\}is non\-increasing intt\. RAG, ReAct\[[18](https://arxiv.org/html/2606.00050#bib.bib14)\], and LangChain\[[3](https://arxiv.org/html/2606.00050#bib.bib15)\]all have constant marginal LM cost per interaction—Theorem[3](https://arxiv.org/html/2606.00050#Thmtheorem3)is the unique mechanism among those considered that breaks this constant\-cost property\.

The elimination rateE​\(𝒲\(t\)\)E\(\\mathcal\{W\}^\{\(t\)\}\)is bounded above by the*structural predictability*PgP\_\{g\}of goal typegg\. For task\-oriented goal types \(capability building, document review, support resolution\),PgP\_\{g\}is expected to be substantial—reflecting the large fraction of interactions that follow predictable structural patterns—though its precise value is domain\-dependent and subject to empirical measurement \(see Discussion, §[XI](https://arxiv.org/html/2606.00050#S11)\)\. The remaining fraction requires genuine synthesis and LM calls indefinitely\.

## VIIDual\-Traversal Ordering

###### Definition 7\(Coherent Generative Task\)\.

A*coherent generative task*over DAGHHassigns artifact content to each nodevvusing the content of its dependencies as shared context, and is*coherent*if all consumers of a shared dependency use the same version of that dependency\.

###### Theorem 4\(Dual\-Traversal Ordering\)\.

LetHHbe a dependency DAG\.

1. 1\.The valid orderings for correct Groker comprehension are precisely the topological sorts ofHHwith leaves first \(bottom\-up\); any ordering not in this class may produce incorrect enrichment for some node\.
2. 2\.The valid orderings for coherent artifact generation overHHare precisely the topological sorts ofHHwith roots first \(top\-down\); any ordering not in this class may produce an incoherent artifact for some node\.
3. 3\.These two classes of orderings are reverses of each other on any DAG: reversing any member of the root\-first class yields a member of the leaf\-first class, and vice versa\. The enriched attributes\{α​\(v\)\}v∈V\\\{\\alpha\(v\)\\\}\_\{v\\in V\}produced by bottom\-up comprehension of a generated artifact constitute, via𝖻𝗎𝗂𝗅𝖽𝖢𝖺𝖼𝗁𝖾𝖽𝖢𝗈𝗇𝗍𝖾𝗑𝗍\\mathsf\{buildCachedContext\}, a byte\-identical \(Theorem[1](https://arxiv.org/html/2606.00050#Thmtheorem1)\) KV\-cached context for subsequent top\-down generation of artifacts over the same schema\.

###### Proof\.

*\(a\)*Let\(v,𝖽𝖾𝗉𝖾𝗇𝖽𝗌​\_​𝗈𝗇,u\)∈E\(v,\\mathsf\{depends\\\_on\},u\)\\in E\. Groker comprehension ofvvrequires enrichedα​\(u\)\\alpha\(u\)\. Ifvvis processed beforeuu,α​\(u\)\\alpha\(u\)is unenriched and correctness fails\. Thereforeuumust precedevv—a topological sort with leaves first\. Sufficiency follows from Theorem[2](https://arxiv.org/html/2606.00050#Thmtheorem2)\.

*\(b\)*Generatingvvrequiresuu’s content as shared context \(definition of coherent generative task\)\. Ifvvis generated beforeuu, shared context is unavailable and coherence fails\. Thereforeuumust precedevv—a topological sort with roots first\.

*\(c\)*Root\-first and leaf\-first topological sorts are reverses of each other on any DAG\. The generation phase \(orderσ\\sigma\) produces content for all nodes, each artifact appended to the graph \(Append\-Only Safety of\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]guarantees no loss\)\. The comprehension phase \(orderσ−1\\sigma^\{\-1\}\) enriches all nodes\. The resulting\{α​\(v\)\}\\\{\\alpha\(v\)\\\}is assembled by the deterministic𝖻𝗎𝗂𝗅𝖽𝖢𝖺𝖼𝗁𝖾𝖽𝖢𝗈𝗇𝗍𝖾𝗑𝗍\\mathsf\{buildCachedContext\}into a context block that, by Theorem[1](https://arxiv.org/html/2606.00050#Thmtheorem1), is byte\-identical across turns until the artifact changes\. ∎

###### Corollary 3\(Incremental Shared Dependency Expansion\)\.

During top\-down generation, when a leaf encounters a new shared dependencyddnot in the current shared dependency stream: \(i\) Addingdddoes not invalidate leaves not depending ondd; \(ii\) leaves already generated that depend onddare stale \(Proposition[1](https://arxiv.org/html/2606.00050#Thmproposition1)\) and require regeneration; \(iii\) all subsequent leaf generations haveddavailable as shared context\.

This corollary formalizes the website generation pattern: a new CSS variable discovered during page generation is added to the design system stream, requires regenerating only affected pages, and is available to all subsequent pages\.

### VII\-AContext Stability Hierarchy

The dual\-traversal structure maps directly onto the KV\-cache stability hierarchy:

- •PERMANENT: root\-level architecture, stable for the goal type’s lifetime
- •SESSION: section/module context, byte\-identical between semantic changes
- •COLD: multi\-level summary tree, changes every≈102\\approx 10^\{2\}–10310^\{3\}messages
- •DYNAMIC: wisdom\-program\-selected context, varying per turn

Root\-level architecture→\\toPERMANENT; section context→\\toSESSION; leaf requirements→\\toDYNAMIC\. Shared context at each hierarchy level is computed once and reused across all descendant leaf generations\.

## VIIIObservations Schema and Write\-Time Extraction

An*observations schema*maps stream typesTTto extraction specificationsσ​\(T,C\)\\sigma\(T,C\):*instruction clauses*\[ℓ1,…,ℓk\]\[\\ell\_\{1\},\\ldots,\\ell\_\{k\}\]\(natural\-language directives constraining LM extraction output, constituting the stable portion of the system prompt for typeTT\), output schemaΣ\\Sigma, attribute namespaceπ\\pi, and staleness conditionδ\\delta\.

Extraction uses a two\-part call: the*stable system prompt*\(constant across all extractions of typeTTand categoryCC, KV\-cached by Theorem[1](https://arxiv.org/html/2606.00050#Thmtheorem1)\) and the*dynamic user content*\(the specific stream’s content and existing attributes, varying per stream\)\.

###### Proposition 2\(Extraction Amortization\)\.

Withksk\_\{s\}stable system prompt tokens andk¯u\\bar\{k\}\_\{u\}mean dynamic content tokens, the mean per\-extraction prefill cost at 10% cache pricing isC¯=0\.1​ks\+k¯u\\bar\{C\}=0\.1k\_\{s\}\+\\bar\{k\}\_\{u\}\. The asymptotic reduction factor relative to no caching is:R=\(ks\+k¯u\)/\(0\.1​ks\+k¯u\)→ks/k¯u→∞10\.R=\(k\_\{s\}\+\\bar\{k\}\_\{u\}\)/\(0\.1k\_\{s\}\+\\bar\{k\}\_\{u\}\)\\xrightarrow\{k\_\{s\}/\\bar\{k\}\_\{u\}\\to\\infty\}10\.

###### Proof\.

Without caching, the per\-extraction prefill cost isks\+k¯uk\_\{s\}\+\\bar\{k\}\_\{u\}\. With 10% cache\-hit pricing on the stable prefix, the effective cost is0\.1​ks\+k¯u0\.1k\_\{s\}\+\\bar\{k\}\_\{u\}\. The reduction factor isR=\(ks\+k¯u\)/\(0\.1​ks\+k¯u\)R=\(k\_\{s\}\+\\bar\{k\}\_\{u\}\)/\(0\.1k\_\{s\}\+\\bar\{k\}\_\{u\}\)\. Dividing numerator and denominator byk¯u\\bar\{k\}\_\{u\}:R=\(ks/k¯u\+1\)/\(0\.1​ks/k¯u\+1\)→1/0\.1=10R=\(k\_\{s\}/\\bar\{k\}\_\{u\}\+1\)/\(0\.1k\_\{s\}/\\bar\{k\}\_\{u\}\+1\)\\to 1/0\.1=10asks/k¯u→∞k\_\{s\}/\\bar\{k\}\_\{u\}\\to\\infty\. ∎

Extracted attributes, once written, serve context assembly, preprocessing, relevance scoring, autosuggest, and search indexes—all at zero additional LM cost\.

## IXDeterministic Semantic Search

### IX\-AThe Precision Problem with Embedding Similarity

For structured domains, typed attribute queries \(“streams withα​\(v\)​\[𝗌𝗍𝖺𝗍𝗎𝗌\]=𝚞𝚗𝚌𝚘𝚗𝚏𝚒𝚐𝚞𝚛𝚎𝚍\\alpha\(v\)\[\\mathsf\{status\}\]=\\mathtt\{unconfigured\}”\) have zero false positives by construction—a precision level unachievable by embedding similarity, which computes undifferentiated similarity over raw content\.

### IX\-BWrite\-Time Indexing and Query Expansion

Keyword extraction is a first\-class observation category: the Groker extracts normalized keyword setK​\(v\)K\(v\)for eachvv, stored asα​\(v\)​\[𝗄𝖾𝗒𝗐𝗈𝗋𝖽𝗌\]\\alpha\(v\)\[\\mathsf\{keywords\}\]and indexed into inverted indexIdx​\[k\]=\{v∣k∈K​\(v\)\}\\mathrm\{Idx\}\[k\]=\\\{v\\mid k\\in K\(v\)\\\}\. Query terms are expanded via: stemming; domain ontology traversal \(IS\-A, SYNONYM\-OF edges as typed graph relations\); thesaurus lookup; and cached prior LM expansions\. Search resolves by set intersection:score​\[v\]=\|Q∩K​\(v\)\|\\mathrm\{score\}\[v\]=\|Q\\cap K\(v\)\|\.

###### Theorem 5\(Synonym Cache Convergence\)\.

Let𝒲\\mathcal\{W\}be the finite domain vocabulary with\|𝒲\|<∞\|\\mathcal\{W\}\|<\\infty\. Assume each query contains at least one term \(\|𝒯q​\(n\)\|≥1\|\\mathcal\{T\}\_\{q\}\(n\)\|\\geq 1for allnn\)\. The LM fallback rateρ​\(n\)=\|𝒯qnew​\(n\)\|/\|𝒯q​\(n\)\|→0\\rho\(n\)=\|\\mathcal\{T\}\_\{q\}^\{\\mathrm\{new\}\}\(n\)\|/\|\\mathcal\{T\}\_\{q\}\(n\)\|\\to 0asn→∞n\\to\\infty\.

###### Proof\.

LetSn=⋃i=1n𝒯q​\(i\)S\_\{n\}=\\bigcup\_\{i=1\}^\{n\}\\mathcal\{T\}\_\{q\}\(i\)be the set of all vocabulary terms seen in queries11throughnn\. SinceSn⊆𝒲S\_\{n\}\\subseteq\\mathcal\{W\}and\|𝒲\|<∞\|\\mathcal\{W\}\|<\\infty, the sequence\|Sn\|\|S\_\{n\}\|is non\-decreasing and bounded above by\|𝒲\|\|\\mathcal\{W\}\|, so it converges\. Therefore\|𝒯qnew​\(n\)\|=\|Sn∖Sn−1\|→0\|\\mathcal\{T\}\_\{q\}^\{\\mathrm\{new\}\}\(n\)\|=\|S\_\{n\}\\setminus S\_\{n\-1\}\|\\to 0\. Since\|𝒯q​\(n\)\|≥1\|\\mathcal\{T\}\_\{q\}\(n\)\|\\geq 1by assumption, the denominator is bounded below, andρ​\(n\)=\|𝒯qnew​\(n\)\|/\|𝒯q​\(n\)\|≤\|𝒯qnew​\(n\)\|→0\\rho\(n\)=\|\\mathcal\{T\}\_\{q\}^\{\\mathrm\{new\}\}\(n\)\|/\|\\mathcal\{T\}\_\{q\}\(n\)\|\\leq\|\\mathcal\{T\}\_\{q\}^\{\\mathrm\{new\}\}\(n\)\|\\to 0\. ∎

Theorem[5](https://arxiv.org/html/2606.00050#Thmtheorem5)mirrors Theorem[3](https://arxiv.org/html/2606.00050#Thmtheorem3): in both cases, LM cost per operation converges to zero for finite\-domain workloads as the relevant cache accumulates coverage\.

## XImplementation

The architecture is realized across three plugins\.Qbixprovides the stream graph substrate:Streams::relate\(\),Streams\_Category\(maintained transactionally\),Users\_Vote\.Safeboxprovides the sandbox \(realizing SPACER’sCC/Capabilities andEE/Execution components\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]\):Protocol\.LLMwith Anthropic and OpenAI adapters,Action\.proposegoverned write pipeline,Protocol\.Transcription,Protocol\.Image\.Safebotsprovides the Groker enrichment layer, goal stream architecture, multi\-level summary tree, wisdom library management, and preprocessing pipeline\.

KV\-cache support: For Anthropic models, PERMANENT and SESSION blocks are passed as system content blocks withcache\_control:\{type:"ephemeral"\}, establishing explicit breakpoints\. For OpenAI models, they are merged into a single system message to maximize stable prefix length for automatic prefix caching\.

## XIDiscussion

Correctness assumptions\.Theorem[2](https://arxiv.org/html/2606.00050#Thmtheorem2)assumes individually correct Groker invocations\. In practice, Grokers use LMs and are stochastic; the governed write pipeline mitigates this via schema validation and fitness\-based evolution\. “Correct” means operationally: passes schema validation and maintains fitness≥fmin\\geq f\_\{\\min\}over the evaluation window\.

Finite vocabulary assumption\.Theorem[5](https://arxiv.org/html/2606.00050#Thmtheorem5)requires\|𝒲\|<∞\|\\mathcal\{W\}\|<\\infty\. For open\-domain deployments the vocabulary is unbounded; in this caseρ​\(n\)\\rho\(n\)converges to the fraction of novel user terms that have no ontology expansion, which may remain bounded away from zero indefinitely\. The theorem applies to structured knowledge domains \(product catalogs, codebases, organizational ontologies\) where the domain vocabulary is finite and indexable\.

Acyclic dependency assumption\.Theorem[4](https://arxiv.org/html/2606.00050#Thmtheorem4)assumes acyclic dependency graphs\. Cyclic dependencies \(mutual imports, circular references\) require cycle\-breaking heuristics; we leave this extension to future work\.

Architecture stack\.The full Magarshak Architecture is:ℳ​ℳ⏟substrate→Grokers⏟comprehension→Context⏟intelligence\\underbrace\{\\mathcal\{MM\}\}\_\{\\text\{substrate\}\}\\to\\underbrace\{\\text\{Grokers\}\}\_\{\\text\{comprehension\}\}\\to\\underbrace\{\\text\{Context\}\}\_\{\\text\{intelligence\}\}\[[12](https://arxiv.org/html/2606.00050#bib.bib2)\]\. This paper covers the comprehension layer; the intelligence layer—proactive goal\-directed agents, organizational efficiency theorems, cross\-platform governance—is treated in\[[12](https://arxiv.org/html/2606.00050#bib.bib2)\]\.

## XIIConclusion

Grokersprovides a write\-time intelligence architecture with three proved properties\. The Byte\-Identity Theorem establishes near\-100% KV\-cache hit rates on deterministically assembled stable prefixes—fundamentally unachievable by retrieval\-based systems\. The Accumulation Monotonicity Theorem establishes that the wisdom library covers an increasing fraction of interactions without LM calls, with marginal LM cost non\-increasing over time—a property no existing chatbot architecture exhibits\. The Dual\-Traversal Ordering Theorem establishes the duality of generation and comprehension as the unique correct orderings for their respective tasks, composing into a complete cycle in which comprehension output becomes cached context for future generation\. These properties constitute a formal basis for write\-time intelligence as a strictly more efficient alternative to query\-time retrieval for structured knowledge domains with recurring interaction patterns\.

## Acknowledgements

The stream graph substrate is implemented in the Qbix open\-source framework, which realizes the Magarshak Machine SPACER substrate\[[13](https://arxiv.org/html/2606.00050#bib.bib1)\]\.

## References

- \[1\]Anthropic\(2024\)Prompt caching\.External Links:[Link](https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching)Cited by:[§I](https://arxiv.org/html/2606.00050#S1.SSx2.p1.1),[§II](https://arxiv.org/html/2606.00050#S2.p4.1),[§IV\-C](https://arxiv.org/html/2606.00050#S4.SS3.p1.2)\.
- \[2\]T\. J\. Biggerstaff, B\. G\. Mitbander, and D\. Webster\(1993\)Program understanding and the concept assignment problem\.Communications of the ACM37\(5\),pp\. 72–82\.Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p2.1)\.
- \[3\]H\. Chase\(2022\)LangChain\.External Links:[Link](https://langchain.com/)Cited by:[Corollary 2](https://arxiv.org/html/2606.00050#Thmcorollary2.p1.2.1)\.
- \[4\]M\. Chen, J\. Tworek, H\. Jun,et al\.\(2021\)Evaluating large language models trained on code\.arXiv preprint arXiv:2107\.03374\.Cited by:[§I](https://arxiv.org/html/2606.00050#S1.SSx2.p1.1)\.
- \[5\]Cognition AI\(2024\)Devin: AI software engineer\.External Links:[Link](https://cognition.ai/)Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p2.1)\.
- \[6\]D\. Edge, H\. Trinh, N\. Cheng,et al\.\(2024\)From local to global: A Graph RAG approach to query\-focused summarization\.arXiv preprint arXiv:2404\.16130\.Cited by:[§I](https://arxiv.org/html/2606.00050#S1.SSx2.p1.1),[§II](https://arxiv.org/html/2606.00050#S2.p1.1)\.
- \[7\]A\. Fan, M\. Lewis, and Y\. Dauphin\(2018\)Hierarchical neural story generation\.InACL,Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p6.1)\.
- \[8\]GitHub\(2023\)GitHub Copilot\.External Links:[Link](https://github.com/features/copilot)Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p2.1)\.
- \[9\]R\. Jinet al\.\(2024\)LMCache: an efficient KV cache layer for enterprise\-scale LLM inference\.arXiv preprint arXiv:2410\.XXXXX\.Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p4.1)\.
- \[10\]P\. Lewis, E\. Petriv, A\. Piktus,et al\.\(2020\)Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.NeurIPS33,pp\. 9459–9474\.Cited by:[§I](https://arxiv.org/html/2606.00050#S1.p1.1),[§II](https://arxiv.org/html/2606.00050#S2.p1.1)\.
- \[11\]Y\. Ma, Z\. Zhang,et al\.\(2024\)RepoUnderstander: agent\-enhanced repository\-level code comprehension\.arXiv preprint arXiv:2408\.02563\.Cited by:[§I](https://arxiv.org/html/2606.00050#S1.SSx2.p1.1),[§II](https://arxiv.org/html/2606.00050#S2.p2.1)\.
- \[12\]G\. Magarshak\(2026\)Context: proactive goal\-directed intelligence via composable sandboxed programs, declarative wiring, and structured interaction\.arXiv preprint arXiv:2503\.XXXXX\.Cited by:[§XI](https://arxiv.org/html/2606.00050#S11.p4.1)\.
- \[13\]G\. Magarshak\(2026\)The Magarshak machine: A stream\-partitioned model for governed state evolution\. the SPACER framework\.arXiv preprint arXiv:2501\.XXXXX\.Cited by:[§I](https://arxiv.org/html/2606.00050#S1.p3.3),[§X](https://arxiv.org/html/2606.00050#S10.p1.2),[§II](https://arxiv.org/html/2606.00050#S2.p5.2),[§IV\-B](https://arxiv.org/html/2606.00050#S4.SS2.1.p1.10),[§V\-A](https://arxiv.org/html/2606.00050#S5.SS1.p1.2),[§V](https://arxiv.org/html/2606.00050#S5.p1.4),[§VI\-A](https://arxiv.org/html/2606.00050#S6.SS1.p1.10),[§VII](https://arxiv.org/html/2606.00050#S7.3.p3.4),[Acknowledgements](https://arxiv.org/html/2606.00050#Sx1.p1.1),[Definition 1](https://arxiv.org/html/2606.00050#Thmdefinition1.p1.8.8),[Remark 1](https://arxiv.org/html/2606.00050#Thmremark1.p1.8.8)\.
- \[14\]Mem0 AI\(2024\)Mem0: the memory layer for personalized AI\.External Links:[Link](https://mem0.ai/)Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p3.1)\.
- \[15\]C\. Packer, S\. Wooders, K\. Lin,et al\.\(2023\)MemGPT: towards LLMs as operating systems\.arXiv preprint arXiv:2310\.08560\.Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p3.1)\.
- \[16\]H\. Rashkin, A\. Celikyilmaz, Y\. Choi, and J\. Gao\(2020\)PlotMachines: outline\-conditioned generation with dynamic plot state tracking\.InEMNLP,Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p6.1)\.
- \[17\]K\. Yang and D\. Klein\(2022\)DOC: improving long story coherence with detailed outline control\.InACL,Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p6.1)\.
- \[18\]S\. Yao, J\. Zhao, D\. Yu,et al\.\(2023\)ReAct: synergizing reasoning and acting in language models\.ICLR\.Cited by:[Corollary 2](https://arxiv.org/html/2606.00050#Thmcorollary2.p1.2.1)\.
- \[19\]L\. Zheng, L\. Yin, Z\. Xie,et al\.\(2024\)Efficiently programming large language models using SGLang\.InNeurIPS,Cited by:[§II](https://arxiv.org/html/2606.00050#S2.p4.1)\.