When Summaries Distort Decisions: Information Fidelity in LLM-Compressed Financial Analysis

# When Summaries Distort Decisions: Information Fidelity in LLM-Compressed Financial Analysis
Source: [https://arxiv.org/html/2606.29251](https://arxiv.org/html/2606.29251)
Hoyoung Lee1, 10Suhwan Park1Seunghan Lee2Jun Seo2Jaehoon Lee2 Sungdong Yoo2Minjae Kim2CheolWon Na3Zhangyang Wang4Zach Golkhou5 Minkyu Kim6Sotirios Sabanis7Alejandro Lopez\-Lira8Dhagash Mehta9 Soonyoung Lee2Chanyeol Choi10Wonbin Ahn2,†Yongjae Lee1, 10† 1UNIST2LG AI Research3Sungkyunkwan University4University of Texas at Austin 5J\.P\. Morgan Chase6State Street7University of Edinburgh8University of Florida 9BlackRock10LinqAlpha

###### Abstract

Financial decision\-makers face more information than they can directly inspect, making context compression necessary\. Yet when large language models \(LLMs\) compress financial source material, they can alter the investment judgment supported by the original source\. We frame this problem as information fidelity: compression loses fidelity when it changes the decision induced by the source\. In agentic systems, such losses may recur across intermediate steps and amplify throughout the decision process\. Across financial filings and earnings\-call transcripts, we find that LLM\-based compression can produce fluent and factually plausible compressed contexts that nevertheless alter downstream decisions\. We analyze two diagnostic patterns associated with fidelity loss: decontextualization, where salient evidence is retained but separated from the caveats and contextual qualifiers needed for correct interpretation, and model dependency, where different compressors expose different views of the same source\. We then propose Agentic Context Compression, which generates multiple candidate compressions and audits their disagreements against the original source\. Our results suggest that financial compression should be evaluated not only by efficiency or factuality, but also by its ability to preserve decision\-relevant context\.

When Summaries Distort Decisions: Information Fidelity in LLM\-Compressed Financial Analysis

Hoyoung Lee1, 10Suhwan Park1Seunghan Lee2Jun Seo2Jaehoon Lee2Sungdong Yoo2Minjae Kim2CheolWon Na3Zhangyang Wang4Zach Golkhou5Minkyu Kim6Sotirios Sabanis7Alejandro Lopez\-Lira8Dhagash Mehta9Soonyoung Lee2Chanyeol Choi10Wonbin Ahn2,†Yongjae Lee1, 10†1UNIST2LG AI Research3Sungkyunkwan University4University of Texas at Austin5J\.P\. Morgan Chase6State Street7University of Edinburgh8University of Florida9BlackRock10LinqAlpha

††footnotetext:†Correspondence to:[yongjaelee@unist\.ac\.kr](https://arxiv.org/html/2606.29251v1/[email protected]);
[wonbin\.ahn@lgresearch\.ai](https://arxiv.org/html/2606.29251v1/[email protected])## 1Introduction

![Refer to caption](https://arxiv.org/html/2606.29251v1/x1.png)Figure 1:Compression\-induced decision flip\.Context compression can cause a decision\-maker to reach a different decision from the one supported by the original source\. For example, information that supports abullishdecision in the original source may lead to abearishdecision after compression\.Large language models \(LLMs\) are increasingly used to compress long documents into shorter contexts before those contexts are passed to another model or a human decision\-maker\. In high\-stakes domains such as finance, the resulting short context is useful only if it preserves the source evidence needed to support the decision\.

Long\-context compression is also inherently open\-ended\. A financial source text has no single correct summary because different summaries can select different evidence while remaining fluent and plausible\. This paper studies a simple but consequential failure mode: an LLM\-compressed representation can lead a decision\-maker to make a financial judgment that differs from the judgment supported by the full source text, as shown in Figure[1](https://arxiv.org/html/2606.29251#S1.F1)\. The issue is not limited to hallucination or factual inconsistency\. It also depends on which evidence remains visible, how it is contextualized, and how competing signals are balanced\.

Existing compression evaluations often measure reconstruction, factual consistency, or surface\-level information preservation\(Łajewskaet al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib17)\), but they do not directly test whether compression preserves the decision supported by the original source\. We frame this problem as information fidelity, where a compressed representation has high information fidelity if it preserves the decision implied by the original source\. A summary may retain headline facts but omit the caveats or offsets needed to interpret them, remaining factually correct while inducing a different decision\.

We examine two structured patterns that help characterize fidelity loss\. Decontextualization builds on work showing that isolated claims need context to remain meaningful\(Choiet al\.,[2021](https://arxiv.org/html/2606.29251#bib.bib20); Gunjal and Durrett,[2024](https://arxiv.org/html/2606.29251#bib.bib21)\)\. Compression may preserve headline claims while stripping away decision\-relevant caveats, comparisons, or expectation framing\. When such context is removed, a factually correct summary can still carry a different decision implication\.

Model dependency is motivated by evidence that investment analyses vary across LLMs\(Leeet al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib11)\)\. In our setting, the relevant variation is not only how models analyze the same evidence, but also which evidence each compressor preserves from the same source\. These patterns suggest that decision change is not merely random degradation\. Instead, it can reflect systematic differences in the evidence made visible after compression\.

These observations motivate a source\-grounded compression approach that preserves contextual qualifiers and compares multiple compressed views of the same source\. We instantiate this approach as agentic context compression, in which an agent generates competing compressed contexts and audits their disagreements against the original source\.

We evaluate information fidelity and compare several compression methods on quarterly filings and earnings\-call transcripts\. Our main contributions are as follows:

- •We introduce*information fidelity*, a decision\-centered criterion for evaluating LLM context compression against the original source\.
- •Using real\-world quarterly filings and earnings\-call transcripts, we show that LLM\-based compression can alter decisions, with decontextualization and model dependency emerging as recurring diagnostic patterns\.
- •We propose agentic context compression, a source\-grounded procedure that audits multiple candidate compressed contexts, and show that it reduces compression\-induced decision flips\.

## 2Related Work

### 2\.1Context Compression

Context compression has become a central problem in LLM\-based systems\. Prior work shows that LLM\-mediated representations can degrade source information during agentic context updating\(Zhanget al\.,[2026](https://arxiv.org/html/2606.29251#bib.bib2)\)and delegated document workflows\(Labanet al\.,[2026](https://arxiv.org/html/2606.29251#bib.bib3)\)\. Compressed contexts may preserve the general topic while dropping essential details, motivating grounding and information\-preservation metrics beyond efficiency\(Łajewskaet al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib17)\)\. A related line asks what compression should retain: LLM salience aligns only weakly with human judgment\(Trieneset al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib23)\), and conformal summarization formalizes coverage of critical content in high\-stakes settings\(Kuwaharaet al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib22)\)\. Even recent generative prompt compression that targets coherence and key information\(Zhanget al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib30)\)is evaluated mainly by task quality or summary metrics, not by preservation of the source\-implied decision\.

### 2\.2Summarization Distortion

Repeated generation through LLMs can cumulatively distort source information\(Mohamedet al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib1)\)\. Summarization can likewise distort a source by selectively reweighting content and failing to preserve perspectives or polarity balance\. FairSumm measures gaps between source and summary perspectives\(Zhanget al\.,[2024](https://arxiv.org/html/2606.29251#bib.bib6)\); opinion summarization can amplify majority polarity and weaken minority polarity\(Leiet al\.,[2024](https://arxiv.org/html/2606.29251#bib.bib7)\); and LLMs may implicitly resolve disagreement during generation, suppressing minority viewpoints\(Aghaebeet al\.,[2026](https://arxiv.org/html/2606.29251#bib.bib8)\)\. LLM\-generated news summaries also introduce framing through selective emphasis and omission\(Pastorino and Moosavi,[2025](https://arxiv.org/html/2606.29251#bib.bib29)\)\.

LLM\-generated summaries can also change readers’ decisions while remaining fluent and plausible\.Alessaet al\.\([2025](https://arxiv.org/html/2606.29251#bib.bib27)\)find that LLM\-generated content can induce framing and primacy biases and shift human purchase decisions, whilePeters and Chin\-Yee \([2025](https://arxiv.org/html/2606.29251#bib.bib28)\)find that LLM summaries of scientific studies often omit scope\-limiting details and overgeneralize conclusions\.

Together, these studies show that summarization is not neutral compression, but a process of selecting and aggregating perspectives\(Mayilvaghananet al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib18)\), where selective omission can itself be a form of bias\(Zhukovaet al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib24)\)\. We extend this insight to financial source texts by asking which decision\-relevant evidence is preserved\.

### 2\.3Financial LLM Decision Bias

Research on financial LLMs shows that the investment judgments they produce can be systematically biased by company names, asset classes, prior knowledge, or evaluation design\.Glasserman and Lin \([2023](https://arxiv.org/html/2606.29251#bib.bib9)\)show that company names and general knowledge can interfere with financial sentiment prediction\.Nakagawaet al\.\([2024](https://arxiv.org/html/2606.29251#bib.bib10)\)quantify company\-specific bias in financial sentiment analysis\.Leeet al\.\([2025](https://arxiv.org/html/2606.29251#bib.bib11)\)analyze model\-specific investment preferences and confirmation bias in LLM investment analysis\.Konget al\.\([2026](https://arxiv.org/html/2606.29251#bib.bib12)\)place these issues in a broader evaluation framework and argue that financial LLM evaluation must explicitly account for bias\.

## 3Information Fidelity

### 3\.1Problem Formulation

#### Task\.

We define information fidelity as source\-relative decision preservation: a compression has high fidelity when the decision\-model belief induced by the compressed text remains close to the belief induced by the original source\. Given a source textss, a compression modelℳ\\mathcal\{M\}produces a compressed textcc, measured as the number of bullets, must respect a fixed budgetBB:

c=ℳ​\(s\)subject to\|c\|≤B\.c=\\mathcal\{M\}\(s\)\\quad\\text\{subject to\}\\quad\\lvert c\\rvert\\leq B\.\(1\)We holdBBfixed across methods so that fidelity differences reflect what each method retains rather than how much it writes\. The realized token ratioτ=tok​\(c\)/tok​\(s\)\\tau=\\mathrm\{tok\}\(c\)/\\mathrm\{tok\}\(s\)is reported descriptively\.

A decision modelℰ\\mathcal\{E\}maps any text to a belief distribution over the label setV=\{bear,neutral,bull\}V=\\\{\\text\{bear\},\\text\{neutral\},\\text\{bull\}\\\}, that is,ℰ​\(⋅\)∈Δ​\(V\)\\mathcal\{E\}\(\\cdot\)\\in\\Delta\(V\), the probability simplex overVV\. For each fixed text, we averageR=3R=3independent decision runs, yielding a source beliefpsp\_\{s\}and a compressed beliefpcp\_\{c\},

ps=1R​∑r=1Rℰ\(r\)​\(s\),pc=1R​∑r=1Rℰ\(r\)​\(c\),p\_\{s\}=\\frac\{1\}\{R\}\\sum\_\{r=1\}^\{R\}\\mathcal\{E\}^\{\(r\)\}\(s\),\\qquad p\_\{c\}=\\frac\{1\}\{R\}\\sum\_\{r=1\}^\{R\}\\mathcal\{E\}^\{\(r\)\}\(c\),\(2\)wherec=ℳ​\(s\)c=\\mathcal\{M\}\(s\)is fixed andℰ\(r\)\\mathcal\{E\}^\{\(r\)\}is therr\-th decision\-model run\. Their top decisions are the most likely labels

v^x=arg​maxv∈V⁡px​\(v\),x∈\{s,c\}\.\\hat\{v\}\_\{x\}=\\operatorname\*\{arg\\,max\}\_\{v\\in V\}p\_\{x\}\(v\),\\qquad x\\in\\\{s,c\\\}\.\(3\)
We quantify its loss with two metrics\. The Decision Flip rate is the fraction of source documents whose top decision changes after compression,

Flip=1N​∑i=1N𝟏​\[v^c\(i\)≠v^s\(i\)\],\\mathrm\{Flip\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\mathbf\{1\}\\\!\\left\[\\,\\hat\{v\}\_\{c\}^\{\(i\)\}\\neq\\hat\{v\}\_\{s\}^\{\(i\)\}\\,\\right\],\(4\)
whereiiindexes theNNsource documents, while Total Variation Distance \(TVD\),dTV​\(pc,ps\)d\_\{\\mathrm\{TV\}\}\(p\_\{c\},p\_\{s\}\), measures how far the belief distribution moves even when the top decision is unchanged,

dTV​\(pc,ps\)=12​∑v∈V\|pc​\(v\)−ps​\(v\)\|\.d\_\{\\mathrm\{TV\}\}\(p\_\{c\},p\_\{s\}\)=\\frac\{1\}\{2\}\\sum\_\{v\\in V\}\\bigl\\lvert p\_\{c\}\(v\)\-p\_\{s\}\(v\)\\bigr\\rvert\.\(5\)
The flip rate captures outright decision changes, while TVD captures smaller belief shifts that may not cross the decision boundary\. We report TVD as its mean over source texts for each method\.

![Refer to caption](https://arxiv.org/html/2606.29251v1/x2.png)Figure 2:Decision change under one\-shot compression\. Bars show Decision Flip and source\-relative TVD for four compressors across MD&A and earnings calls\. Dashed lines show the decision\-model randomness floor: the same decision model reads the original source again, with no compression\.

### 3\.2LLM Context Compression

We realize the compression operatorℳ\\mathcal\{M\}with calls to a language model\. FollowingKanget al\.\([2025](https://arxiv.org/html/2606.29251#bib.bib19)\), we writeLLM​\(x;θ,𝒫\)\\mathrm\{LLM\}\(x;\\theta,\\mathcal\{P\}\)for a model with pretrained weightsθ\\thetathat maps an inputxxto an output under a natural\-language prompt𝒫\\mathcal\{P\}\.

#### Naive prompt\.

The naive baseline is a single passc=LLM​\(s;θ,𝒫comp\)c=\\mathrm\{LLM\}\(s;\\theta,\\mathcal\{P\}\_\{\\text\{comp\}\}\): the model receives the full source and a fixed bullet budget, then selects the facts it judges most material\. The full prompt is shown in Appendix[D](https://arxiv.org/html/2606.29251#A4)\.

#### Contextualization\.

Contextualization uses the same fixed budget but changes what counts as worth keeping\. Instead of choosing facts that look important on their own, it asks the compressor to preserve the details that make each material point interpretable: the relevant comparison, caveat, offsetting signal, or qualifier\. Under the same model, source, and budget, it uses𝒫ctx\\mathcal\{P\}\_\{\\text\{ctx\}\}to generate candidate contextsc~=LLM​\(s;θ,𝒫ctx\)\\tilde\{c\}=\\mathrm\{LLM\}\(s;\\theta,\\mathcal\{P\}\_\{\\text\{ctx\}\}\)\. This target also acts as the stopping condition: if a candidate spends the budget on standalone facts while missing the details needed to judge them, the compressor revisits the source and selects again\. The mechanism is unchanged; only the selection criterion shifts from isolated salience to interpretable evidence\.

### 3\.3Agentic Context Compression

We further consider Agentic Context Compression \(ACC\), which realizes the same operatorℳ\\mathcal\{M\}as a two\-stage agentic compression process\. First, the agent calls LLM candidate\-generation tools to produce contextualized candidate contextsck=LLM​\(s;θk,𝒫ctx\)c\_\{k\}=\\mathrm\{LLM\}\(s;\\theta\_\{k\},\\mathcal\{P\}\_\{\\text\{ctx\}\}\)from the same sourcess, wherekkindexes a distinct compression model\. Second, the agent audits disagreements among the setck\{c\_\{k\}\}against the source\. At steptt, it emits an actionat=LLM​\(ot,𝐡t−1;θ,𝒫agent\)a\_\{t\}=\\mathrm\{LLM\}\(o\_\{t\},\\mathbf\{h\}\_\{t\-1\};\\theta,\\mathcal\{P\}\_\{\\text\{agent\}\}\)from the latest tool observationoto\_\{t\}and history𝐡t−1=\(o1,a1,…,ot−1,at−1\)\\mathbf\{h\}\_\{t\-1\}=\(o\_\{1\},a\_\{1\},\\dots,o\_\{t\-1\},a\_\{t\-1\}\)\. For a claim or omission that differs across candidate contextsckc\_\{k\}, the action can grep the source for the relevant terms and then read only a short source snippet around the matching span, following the direct\-corpus\-interaction view of agentic search\(Liet al\.,[2026](https://arxiv.org/html/2606.29251#bib.bib14)\)\. The agent uses these snippets to test whether a candidate contextckc\_\{k\}overclaims the source, for example by making the evidence more bullish, bearish, or certain than the source supports, then scores eachckc\_\{k\}by overclaim risk and source\-locus coverage, and commits to the intact compressed context that the source best supports\. This connects ACC to multi\-agent factuality verification\(Ninget al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib26)\)and financial atomic\-claim grounding\(Guoet al\.,[2026](https://arxiv.org/html/2606.29251#bib.bib25)\), but targets decision\-changing omissions and overclaims rather than factuality alone\.

## 4Experiments

### 4\.1Experimental Setup

#### Datasets\.

Our main panel covers two source types from S&P 100 firms over fiscal 2025 Q1–Q3: 10\-Q MD&A sections \(N=300N=300\), collected from structured SEC filings following EDGAR\-Crawler\(Loukaset al\.,[2025](https://arxiv.org/html/2606.29251#bib.bib13)\), and quarterly earnings call transcripts \(N=297N=297\)\.

Table 1:Reliability of the decision model\. We report pairwise top\-decision agreement and Fleiss’κ\\kappaacross three independent reads of the same raw source\.
#### Decision model\.

The decision modelℰ\\mathcal\{E\}isGemini\-3\.1\-Flash\-Lite\. Given either a sourcessor compressed textcc,ℰ\\mathcal\{E\}produces the belief distribution used to computepsp\_\{s\}orpcp\_\{c\}under the same decision prompt\.

#### Baselines\.

We compare against compression baselines under the same budget\.Chunkingpartitionsssinto fixed\-size chunks, applies the same compression to each chunk, and concatenates the outputs\. ForToken Pruning, we use LLMLingua\(Jianget al\.,[2023](https://arxiv.org/html/2606.29251#bib.bib15)\), which drops low\-information tokens, and LongLLMLingua\(Jianget al\.,[2024](https://arxiv.org/html/2606.29251#bib.bib16)\), its long\-context extension; both useFinance\-Llama3\-8Bas the token\-scoring model\.Integratorobtains compressed texts from multiple models and combines them into one final context\.

![Refer to caption](https://arxiv.org/html/2606.29251v1/x3.png)Figure 3:Model\-specific decision movement under one\-shot compression\. Each panel shows one compressor; points are source\-relative movements for individual disclosures\. The horizontal axis shows signed decision shift from bearish to bullish, the vertical axis shows TVD magnitude, and the dashed line marks the noise floor\.

### 4\.2Fidelity Loss Under Compression

Compression changes the decision model’s output beyond rereading noise\. In Figure[2](https://arxiv.org/html/2606.29251#S3.F2), the dashed line is the decision\-model randomness floor: the same model rereads the original source, with no compression\. The model still flips on 11\.0% of MD&A documents and 8\.8% of earnings calls, so this floor captures rereading randomness alone\.

All four one\-shot compressors exceed this floor by more than a factor of two on both source types\. The excess is the compression\-induced component: summaries remove or reweight decision\-relevant evidence\. The parallel rise in TVD shows that compression also shifts the decision model’s belief distribution even without a top\-label flip\.

Table[1](https://arxiv.org/html/2606.29251#S4.T1)addresses the possibility that these flips come from an unstable decision model\. Across three independent reads of the same raw source, the model reproduces its top decision at a high rate, with Fleiss’κ\\kappain the substantial\-agreement range on both disclosure types\. This makes it unlikely that the gap in Figure[2](https://arxiv.org/html/2606.29251#S3.F2)is only a judge artifact, and supports interpreting the no\-compression line as a rereading\-noise floor\.

### 4\.3Model Dependency

All compressors lose fidelity in aggregate, but flip rate reduces movement to a binary outcome \(Figure[2](https://arxiv.org/html/2606.29251#S3.F2)\)\. We call this model dependency when the size or direction of belief shifts varies with the compressor\. To examine this pattern, we plot source\-relative movement by ticker for each compressor\.

Figure[3](https://arxiv.org/html/2606.29251#S4.F3)separates these shifts by compressor\. The horizontal axis shows direction, from bearish to bullish, and the vertical axis shows magnitude as source\-relative TVD\. If compression were model\-invariant, the panels would look similar\. Instead, compressors occupy different regions: some move tickers toward bearish decisions, others toward bullish decisions, and high\-TVD points differ across panels\. This pattern suggests that compression\-induced decision change is not merely information loss; single\-model summaries can carry a model\-specific tilt\.

If part of the tilt is compressor\-specific, aggregating models may reduce it\. The Integrator provides a simple check: it differs from single\-model compression only by merging a second model’s summary, and Table[4\.3](https://arxiv.org/html/2606.29251#S4.SS3)shows this alone lowers the MD&A flip rate from 33\.0% to 24\.7%\. ACC reduces it further by drawing on both models and checking their disagreements against the source\.

Table 2:Main results for information fidelity under compression\.No compressionis the identity floor, where the decision model rereads the source\. Flip denotes the rate at which compression changes the source\-based decision\. TVD denotes source\-relative probability drift\. Cost denotes total inference cost\.### 4\.4Decontextualization

We diagnose decontextualization with an offline fact inventory following FActScore and SAFE\(Minet al\.,[2023](https://arxiv.org/html/2606.29251#bib.bib4); Weiet al\.,[2024](https://arxiv.org/html/2606.29251#bib.bib5)\); details of the decomposition and role\-labeling prompts are in Appendix[D\.6](https://arxiv.org/html/2606.29251#A4.SS6)\. We split each source disclosure into atomic facts, link each fact to its source sentence, and tag it as headline, context, or boilerplate\. Headline facts carry the main investment signal; context facts calibrate that signal through caveats, offsets, comparisons, or expectation framing; boilerplate facts are background or procedural text\.

The second diagnostic pattern is decontextualization: under a fixed budget, compression often keeps headline results while dropping the caveats, offsets, comparisons, and forward\-looking qualifiers needed to interpret them\. Figure[4](https://arxiv.org/html/2606.29251#S4.F4)A shows that this stripping is selective: under MD&A compression, the context share falls from 25% in the source to 9% in the one\-shot compression\. This disproportionate loss can make a factually accurate summary read less calibrated than the source it came from\. The pattern is milder for shorter earnings transcripts, where more facts fit under the same budget and context is largely preserved\.

We then run an add\-back diagnostic on flip cases \(Figure[4](https://arxiv.org/html/2606.29251#S4.F4)B\)\. Starting from one\-shot summaries that flipped the decision model’s output, we append omitted facts and have the model reread the summary\. Restoring dropped context facts recovers the source decision in 37% of MD&A flips, compared with 16–19% for boilerplate or random facts from the same filing\. Earnings calls show the same ordering but a narrower gap: 33% for context versus 19–27% for the placebos\. This pattern is consistent with missing context being an important contributor, though the add\-back exceeds the budget and remains diagnostic rather than deployable\.

![Refer to caption](https://arxiv.org/html/2606.29251v1/x4.png)Figure 4:\(A\)Context facts are dropped much more under MD&A than earnings\-call\.\(B\)Re\-adding dropped context recovers more flipped decisions than boilerplate or random facts\.The gain is largest where the decontextualization diagnostic is strongest: for 10\-Q MD&A, where dropped context produces the clearest add\-back recovery, contextualization lowers the flip rate from 33\.0% to 21\.3%\. For earnings calls, where context is already better preserved under the budget, the gain is smaller, from 23\.9% to 22\.2%\.

## 5Industry Case Study

To examine whether the fidelity effects observed in our main experiments translate to production forecasting systems, we apply ACC in a globally deployed commercial equity\-forecasting product used by institutional investors across global markets\. The study is conducted on S&P 100 disclosures from fiscal 2025 Q1–Q3\.

Absolute product metrics are omitted due to commercial confidentiality\. Relative to the original disclosures, both naive compression and ACC reduced token consumption by more than 90%\. However, naive compression degraded forecasting IC, suggesting that compression\-induced decision changes can remove forecast\-relevant disclosure signals\. In contrast, ACC improved forecasting IC by 8\.3% over the original\-source baseline and by 23\.8% relative to naive compression, while reducing source\-relative decision flips by 59\.5%\.

These results suggest that preserving decision\-relevant context is important not only for compression quality, but also for retaining downstream forecasting signal in real\-world financial systems\.

## 6Conclusion

This paper examines fidelity loss in LLM context compression for financial decision\-making\. We show that compressed contexts can remain fluent and factually plausible while still changing downstream investment judgments\. Our results indicate that such loss often arises from decontextualization, where key caveats, offsets, or comparative framing are omitted\. This suggests that preserving isolated facts is insufficient when decisions depend on how those facts are framed and interpreted\. To address this issue, we propose ACC, which generates contextualized candidate compressions and audits them against the original source\. Under the same compression budget, ACC better preserves source\-relative decisions than one\-shot compression baselines\. Overall, our findings highlight that compression should be evaluated not only by efficiency or factuality, but also by its ability to preserve decision\-relevant context\.

## 7Limitations

This study has several limitations\. First, our experiments focus on financial disclosures from S&P 100 firms, so the findings may not directly generalize to other domains, firm types, or market conditions\. Second, fidelity is measured through a decision model rather than human analysts or realized market outcomes, meaning that the results depend on the chosen model, prompt, and label space\. Third, our setting uses a fixed compression budget and a three\-class investment schema, which may not capture other decision tasks or output formats\. Finally, ACC reduces fidelity loss but cannot recover information omitted by all candidate compressors\. Future work should evaluate broader domains, human judgments, and more flexible source\-grounded compression methods\.

## References

- Faithful Summarisation under Disagreement via Belief\-Level Aggregation\.External Links:2601\.04889Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p1.1)\.
- A\. Alessa, P\. Somane, A\. Lakshminarasimhan, J\. Skirzynski, J\. McAuley, and J\. Echterhoff \(2025\)Quantifying Cognitive Bias Induction in LLM\-Generated Content\.External Links:2507\.03194Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p2.1)\.
- E\. Choi, J\. Palomaki, M\. Lamm, T\. Kwiatkowski, D\. Das, and M\. Collins \(2021\)Decontextualization: Making Sentences Stand\-Alone\.Transactions of the Association for Computational Linguistics\.External Links:2102\.05169Cited by:[§1](https://arxiv.org/html/2606.29251#S1.p4.1)\.
- P\. Glasserman and C\. Lin \(2023\)Assessing Look\-Ahead Bias in Stock Return Predictions Generated By GPT Sentiment Analysis\.External Links:2309\.17322Cited by:[§2\.3](https://arxiv.org/html/2606.29251#S2.SS3.p1.1)\.
- A\. Gunjal and G\. Durrett \(2024\)Molecular Facts: Desiderata for Decontextualization in LLM Fact Verification\.External Links:2406\.20079Cited by:[§1](https://arxiv.org/html/2606.29251#S1.p4.1)\.
- D\. Guo, J\. Wu, and S\. M\. Yiu \(2026\)FinGround: Detecting and Grounding Financial Hallucinations via Atomic Claim Verification\.External Links:2604\.23588Cited by:[§3\.3](https://arxiv.org/html/2606.29251#S3.SS3.p1.12)\.
- H\. Jiang, Q\. Wu, C\. Lin, Y\. Yang, and L\. Qiu \(2023\)LLMLingua: compressing prompts for accelerated inference of large language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 13358–13376\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.825/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.825)Cited by:[§4\.1](https://arxiv.org/html/2606.29251#S4.SS1.SSS0.Px3.p1.1)\.
- H\. Jiang, Q\. Wu, X\. Luo, D\. Li, C\. Lin, Y\. Yang, and L\. Qiu \(2024\)LongLLMLingua: accelerating and enhancing LLMs in long context scenarios via prompt compression\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 1658–1677\.External Links:[Link](https://aclanthology.org/2024.acl-long.91/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.91)Cited by:[§4\.1](https://arxiv.org/html/2606.29251#S4.SS1.SSS0.Px3.p1.1)\.
- M\. Kang, W\. Chen, D\. Han, H\. A\. Inan, L\. Wutschitz, Y\. Chen, R\. Sim, and S\. Rajmohan \(2025\)Acon: Optimizing Context Compression for Long\-horizon LLM Agents\.External Links:2510\.00615Cited by:[§3\.2](https://arxiv.org/html/2606.29251#S3.SS2.p1.5)\.
- Y\. Kong, H\. Lee, Y\. Hwang, A\. Lopez\-Lira, B\. Levy, D\. Mehta, Q\. Wen, C\. Choi, Y\. Lee, and S\. Zohren \(2026\)Evaluating LLMs in Finance Requires Explicit Bias Consideration\.External Links:2602\.14233Cited by:[§2\.3](https://arxiv.org/html/2606.29251#S2.SS3.p1.1)\.
- B\. Kuwahara, C\. Lin, X\. S\. Huang, K\. K\. Leung, J\. A\. Yapeter, I\. Stanevich, F\. Perez, and J\. C\. Cresswell \(2025\)Document Summarization with Conformal Importance Guarantees\.External Links:2509\.20461Cited by:[§2\.1](https://arxiv.org/html/2606.29251#S2.SS1.p1.1)\.
- P\. Laban, T\. Schnabel, and J\. Neville \(2026\)LLMs Corrupt Your Documents When You Delegate\.External Links:2604\.15597Cited by:[§2\.1](https://arxiv.org/html/2606.29251#S2.SS1.p1.1)\.
- W\. Łajewska, M\. Hardalov, L\. Aina, N\. A\. John, H\. Su, and L\. Màrquez \(2025\)Understanding and Improving Information Preservation in Prompt Compression for LLMs\.External Links:2503\.19114Cited by:[§1](https://arxiv.org/html/2606.29251#S1.p3.1),[§2\.1](https://arxiv.org/html/2606.29251#S2.SS1.p1.1)\.
- H\. Lee, J\. Seo, S\. Park, J\. Lee, W\. Ahn, C\. Choi, A\. Lopez\-Lira, and Y\. Lee \(2025\)Your AI, Not Your View: The Bias of LLMs in Investment Analysis\.InProceedings of the 6th ACM International Conference on AI in Finance,ICAIF ’25,New York, NY, USA,pp\. 150–158\.External Links:[Document](https://dx.doi.org/10.1145/3768292.3770375),[Link](https://doi.org/10.1145/3768292.3770375),ISBN 979\-8\-4007\-2220\-2Cited by:[§1](https://arxiv.org/html/2606.29251#S1.p5.1),[§2\.3](https://arxiv.org/html/2606.29251#S2.SS3.p1.1)\.
- Y\. Lei, K\. Song, S\. Cho, X\. Wang, R\. Huang, and D\. Yu \(2024\)Polarity Calibration for Opinion Summarization\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),K\. Duh, H\. Gomez, and S\. Bethard \(Eds\.\),Mexico City, Mexico,pp\. 5211–5224\.External Links:[Link](https://aclanthology.org/2024.naacl-long.291/),[Document](https://dx.doi.org/10.18653/v1/2024.naacl-long.291)Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p1.1)\.
- Z\. Li, H\. Zhang, C\. Wei, P\. Lu, P\. Nie, Y\. Lu, Y\. Bai, S\. Feng, H\. Zhu, M\. Zhong, Y\. Zhang, J\. Xie, Y\. Choi, J\. Zou, J\. Han, W\. Chen, J\. Lin, D\. Jiang, and Y\. Zhang \(2026\)Beyond semantic similarity: rethinking retrieval for agentic search via direct corpus interaction\.External Links:2605\.05242Cited by:[§3\.3](https://arxiv.org/html/2606.29251#S3.SS3.p1.12)\.
- L\. Loukas, F\. Billert, M\. Fergadiotis, P\. Malakasiotis, and I\. Androutsopoulos \(2025\)EDGAR\-CRAWLER: From Raw Web Documents to Structured Financial NLP Datasets\.InCompanion Proceedings of the ACM on Web Conference 2025,WWW ’25,New York, NY, USA,pp\. 761–764\.External Links:[Document](https://dx.doi.org/10.1145/3701716.3715289),[Link](https://doi.org/10.1145/3701716.3715289),ISBN 979\-8\-4007\-1331\-6Cited by:[§4\.1](https://arxiv.org/html/2606.29251#S4.SS1.SSS0.Px1.p1.2)\.
- K\. Mayilvaghanan, S\. Gupta, and A\. Kumar \(2025\)Spot the BlindSpots: Systematic Identification and Quantification of Fine\-Grained LLM Biases in Contact Center Summaries\.External Links:2508\.13124Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p3.1)\.
- S\. Min, K\. Krishna, X\. Lyu, M\. Lewis, W\. Yih, P\. Koh, M\. Iyyer, L\. Zettlemoyer, and H\. Hajishirzi \(2023\)FActScore: Fine\-grained Atomic Evaluation of Factual Precision in Long Form Text Generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 12076–12100\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.741/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.741)Cited by:[§4\.4](https://arxiv.org/html/2606.29251#S4.SS4.p1.1)\.
- A\. Mohamed, M\. Geng, M\. Vazirgiannis, and G\. Shang \(2025\)LLM as a broken telephone: iterative generation distorts information\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 7493–7509\.External Links:[Link](https://aclanthology.org/2025.acl-long.371/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.371),ISBN 979\-8\-89176\-251\-0Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p1.1)\.
- K\. Nakagawa, M\. Hirano, and Y\. Fujimoto \(2024\)Evaluating Company\-specific Biases in Financial Sentiment Analysis using Large Language Models\.In2024 IEEE International Conference on Big Data,pp\. 6614–6623\.External Links:[Document](https://dx.doi.org/10.1109/BigData62323.2024.10826008),[Link](https://doi.org/10.1109/BigData62323.2024.10826008)Cited by:[§2\.3](https://arxiv.org/html/2606.29251#S2.SS3.p1.1)\.
- Y\. Ning, X\. Lin, F\. Fang, and Y\. Cao \(2025\)MAD\-Fact: A Multi\-Agent Debate Framework for Long\-Form Factuality Evaluation in LLMs\.External Links:2510\.22967Cited by:[§3\.3](https://arxiv.org/html/2606.29251#S3.SS3.p1.12)\.
- V\. Pastorino and N\. S\. Moosavi \(2025\)Frame In, Frame Out: Measuring Framing Bias in LLM\-Generated News Summaries\.External Links:2505\.05406Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p1.1)\.
- U\. Peters and B\. Chin\-Yee \(2025\)Generalization Bias in Large Language Model Summarization of Scientific Research\.External Links:2504\.00025Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p2.1)\.
- J\. Trienes, J\. Schlötterer, J\. J\. Li, and C\. Seifert \(2025\)Behavioral Analysis of Information Salience in Large Language Models\.External Links:2502\.14613Cited by:[§2\.1](https://arxiv.org/html/2606.29251#S2.SS1.p1.1)\.
- J\. Wei, C\. Yang, X\. Song, Y\. Lu, N\. Hu, J\. Huang, D\. Tran, D\. Peng, R\. Liu, D\. Huang, C\. Du, and Q\. V\. Le \(2024\)Long\-form Factuality in Large Language Models\.InAdvances in Neural Information Processing Systems,External Links:[Link](https://openreview.net/forum?id=4M9f8VMt2C)Cited by:[§4\.4](https://arxiv.org/html/2606.29251#S4.SS4.p1.1)\.
- Q\. Zhang, C\. Hu, S\. Upasani, B\. Ma, F\. Hong, V\. Kamanuru, J\. Rainton, C\. Wu, M\. Ji, H\. Li, U\. Thakker, J\. Zou, and K\. Olukotun \(2026\)Agentic context engineering: evolving contexts for self\-improving language models\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=eC4ygDs02R)Cited by:[§2\.1](https://arxiv.org/html/2606.29251#S2.SS1.p1.1)\.
- T\. Zhang, Y\. Wang, and D\. Z\. Wang \(2025\)SCOPE: A Generative Approach for LLM Prompt Compression\.External Links:2508\.15813Cited by:[§2\.1](https://arxiv.org/html/2606.29251#S2.SS1.p1.1)\.
- Y\. Zhang, N\. Zhang, Y\. Liu, A\. Fabbri, J\. Liu, R\. Kamoi, X\. Lu, C\. Xiong, J\. Zhao, D\. Radev, K\. McKeown, and R\. Zhang \(2024\)Fair Abstractive Summarization of Diverse Perspectives\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),K\. Duh, H\. Gomez, and S\. Bethard \(Eds\.\),Mexico City, Mexico,pp\. 3404–3426\.External Links:[Link](https://aclanthology.org/2024.naacl-long.187/),[Document](https://dx.doi.org/10.18653/v1/2024.naacl-long.187)Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p1.1)\.
- A\. Zhukova, T\. Ruas, F\. Hamborg, K\. Donnay, and B\. Gipp \(2025\)What’s in the News? Towards Identification of Bias by Commission, Omission, and Source Selection \(COSS\)\.External Links:2508\.02540Cited by:[§2\.2](https://arxiv.org/html/2606.29251#S2.SS2.p3.1)\.

## Appendix ABudget Sensitivity

As a brief robustness check, Figure[5](https://arxiv.org/html/2606.29251#A1.F5)repeats one\-shot compression with budgets from 5 to 40 bullets on a 150\-document subset\. Larger budgets reduce source\-relative TVD, especially from 20 to 40 bullets, but decision flips remain well above the no\-compression floor\. Extra space therefore smooths belief drift more reliably than it eliminates top\-label changes\.

![[Uncaptioned image]](https://arxiv.org/html/2606.29251v1/x5.png)![[Uncaptioned image]](https://arxiv.org/html/2606.29251v1/x6.png)

Figure 5:Budget sensitivity underGPT\-5\.4\-Minione\-shot compression\. Top: decision flips remain above the analyst noise floor even at larger budgets\. Bottom: source\-relative TVD declines with budget but plateaus above the floor\.

## Appendix BInter\-Model Agreement

Aggregate flip rates can hide whether compressors fail on the same filings\. Figure[6](https://arxiv.org/html/2606.29251#A2.F6)reports pairwise top\-decision agreement across four one\-shot compressors on the 597 source documents common to MD&A and earnings\-call Q&A\. The mean off\-diagonal agreement is 0\.75, so compressors disagree on roughly one filing in four even when their marginal flip rates are similar\. This supports treating compression fidelity as model\-dependent rather than as a property of the source alone\.

![[Uncaptioned image]](https://arxiv.org/html/2606.29251v1/x7.png)

Figure 6:Pairwise top\-decision agreement among one\-shot compressors on the same source filings\. Cells show the share of filings for which two compressors induce the same downstream bear/neutral/bull decision; off\-diagonal values near 0\.75 indicate that compressor choice changes the induced decision on about one quarter of filings\.

## Appendix CDecision Movement

Figures[7](https://arxiv.org/html/2606.29251#A3.F7)and[8](https://arxiv.org/html/2606.29251#A3.F8)show the same movement diagnostic separately for MD&A and earnings\-call Q&A\. The black vertical marker is the source\-implied stance for each ticker, and each colored arrow shows where a one\-shot compressor moves the downstream decision\. Wider horizontal spread within a ticker indicates stronger model dependence\.

![Refer to caption](https://arxiv.org/html/2606.29251v1/x8.png)Figure 7:Decision movement for MD&A disclosures\.![Refer to caption](https://arxiv.org/html/2606.29251v1/x9.png)Figure 8:Decision movement for earnings\-call Q&A disclosures\.
## Appendix DPrompt Templates

The boxed templates below reproduce the prompts used in our experiments\. Template variables are instantiated with the values used for the main results: the bullet budget isB=20B\{=\}20and word caps are stated per box\. Compression prompts are sent as system messages; the user message supplies the source document and, for the integrator, the model views\.

### D\.1One\-Shot Compression Prompt

One\-shot compressionSystem\.Compress the document below into exactly 20 bullets\. This count is STRICT; do not exceed it\.Each bullet must be on its own line starting with ‘\- ’\. Do not write headers or blank lines\.Each bullet captures exactly ONE fact or figure\. Do not combine multiple facts into a single bullet\.If the document contains more facts than 20, select only the 20 most material facts and omit the rest\.Output only the bullets\.User\.\[\{ticker\}\] \{company\} \{document\_label\}<document\>\{source\}</document\>

### D\.2Contextualization Compression Prompt

Contextualization compressionCompress the document below into exactly 20 bullets\. This count is STRICT; do not exceed it\.Each bullet must be on its own line starting with ‘\- ’\. Do not write headers or blank lines\.Each bullet captures exactly ONE decision\-relevant point\. Do not combine unrelated points into one bullet\.The 20 bullets together should be about 240\-250 words total\. Do not exceed 250 words, but do not make the summary much shorter unless the source lacks material decision\-relevant information\.A downstream decision model will make a bear / neutral / bull investment judgment from your bullets alone\. Your goal is to preserve the source’s decision direction and confidence as closely as possible\. The compressed bullets should induce the same overall investment interpretation as the full source\.Selection rules:\- Preserve the source’s dominant signal when the source is directionally positive, negative, or neutral\.\- Preserve facts that explain why that signal is strong, weak, improving, deteriorating, uncertain, or mixed\.\- When including a number, keep the comparison or frame that gives it meaning: prior year, prior quarter, guidance, consensus, management expectation, segment trend, margin impact, liquidity impact, or risk exposure\.\- Do not write standalone figures without explaining what they imply\.\- Preserve management’s framing when it changes interpretation\.\- Preserve forward\-looking guidance, demand signals, margin or cost pressure, capital allocation, liquidity, and material risks when they affect the investment judgment\.\- Include caveats or offsetting evidence only when they materially change the interpretation or confidence\.\- Do not mechanically include both positive and negative facts\.\- Do not force neutrality if the source has a clear directional signal\.\- Avoid boilerplate, immaterial details, and facts that are vivid but unlikely to affect the investment judgment\.Use only what the source actually says; do not add outside information\.Output only the bullets\.

### D\.3Multi\-LLM Integrator Prompt

Multi\-LLM integratorYou are a view\-only financial\-disclosure compression integrator\.You will receive one\-shot compressed summaries from different LLMs\. Treat the summaries as compressed model views, not as truth\. Integrate them into one 20\-bullet summary that keeps the decision\-relevant information visible across the views\.Rules:\- Use only the provided compressed model views and task metadata\.\- Do not use source decision labels, decision\-model probabilities, realized outcomes, or external knowledge\.\- Do not take a majority vote across views\.\- Do not invent facts to fill gaps that no view mentions\.\- If views differ, use cautious wording or keep both material perspectives\.\- Output exactly 20 bullets\.\- Each bullet must start with ‘\- ’ and capture one view\-supported fact or qualifier\.\- Output only the bullets, with no heading or commentary\.\- The entire 20\-bullet summary must total at most 330 words\. Keep every bullet concise; drop lower\-value detail rather than exceed the word budget\.

### D\.4Agent Context Compression

Agent runtime contractCandidates\.The agent adjudicates two one\-shot Contextualization summaries of the same disclosure\. Candidate A and candidate B are produced independently by two compression models\.Source access\.The rawsource\.txtis not placed in the prompt, and the agent does not read the disclosure end\-to\-end\. It uses targeted tools:get\_metadataidentifies the document and source length,grep\_sourcesearches for terms or claims where candidates disagree, andread\_source\_spanopens short local passages around retrieved evidence\. These inspections are span\-limited; the agent must perform at least four targeted source checks before deciding\.Audit criteria\.The audit focuses on disagreements between candidates and checks them against retrieved source spans for*overclaim risk*and*source\-locus coverage*\.Selection rule\.The agent must select exactly one candidate, A or B\. It compares the two candidates using retrieved source spans, chooses the candidate with lower overclaim risk, and uses broader source\-locus coverage as the deciding criterion when overclaim risk is comparable\.

### D\.5Decision Model Prompt

Decision modelSystem\.You are a financial decision model\. You are given a corporate disclosure \(or a compressed summary of one\)\. Do not use prior knowledge of the company or any external context\.Estimate the disclosure\-implied update, not a full investment recommendation\. Use ONLY the provided disclosure evidence\. Do not use valuation, stock price, prior returns, external company knowledge, or market context\.Based on the disclosure evidence alone, estimate how the disclosure would update an investor’s belief about the stock’s NEXT\-QUARTER total\-return regime:bear = disclosure evidence points to downside risk \(< \-5% if realized\)neutral = evidence is mixed, immaterial, or balanced \(\-5% to \+5% if realized\)bull = disclosure evidence points to upside potential \(\> \+5% if realized\)Output ONLY a JSON object with three probabilities that sum to 1\.0, no other text:\{"bear": 0\.xx, "neutral": 0\.xx, "bull": 0\.xx\}User\.Ticker: \{ticker\} \(\{company\}, \{sector\}\)Period of report: \{period\} \(\{q\_label\} \{year\}\)Disclosure type: \{level\_label\}Disclosure:\{material\}Disclosure\-implied probability distribution over next\-quarter regime \(JSON only\):

### D\.6Fact Decomposition and Role Labeling

Fact decompositionYou are a financial\-document fact extractor\. Decompose the 10\-Q MD&A below into atomic facts\. For each fact, return both the atomic claim and the source sentence it was extracted from\.Rules:\- Each fact is ONE atomic claim or figure\. Do not combine multiple claims\.\- Preserve quantitative detail verbatim \(numbers, percentages, dates, entities\)\.\- "source" is the full sentence \(or short adjacent span\) from the MD&A that contains the fact\. Copy it VERBATIM from the document – do not paraphrase or summarize\.\- Do NOT assign sentiment, polarity, or positivity/negativity\. Sentiment is labeled in a separate step\.\- Skip boilerplate \(forward\-looking disclaimers, table\-of\-contents text, GAAP definitions\) and strictly neutral structural sentences\.\- Return ONLY a JSON object: \{"facts": \[\{"content": str, "source": str\}\]\}\. No prose, no markdown fences\.

Role labelingYou are a financial fact\-role labeling assistant\.You are labeling atomic facts from a financial disclosure for an information\-fidelity study\.The study asks whether compression preserves the context needed to interpret headline investment signals\.You will receive a numbered list of atomic facts from one disclosure\. For each atomic fact, assign exactly one evidence\-role label:Use exactly one label:\- "headline": a primary investment signal that could directly move the decision model’s assessment\. This includes standalone facts about revenue, sales, demand, margins, profitability, costs, cash flow, guidance, growth, decline, market share, customer traction, operational performance, capital allocation, liquidity, or material risk when the fact itself is the main directional signal\.\- "context": a fact that calibrates how a headline should be read\. Label a fact as context only if it changes the interpretation of a headline signal by providing a caveat or limitation, offsetting evidence, year\-over\-year/sequential/segment/geography/customer/product comparison, magnitude/baseline/expectation\-relative framing, causal driver, timing/durability/one\-time nature, forward\-looking qualifier, or risk condition/dependency\.\- "boilerplate": generic background, corporate description, accounting/legal boilerplate, procedural text, vague management language, or scene\-setting that does not materially affect how a headline investment signal should be interpreted\.Important: Do not label a fact as context merely because it is detailed, neutral, hedged, explanatory, legal/regulatory, or generally relevant to the business\. A context fact must help interpret the direction, strength, scope, durability, cause, or expectation\-relative meaning of a headline\.Decision rule: Ask: "If this fact were removed, would the decision model interpret the relevant headline signal differently?" If yes, and the fact calibrates another signal, label it "context"\. If the fact itself is the main directional signal, label it "headline"\. If it does not affect interpretation of a decision\-relevant signal, label it "boilerplate"\.Tie\-break rules:1\. If a fact contains both a headline result and a qualifier that changes how the result should be read, choose "context"\.2\. If it is a standalone KPI/result/metric/guidance/capital\-allocation fact, choose "headline"\.3\. If it is generic, procedural, definitional, or only broadly relevant, choose "boilerplate"\.Output exactly one JSON object and no prose:\{"labels": \[\{"fact\_id": 0, "role": "<headline\|context\|boilerplate\>", "context\_function": "<none\|caveat\|offset\|comparison\|expectation\_framing\|causal\_driver\|timing\_durability\|scope\_magnitude\|risk\_condition\>", "target\_headline": "<short description or null\>", "rationale": "<one short sentence\>"\}\]\}