In brief: An AI analyst can accurately find a risk in a long document and quote it, but barely change the final recommendation. A new study of financial reports at context lengths from 2,000 to 128,000 tokens calls this the gap between retrieval and integration. So system acceptance testing should check not only retrieval and citations, but also the decision’s sensitivity to a material fact: if the fact is added, removed, or strengthened, the conclusion should change in the expected direction and in proportion to the risk.
This article is intended for executives, finance and risk teams, owners of document workflows, RAG teams, and AI engineers. We review a study published on August 25, 2026, and discuss pilot design. This is not investment, accounting, or legal advice. The experiment results do not prove the frequency of error in any specific company and do not automatically transfer from U.S. 10-K filings to contracts, procurement, or Russian financial reporting.
Contents
- What the new study found
- What the retrieval-integration gap is
- How the experiment was structured
- What results were obtained
- Why a citation does not yet prove analysis
- Why long context is not a guarantee
- Which architecture worked
- The SHIFT method for testing an AI analyst
- How to run a pilot on company documents
- Which metrics to track
- When a human is needed
- Study limitations
- Sources
- Frequently asked questions
- How AI Dawn helps evaluate document AI
- Conclusion
What the new study found
On August 25, 2026, Miao Liu and Zhizhe Liu published a preprint Reading Is Not Using. The authors separated two capabilities of an AI analyst: finding a specific risk disclosure in a document and using that risk in an investment judgment.
The main result is uncomfortable for standard RAG acceptance testing. In a long document, the model could identify a covenant threshold, settlement amount, and breach consequences, but its recommendation barely responded to those same facts. In other words, a correct answer to the question “what is written?” did not guarantee the correct reaction to the question “what does that imply?”.
This extends an earlier peer-reviewed result from Context Length Alone Hurts LLM Performance Despite Perfect Retrieval. In that study, across five open and closed models and tasks in math, question answering, and code, performance dropped by 13.9–85% as input grew, even with perfect retrieval. The new paper moves the problem from lab accuracy into how much weight the AI gives a fact in a decision.
What the retrieval-integration gap is
The retrieval-integration gap is a situation where a system can correctly retrieve a relevant fact, but that fact has little or no effect on the final judgment. Retrieval answers “did it find the evidence?” while integration answers “did it change the decision?”.
| Evaluation level | Question | What it confirms | What it does not confirm |
|---|---|---|---|
| Retrieval | Did the model find the needed passage? | the fact is available for a direct question | that the fact participated in the conclusion |
| Grounding | Is there a citation, and does it match the source? | traceability of the answer back to the document | the correct weight of the fact |
| Judgment direction | Did the conclusion change in the expected direction? | directional sensitivity of the decision | a sufficient magnitude of change |
| Judgment magnitude | Did the conclusion change in proportion to the strength of the fact? | calibration of the response in the test scenario | the correctness of the real business decision |
| Outcome | Did the decision help the process? | the observed effect in the given pilot | the universal benefit in other conditions |
That is why a report of retrieval accuracy = 100% can be true and still be insufficient. It measures an intermediate operation, not the delegated objective.
How the experiment was structured
The core design used annual reports from 12 U.S. registrants. For each company, the researchers created a quantitatively verifiable risk disclosure: for example, an obligation threshold, settlement payment, or reimbursement cap. Each target paragraph had five neutral same-length replacements.
The authors changed only the economically irrelevant surrounding text, extending context to 2,000, 8,000, 32,000, and 128,000 tokens. The material information about the selected company remained fixed. Impact was calculated as the difference between the judgment with the risk present and the average judgment for the five neutral replacements.
To make sure a simple paragraph insertion would not look like impact, the researchers also used 10 pseudo-insertions with no meaningful content and built an empirical noise baseline. Retrieval was checked with a separate call using pre-frozen answer keys. This way, the authors did not mix the ability to quote a fact with its causal contribution to the judgment.
Separate tests included three model families, a seven-level risk severity ladder, and 20 real disclosures from actual 10-K filings, where the source passage was surgically removed. The authors describe the final part as exploratory and capability-conditional, not as a universal confirmation.
What Results Were Obtained
| Observation | Short Context | Long Context | Inference Boundary |
|---|---|---|---|
| Risk Impact on the Main 9B Model | +3.2 pp sell probability at 2k | +0.4 pp at 128k; indistinguishable from an irrelevant insertion | specific model and experimental readout |
| Retrieval for the Main Model | 12/12 companies at 2k | 12/12 at 128k, with no false positives on neutral files | direct question about a frozen key |
| Response range across 7 risk levels | 0.23 | 0.04 | 5.6x compression |
| Pairwise severity ranking | 0.93 | 0.79 | order was preserved better than response magnitude |
| Real disclosures | +0.042 effect for 16 clearly negative excerpts | near 0 in the full document; retrieval 20/20 | exploratory, most capable open-weight model |
| Larger open-weight model | +0.105 at 2k | +0.034 at 128k | capability pushed the boundary farther out, but did not remove it |
For the production API model, risk increased the stated default probability by 6.06 pp at 2,000 tokens and by 3.15 pp at 128,000. On the full context, the effect remained distinguishable from zero, but it was no longer statistically distinguishable from the noise of an irrelevant insertion. The authors explicitly stress: these are two different checks, and a weak result against the noise control cannot be turned into an absolute "zero."
The most practical pattern is the separation of order and magnitude. At 128,000 tokens, the model still more often understood which risk was worse, but it almost stopped scaling the recommendation according to severity. For business, that means it may produce a plausibly ranked but economically flattened output.
Why a quote does not yet prove analysis
A quote solves an important task: it lets a person open the source and check whether the model made up a fact. But it does not show the internal or system-level weight of the evidence. The same fragment can be cited in an explanation and have almost no effect on the score, recommendation, or chosen action.
Four especially dangerous false jumps:
Fragment found → document understood.Retrieval can be accurate only for the explicitly named question.Fragment quoted → fact accounted for.The source is visible, but causal sensitivity is not measured.Risks listed → they are weighted.The model may rank risks correctly and compress the differences between them.The answer looks reasonable → the process is reliable.Plausibility does not replace a test on modified input.
Therefore business RAG architecture should have separate evals for retrieval and the final decision. One aggregated score hides where the error occurred.
Why long context is not a guarantee
The context window describes the maximum amount the model technically accepts under given conditions. It does not promise that every fact will retain the same influence on the final output.
In the study, more capable models preserved influence deeper into the context, but they did not become demonstrably invulnerable. For the largest open-weight model, the effect at 128k was roughly equal to the effect of the mid-sized model at 2k. One cheap commercial model on long input also lost part of retrieval: from 12/12 to 7/12, while the influence shrank even more.
Practical phrasing:
The stated context window is an input limit, not a certificate of effective decision context.
Adding a reasoning budget also was not a universal fix. In the tested modes, extended reasoning did not restore influence at 128k, and on short context it even reduced it. This does not prove that reasoning never helps; it shows that in this experiment, computation did not replace proper evidence routing.
What Architecture Worked
The authors compared several ways of bringing risk into the decision. The result depends not only on the model and the document, but also on the form of the intermediate representation.
| Workflow | What It Does | Result in the Study | Practical Risk |
|---|---|---|---|
| Direct reading | the entire document is fed in at once before the decision | influence fades with length | the retrieved fact loses weight |
| Chunk-and-summarize | each chunk is compressed, then the summaries are merged | the target fact often disappeared from the notes; influence was not restored | eviction due to a limited summary budget |
| Repeat raw excerpt | the original paragraph is placed before the question | restored very little | proximity without structure is insufficient |
| Targeted extract-then-decide | answers a specific risk question, structures the fact, and places it next to the decision while preserving the source document | restored impact better than the other options | quality depends on the question and fidelity of extraction |
In a fully synthetic context, targeted restatement increased impact to 8.5 percentage points, and all 12/12 companies moved in the expected direction. On 12 full real-world reports, this workflow delivered +0.043 with 11/12 expected directions, while direct reading was near zero. The experimenter-style structured extract worked no worse than self-extraction at a comparable format, so the benefit did not require the same model’s “introspection.”
But the result cannot be reduced to “just make a summary.” In this study, a generic summary actually removed a material risk. What was useful was a targeted structured restatement for a specific decision, placed next to the decision point while keeping the original document available for verification.
The SHIFT method for evaluating an AI analyst
SHIFT is a practical counterfactual test that checks not the beauty of the answer, but the causal link between evidence and conclusion.
| Step | Question | Artifact |
|---|---|---|
| S — Scenario | What specific decision does the system support? | decision statement, owner, allowed actions |
| D — Data | Which facts must affect the decision, and why? | evidence matrix with source, version, and expected direction |
| H — Hypothesis | What changes if a fact is removed, replaced with a neutral one, or strengthened? | paired counterfactual cases with no change in the rest of the context |
| I — Impact | Did the conclusion move in the right direction and by the right amount? | direction, magnitude, confidence/abstention, and citation diff |
| F — Failure boundary | When should the system stop or hand off to a human? | acceptance threshold, escalation, audit record, and stop condition |
The method does not require knowing the model’s hidden reasoning. It treats the system as a test object: change one meaningful input and observe the final output or action. That is closer to testing a business function than asking “find clause 7.3.”
How to run a pilot on company documents
1. Choose one decision
Do not start with the task “analyze all documents.” Pick one bounded outcome: flag contracts for mandatory legal review, draft a risk class, or prepare a list of exceptions for an analyst. Record the current baseline and the decision owner.
2. Freeze the corpus and answers
Collect approved historical or synthetic cases, document versions, and reference material facts. The team should choose the size of the set based on risk and diversity; the study does not give a universal number.
3. Build pairs
For each critical fact, create the original and a controlled modified version: removal, neutral replacement, strengthening, or weakening. All other decision-relevant information must stay the same. Otherwise, the shift cannot be attributed to the chosen fact.
4. Separate four evaluations
Measure retrieval, citation-to-source alignment, decision direction, and magnitude of change separately. Do not average them into a single score before analyzing errors.
5. Compare workflows
Run the same pairs through direct long context, the current RAG, generic summarization, and targeted structured extraction. Keep the model version, prompt, retriever, reranker, chunking, tool graph, and decision template.
6. Test long and dense cases
Add only controlled irrelevant or genre-similar context. Test document conflicts, tables, OCR, late appendices, and updated versions separately. These scenarios were not fully covered by the new study, but they are critical for a real document workflow.
7. Bring in a domain expert
The expert confirms not style, but expected direction, materiality, acceptable magnitude, and cases where there is no clear answer. Mark disputed pairs as requires review, rather than forcing the model to guess.
8. Run shadow mode
The AI produces a recommendation alongside the live process, but does not trigger an irreversible action. After the frozen evaluation, compare discrepancies, overrides, and missed material facts. A separate article on asynchronous AI agentscovers long-task architecture and independent result verification in more detail.
Which metrics to track
| Metric | Formula or rule | What it shows |
|---|---|---|
| Retrieval accuracy | correctly extracted facts / verified facts | whether the system finds the evidence |
| Citation fidelity | citations that match the source / verified citations | whether it distorts the source |
| Directional sensitivity | pairs with expected direction / valid counterfactual pairs | whether the decision responds correctly |
| Magnitude calibration | deviation of actual shift from the stated scale | whether the system flattens severity |
| Evidence survival | critical facts that reached the decision representation / required facts | where summary/RAG loses signal |
| Unsupported movement | shifts on neutral replacements / neutral controls | how much noise the pipeline itself creates |
| Abstention adequacy | correct requires review / ambiguous cases |
can the system avoid unnecessary decisions |
| Human override rate | material edits / verified decisions | where the person does not accept the output |
The threshold for each metric is approved by the process owner together with the domain expert, risk/compliance, and the technical team. The study does not set a universal pass rate and does not allow claims of financial impact.
NIST AI RMF Core recommends documenting test sets, metrics, and deployment-like conditions, as well as monitoring system behavior in production. The draft published on August 7, 2026, TEVV-Athlon also focuses on real-world impact and outcomes rather than a single convenient proxy. These documents are voluntary and do not certify any specific AI analyst.
When human review is needed
Human review is needed not because “AI is always bad,” but because the cost of error, reversibility, and authority differ. A person should confirm the conclusion or action when:
- the decision affects money, rights, obligations, access, or reporting;
- the document contains conflicting versions or an ambiguous rule;
- a critical fact is found, but the causal test shows a weak or unstable response;
- the action is irreversible or exceeds the approved threshold;
- evidence survival, citation fidelity, or directional sensitivity fail the local threshold;
- the model cannot decline appropriately.
For the financial sector, the Bank of Russia recommended on June 16, 2026, that an employee confirm operations in critical processes with high information security risk, including payment transactions. This is a specific information security recommendation from the regulator, not a universal rule for all document assistants; the company should determine applicability with its responsible functions.
Study limitations
The paper Reading Is Not Using — very preliminary draft, published as arXiv v1. The authors plan to release a replication package around the end of September; as of the date of this article, it is not yet available. The main design uses 12 U.S. companies and research disclosures, and the experiment with 20 real disclosures is described as exploratory.
The behavioral gap repeated across several model families, but the causal analysis of channels was performed on a single hybrid architecture. Models, prompts, numerical readouts, and context settings limit portability. One post-trained 9B model showed an unexplained dependence of the sign of the effect on document placement; the authors do not hide this anomaly.
The study concerns financial disclosures and investment judgment. Applying the method to contracts, tenders, insurance cases, or compliance is our engineering hypothesis based on the general structure of “evidence → decision,” not a measured result of the paper. Production prevalence, user impact, cost, latency, ROI, Russian legal applicability, or the effect for a specific company have not been established.
Search volume, keyword difficulty, rankings, traffic, CTR, backlinks, and AI citations for this page are Unknown.
Sources
- Miao Liu, Zhizhe Liu. Reading Is Not Using: Retrieval, Judgment, and the Design of AI Financial Research Workflows, arXiv v1, 08/25/2026.
- Yufeng Du et al. Context Length Alone Hurts LLM Performance Despite Perfect Retrieval, Findings of EMNLP 2025.
- NIST. AI RMF Core: Measure, observed 08/26/2026.
- NIST. The TEVV-Athlon Framework for Evaluating AI Systems, initial public draft announcement, 08/07/2026.
- Bank of Russia. Recommendations on the Safe Use of Artificial Intelligence in Financial Services, 06/16/2026.
Frequently asked questions
If the AI quoted the document correctly, can you trust the conclusion?
A quote confirms that the source is available for verification, but it does not show how strongly the fact influenced the conclusion. For consequential use, you need a paired test: change one material fact and check the direction and magnitude of the decision shift.
Does a larger context window solve the problem?
Not necessarily. In the study, the model’s capability pushed the boundary, but the influence of the fact still decreased with length. The stated context window is an input limit, not proof that all facts are used equally.
Will a standard chunk-and-summarize approach help?
In the tested implementation, a generic summary performed worse: critical risk dropped out of the limited notes before the decision stage. That does not mean every summary is harmful; the summary needs to be tested for evidence survival and designed for the specific decision question.
How is targeted extraction different from RAG?
RAG finds relevant passages. Targeted extraction additionally turns facts into a structured representation for a specific decision and places it next to the decision step, preserving the source link. These layers can work together.
How do you test AI without access to chain of thought?
Use counterfactual pairs: change one material fact, keep everything else the same, and measure the shift in output. The method tests observable system behavior and does not require hidden reasoning.
What is the first test a company should run?
Choose one decision and one class of documents, record the baseline, mandatory facts, and stop condition. Create several validated pairs with one fact changed, then compare retrieval, citation fidelity, and decision sensitivity in shadow mode.
How AI Dawn helps validate document AI
AI Dawn can connect document analysis to a verifiable business decision:
- assess the process, sources, current baseline, permissions, and points of human confirmation;
- design a corporate knowledge base or RAG system with targeted extraction and traceability back to the document version;
- build an MVP AI agent or AI workspace with a structured decision step, audit trail, and escalation;
- prepare a frozen eval using the SDVIG method, integrate the system, and carry out testing, launch, team training, and support.
The safest first step is to choose one process, its baseline, sources, constraints, and acceptance criteria. Then test in shadow mode whether the decision changes when a material fact is deliberately changed. Discuss the project.
Conclusion
You should not approve a document AI analyst based only on search, citations, and a polished summary. These signals show that the system reads and explains, but they do not prove that the material fact was given the correct weight in the decision.
Test the full chain: source → retrieval → structured evidence → decision → human/action. The SDVIG method adds the missing counterfactual: if the fact changes, the conclusion should also change in the expected direction, and at the critical threshold the system should stop and hand the decision to a person.