acceptodds
Under review as a conference paper at ICLR 2027

What Did the Verifier Verify? Auditing Evidence Attribution in Scientific Agents

Abstract

Scientific agents can reproduce numerical results while assigning them to unsupported claims. We audit their verification pipelines by separating execution-evidence representation, original-claim representation, and deterministic checking. Historical workflows motivate a controlled follow-up with 24 assertions arranged in 12 workflow-level pairs, two model configurations, three review interfaces, and two repetitions. All interfaces can access the same execution artifacts. Luna's direct review makes one false authorization per repetition, whereas full structured review makes none before deterministic checking. GPT-5.5 direct review is correct on every item in both repetitions. Ordinary and specialized implementations of the same checking contract agree on all 192 structured decisions. Replacing extracted evidence with executor records repairs one workflow-level error in one compact-extraction run, without a repeated advantage over full structured review. The follow-up does not reproduce the extraction-induced degradation observed in exploratory cases. Our component-audit framing is a retrospective synthesis of these results, not confirmation of that hypothesis. It makes performance attribution auditable through matched ordinary controls, intermediate judgments, and field-level traces. In this controlled suite, the observed safety improvement is already present upstream of deterministic checking: evaluating only final verdicts would obscure both its location and remaining representation errors.

Primary area:
foundation or frontier models, including LLMs
Keywords:
scientific agents, claim verification, evidence attribution, agent evaluation, reproducibility

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QQW8OpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a fundamental question: when a scientific agent’s verifier deems a conclusion verified, what is it actually checking—numerical consistency, the alignment between evidence and conclusions, or the validity of the scientific conclusion itself? The authors propose a four-layered auditing framework: 1. Historical audit (can numerical self-consistency coexist with misattribution?); 2. Contract testing (do two different checkers implement the same formal contract?); 3. Exploratory comparison (which failures warrant follow-up frozen experiments?); and 4. Frozen follow-up (at which layer do observed errors or improvements originate?). The experiments yield several key findings: 1. Safety improvements from a comprehensive, structured review process emerge even before the rule-based checks are applied. 2. No significant performance difference is observed between the two rule-based checkers. 3. Utilizing raw execution logs can rectify localized errors in evidence extraction. 4. A correct final label does not guarantee the correctness of the intermediate representations. 5. The authors' initial hypothesis was not supported by the experimental results.

Critical strengths
  1. The research addresses an important problem, and the conceptual distinctions made are of practical value.
  2. The experimental design is well-structured, effectively preserving intermediate judgments for analysis.
Critical weaknesses
  1. There is a significant lack of engagement with related work. The paper should cite more existing research on failure modes in AI-based review and verification systems [1–8].

[1] When AI Co-Scientists Fail: SPOT—a Benchmark for Automated Verification of Scientific Research.

[2] FLAWS: A Benchmark for Error Identification and Localization in Scientific Papers.

[3] LLMs Assist NLP Researchers: Critique Paper (Meta-)Reviewing.

[4] AI Review Is a Systemic Risk to Peer Review: Toward a Blockchain-Supported Claim-Level Ledger for Accountability.

[5] PRISMM-Bench: A Benchmark of Peer-Review Grounded Multimodal Inconsistencies.

[6] Position: Stop Automating Peer Review Without Rigorous Evaluation.

[7] No Hidden Prompts Needed! You Can Game AI Peer Review with Presentation-Only Revisions.

[8] ProcessBench: Identifying Process Errors in Mathematical Reasoning.

  1. The methodological contributions and novelty require further substantiation. The phenomenon described is already known (having been documented in several existing studies, such as those concerning ScientistOne). The authors' contributions are not yet sufficiently demonstrated.

  2. The scale of the core experiments is very limited, consisting of only 4 data sources, 3 types of templates, and 12 workflows.

  3. There is no explanation as to why the comprehensive structured review process is effective.

Questions for authors
  1. Is it possible to incorporate a larger and more naturalistic test set?
  2. Can an independent expert verification be provided? The reference labels and the review of historical cases relied on AI assistance, lacking adjudication by independent domain experts.
3 comments
AuthorsOpenReview ↗

Response to QQW8

The revised contribution is evidence about where checking and repair break down. We follow a calculation, the model's description of it, the rule's decision, and the final answer to the original request. The larger studies show both useful corrections and cases where added checking fails.

  1. Related work and what this study adds

We inspected relevant primary sections for seven of your eight suggested works (p. 2):

  • SPOT and FLAWS locate scientific errors.
  • LLMs Assist NLP Researchers (ReviewCritique) studies deficiencies in review text using researcher annotations.
  • PRISMM-Bench studies evidence-claim records, multimodal inconsistencies, and preservation of meaning.
  • Position: Stop Automating Peer Review Without Rigorous Evaluation and No Hidden Prompts Needed! distinguish better presentation or higher review scores from better science.
  • ProcessBench locates reasoning errors beyond the final answer.

For AI Review Is a Systemic Risk to Peer Review, the claim-level ledger paper, the official workshop page confirms its title and three authors. Its primary full text remains unverified, so the technical comparison remains open. We also discuss CogEEGAgent, ClaimReceipt, Fisher-R1/P-Bench, extraction and alignment failures, and Steed et al.'s claim-to-code checks.

These works already establish many of the problems above. Our comparison asks which step changes the answer: the model's review, the checking rule, the supplied evidence, or a new analysis. We compare ordinary and RIEC checkers on the same records and requirements. Their agreement does not show a new checking advantage. Equal, reversed, and inconclusive results remain in the paper.

  1. Larger and more complex tests

E01 uses eight datasets, 24 workflows, and 48 assertions: 648 contexts and 5,184 judgments across three xhigh configurations. For Luna, writing full records gives 3/60 false approvals versus direct review's 0/60 when short evidence cards include a checklist. With the same facts listed field by field, the counts are 0/60 versus 1/60, with or without a checklist. Writing out the evidence therefore helps or hurts depending on the condition (p. 5).

E02 adds 19 research projects, 33 workflows, and 62 original sentences paired with 62 predefined changes. We group shared authors into 17 groups and also check other source relationships. Some sentences admit more than one reasonable reading. Ambiguous references stay visible, and 1,775 of 3,348 answers cannot be used as valid whole-sentence predictions. In all 12 main comparisons, structured reviews matched the draft reference answers less often. Once missing or unusable answers are allowed for, either review approach could score higher (p. 6).

E03 uses 11 projects: nine shared-author groups, or eight with an additional shared-article rule. Ten projects overlap with E02, giving 20 distinct projects across both studies. Tables include coded, simulated, illustrative, and measurement-derived data. Repeated reviews remain grouped with their sources.

All 11 initial answers meet their specified analysis requests, so their 297 branches test whether checking preserves correct work. A separate 162-branch layer starts from six deliberately wrong analyses. Three illustrative new executions restore the requested coefficient, denominator filter, or training-only scaler fitting. Complete results retain technical failures and incomplete stages (p. 7; examples p. 49).

  1. What sometimes makes checking useful

We retain the model's judgment before applying rules. Four E01 evidence records misstated how evaluation data were used. The ordinary rules withdrew none of the wrong approvals; replacing the evidence with actual execution records corrected 4/4 without changing the model's account of the sentence. These cases cover three workflows in one model/presentation setting. An extra validity flag also caused a correct model judgment to be lost. These traces show what went wrong and what information corrected it. The earlier exploratory drop in performance remains unreplicated, and its historical cause remains unresolved.

  1. Independent expert verification

This request remains open: there are zero completed independent domain-expert judgments. We report the current references as AI-assisted drafts. Source-linked materials and blank forms omit model answers and expected labels. Numerical checks and offline replay support the specified calculations; independent review of sentence meaning is still outstanding.

Reviewer_QQW8OpenReview ↗

lower my score from 2 to 1

The authors' use of AI to generate their rebuttal has resulted in a response consisting largely of vague and abstract statements that fail to address my concerns, such as how their work differs from the issues explored by ScientistOne. Therefore, I have decided to lower my score from 2 to 1.

AuthorsOpenReview ↗

A Final Clarification

I am a PIG. I only realized at the last moment that I had submitted my reviewer responses but had not submitted the revised manuscript.

This was my mistake, not an AI problem. At this stage, I used AI to help organize my rebuttal, but I did not ask it to check whether I had completed the submission process. Even a simple final check with AI could probably have prevented this stupid mistake. I take full responsibility.

Actually, the revised manuscript I prepared does include a direct comparison with ScientistOne.

In Section 2, we discuss ScientistOne and explain the differences between the two studies. ScientistOne focuses on building and checking evidence chains in autonomous research. Our work focuses on identifying where verification errors come from. We test what happens when only the evidence record is replaced while the claim stays unchanged.

In Section 7, we also explain how our repair experiments differ from ScientistOne. We test whether a repaired analysis actually answers the original research request, rather than simply producing a statement that matches the available evidence.

The revised manuscript also includes three extended studies, E01, E02, and E03, with additional experiments and results.

Unfortunately, I failed to submit this version. I understand that the reviewer could not see these changes. This was entirely my responsibility.

However, I would also like to raise one point about the ICLR reviewing rules.

ScientistOne is currently available as an arXiv preprint. The ICLR 2027 Reviewer Guidelines clearly state that authors are not required to compare their work with papers available only on arXiv, and that the lack of such comparisons cannot be a basis for rejection.

I agree that ScientistOne is relevant, and I have prepared a direct technical comparison. However, I believe that a missing comparison with an unpublished preprint should not itself be treated as grounds for rejection.

Again, I take full responsibility for my submission mistake. I apologize for it.

Reviewer_B5NMOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates evidence attribution in scientific-agent verification, arguing that correct final judgments may conceal errors in intermediate evidence and claim representations. The authors compare direct review, structured review, and deterministic checking to identify where verification errors originate. The problem is relevant, and the effort to distinguish intermediate errors from final verdict accuracy is valuable.

Critical strengths
  1. The problem itself is interesting.
Critical weaknesses
  1. The paper does not clearly explain the actual evaluation procedure. The reviewer function is introduced through an abstract input-output formulation without a concrete description in the main text. The four judgment categories have conceptual definitions, but the complete decision criteria are not explicit. This makes the method difficult to assess. An end-to-end example would be helpful.
  2. The main evaluation contains only 12 workflows and 24 paired assertions, covering three statistical analysis families. Most configurations achieve nearly perfect accuracy, leaving little room for failure analysis. The controlled assertions cannot efficiently reflect the complexity of realistic scientific claims.
  3. The authors claim that error can occur in intermediate representations even when final judgements are correct. However, the experiments do not provide a systematic charaterization of these failures. Furthermore, the analysis does not show that how specific misattribution errors occur and under which conditions deterministic checking can effectively correct those errors.
Questions for authors
  1. How are the incomplete or ambiguous representation handled?
  2. How is semantic fidelity validated beyond the predefined fields, and how can a reader independently establish that the reference labels reflect the original scientific claims?
2 comments
AuthorsOpenReview ↗

Response to B5NM

We now show what the model reads, what it writes, how each label is decided, and whether its final answer still addresses the original request. We also separate meanings covered by the listed fields from parts of a sentence that need further review.

  1. How a decision is made

The model reads the calculation and original sentence. Direct review gives a label and reason. Structured review also writes what the calculation did and what the sentence claims. We save the model's own decision before ordinary and RIEC rules compare those records (p. 2).

The rules first check formats and required information, then whether the calculation matches the statement:

  • UNKNOWN: necessary information is unavailable, or the listed fields cannot represent the case.
  • INSUFFICIENT: the evidence concerns a different outcome, coefficient/comparison, adjustment variables, tested hypothesis, or required use of evaluation data.
  • SUPPORTED: those details match and the sentence's rejection claim agrees with p <= alpha.
  • CONTRADICTED: the same details match but the rejection claim is opposite. A malformed answer is a format failure, not a correct UNKNOWN answer. Full rules appear on p. 12.

For example, the cyl coefficient in carb ~ cyl + mpg has p = .4800706408. At a .05 threshold, the test does not reject a zero coefficient. A sentence saying it does is CONTRADICTED. A sentence naming another coefficient is INSUFFICIENT; unavailable required adjustment information gives UNKNOWN. This walks through the procedure rather than adding another observed model failure.

  1. How wrong evidence passes a correct rule

Four Luna evidence records misstated how evaluation data were used. Both checkers approved the internally matching records. Replacing the evidence with the actual execution record corrected all four, leaving the model's account of the sentence fixed. These cases cover three workflows and three datasets (p. 5).

Another airquality answer matched all six measured evidence fields and five measured claim fields, yet an extra flag, e.valid=false, turned a correct SUPPORTED decision into UNKNOWN. This shows why the audit names the fields it checks. Ordinary and RIEC agreed in all 2,869 comparisons; both can receive the same mistaken account of a calculation.

  1. Meaning outside the fields

E02 links each sentence to quotations, code/output locations, omitted qualifications, and possible readings. In S015, “regardless of memory collection” can mean an association after accounting for collection, or an association within every collection. The mixed model gives different support for those readings (p. 49).

Whole-sentence references have 68 resolved AI-assisted draft labels, 54 ambiguous versions, and two unresolved versions. The narrower reading has 105 resolved, 17 ambiguous, and two unresolved. Both assess the same whole-sentence answers; the narrower result does not separately score individual claim parts. An ambiguous reference is separate from a settled UNKNOWN label. For E02, we have not separately checked that extracted fields faithfully reproduce the source. A match on listed fields leaves unrepresented clauses unchecked.

  1. More realistic tasks and repair

E02 adds 19 projects, 33 workflows, and 124 sentence versions in 486 contexts. Under the frozen rules, 1,775 of 3,348 answers cannot be used as valid whole-sentence predictions. In all 12 main comparisons, structured reviews matched the draft reference answers less often. Once missing or unusable answers are allowed for, either review approach could score higher (p. 6).

E03 adds 11 specified analysis requests. Every initial answer meets its request, so later branches test whether correct work is preserved. Six deliberately wrong starting analyses test repair separately. A repair must keep the requested object, data rows, and analysis conditions, and report values and a conclusion tied to what was actually run. The tables include technical failures, uncertain compliance with tool/data instructions, and unstarted stages; completion therefore combines task performance with those requirements (p. 7).

  1. Reliability of the reference answers

The anonymous supplement recomputes tables from saved scored rows. Full rescoring and numerical replay packages remain in the author archive. A separate NumPy/SciPy implementation checked 23 requests: 412 predefined comparisons matched; 26 later alias checks are separate; 36 records could not be compared directly. Reference decisions remain AI-assisted, with zero completed independent domain-expert judgments. Blank expert forms omit model answers and expected labels.

Thanks for your rebuttal and additional results. The clarification of the evaluation procedure makes the method easier to understand. I also appreciate the additional failure case analysis, showing that incorrect evidence representations can lead to incorrect verification results.

Hoever, the additional results do not provide under which conditions the failure occurs. The expanded E02 evaluation also shows that a large proportion of sentences are marked as ambiguous and over 50% of answers cannot be treated as whole-sentence predictions. These limitations make it difficult to compare the verification methods.

I will keep my score unchanged.

Reviewer_w5o9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper looks at post-execution attribution in scientific agents, which is whether a sentence written about an executed computation actually describes that computation and not just its numbers. The verifier is split into an evidence record from the execution, a claim record from the sentence, and a deterministic checker over the two (Section 3.1). The authors compare direct review, full extraction and compact extraction for two model configurations, with an ordinary checker and a RIEC checker on the same contract, plus an intervention which swaps the model’s evidence record for the executor’s record. Evidence comes in four tiers (Table 1): a historical audit of four data-to-paper runs, 114 constructed contract checks, 19 exploratory sessions on 42 cards, and a frozen follow-up of 24 assertions from 12 scikit-learn workflows run twice. The predeclared degradation hypothesis did not recur, the two checkers agree on everything, and the one gain for Luna is in the model’s own structured review, before the rules. The paper calls this a component audit and a reporting discipline, not a new method.

Critical strengths

The problem is real. The D2 case in Section 1 and Appendix B is exactly the failure which gets hidden in practice. The coefficient 34.9 and p of .0196 are reproduced correctly, the numeric check passes 17 out of 17, but the prose attaches them to policy change when the table says intercept. So separating the evidence record, the claim record and the rule (Section 3.1), and keeping the model’s provisional label as its own row in Table 2, is the right discipline. The restatement audit in Section 7.2 is actually the sharpest point for me, all 57 restatements pass the ordinary checker but only 3 keep the original testing object. That basically tells you a checker pass says nothing about task completion.

Critical weaknesses

There is almost no signal in the frozen follow-up, and the abstract reads as if there is. In Table 2 Luna D has 1 false authorization per repetition, every other row has 0, and GPT-5.5 is 24 out of 24 in every row. So the “safety improvement already present upstream of deterministic checking” in the abstract is one assertion, wf11 (Section 6), on one model at low effort. The “192 structured decisions” is 24 assertions times 2 models times 2 extraction interfaces times 2 repetitions, and the paper itself calls the ordinary versus RIEC agreement a conformance control, not a finding (Section 3.3, Section 5.2). The interesting effect is in the exploratory stage, where Luna low goes from 4 false authorizations under direct review to 15 under structured self-review (Table 4), and that is basically one session out of 19. The follow-up did not reproduce it and the paper never says why. So what is left is a reporting discipline plus a null result on 24 toy items from 4 scikit-learn datasets, with wf01/wf05 and wf03/wf07 sharing one OLS fit (Appendix E). I would like to see a follow-up which is large and hard enough to produce signal, or a diagnosis of why the exploratory degradation vanished.

Questions for authors

What exactly did the injected project memory contain in the twelve runs? Did it include the 42 exploratory cards, the degradation hypothesis, the checker code or the follow-up task definitions?

1 comment
AuthorsOpenReview ↗

Response to w5o9

The frozen follow-up did not support the broad safety claim in the original abstract. We have changed that claim and tested when asking a model to write out its evidence helps, fails, or changes the question being answered.

  1. The earlier drop in performance

The drop did not recur. The historical stages changed the tasks, prompts, memory, model settings, and timing together, so the records cannot identify one cause.

The new controlled study (E01) compares three ways of reviewing the same facts, two ways of presenting them, and an explicit checklist. Three xhigh configurations and three repetitions give 648 contexts and 5,184 judgments on eight datasets, 24 workflows, and 48 assertions. For Luna, full records produced 1/60 false approvals versus direct review's 2/60 with short evidence cards alone, but 3/60 versus 0/60 when those cards included a checklist. With facts listed field by field, the counts were 0/60 versus 1/60, with or without the checklist. Writing out evidence therefore helped in some conditions and hurt in others. Full results start on p. 5; all conditions and source checks are on p. 13.

  1. What the checking rules changed

We save the model's decision before applying automated rules. Ordinary and RIEC checkers receive the same records and apply the same requirements. They give the same decision in all 2,869 comparisons.

Four Luna answers wrongly approved unsupported statements because their evidence records misstated how evaluation data were used. Ordinary rules removed none of these approvals. Replacing the model-written evidence with the actual execution record corrected 4/4, keeping the model's account of the sentence unchanged. These were three workflows and three datasets, with 2/0/2 cases across repetitions. The rules needed an accurate account of the calculation before they could expose the mismatch. The tables also retain correct statements lost during checking and answers that could not be scored.

  1. The injected memory

All 12 frozen runs received the same recovered 7,426-character summary. In that readable summary:

  • The earlier 42-card benchmark, scores, and project framing were mentioned; complete cards and the answer table were absent.
  • The specific hypothesis that full extraction would produce more false approvals than direct review, and its stop/go threshold, were absent.
  • Checker code was absent; ordinary/RIEC outputs and benchmark work were mentioned.
  • The new 24-assertion task definitions and workflow IDs were absent.

The 38 exploratory child runs received a different, earlier 7,954-character summary. Some task messages remain encrypted. The recovered summaries therefore show what memory was present, without establishing complete isolation or a causal memory effect (p. 26).

We also corrected the timing record: the first known prompt/request hashes were recorded after the first dispatch and submission. All 12 first submissions were reconstructed byte for byte. The later luna_D_r2 edit stays separate: the main analysis uses 22/24, not the later 23/24. The partial exploratory run and seven structured retries also stay separate.

  1. More complex sentences and actual repair

E02 adds 19 research projects, 33 workflows, and 124 original/changed sentence versions, grouped by shared sources and authors. In all 12 main comparisons, structured reviews matched the draft reference answers less often. Once missing or unusable answers are allowed for, either review approach could score higher (p. 6).

E03 follows your point about a correct replacement answering the wrong question. All 11 initial answers already met their specified analysis requests, so later branches test whether checking preserves correct work. Six deliberately wrong starting analyses test repair separately. Examples rerun the requested coefficient, restore denominator filters, or keep evaluation data out of scaler fitting (p. 7; examples p. 49).

The original follow-up had 24 assertions, or 12 pairs. We correct that count while retaining your criticism about its limited coverage and signal. The abstract now follows the findings above.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

Thank you for the reviews. We have revised the paper to explain a simple problem: an AI can report the right numbers while answering the wrong question. We now start with a concrete example, follow the mistakes through three sets of tests, and show what successful checking and repair actually change.

In the historical example, 34.9 and p=.0196 were copied correctly. All 17 checked numbers matched their sources. But these two numbers belonged to the regression's constant term, while the text used them to describe a policy association. The mistake was which conclusion the numbers were used to support.

  1. Why can a wrong statement pass a check?

We compared direct review with reviews that ask the model to describe the evidence and the sentence separately. Asking for these details helped in some settings and produced more errors in others. The ordinary checker and RIEC gave the same decisions in all 2,869 comparisons where both were run with the same inputs and rules.

Four wrong statements still passed because the model's description of the analysis gave the wrong role to the evaluation data. When we replaced that description with the actual execution record, the checks stopped approving all four. The extracted description of the sentence stayed fixed. These four judgments came from three workflows and three datasets, using one model and review format. This shows where those corrections came from: the information given to the checker. Results are on p. 5; the full matrix starts on p. 13.

  1. What happens with sentences from real papers?

We added 62 original sentences and 62 predefined changes from 19 research projects and 33 workflows. One sentence says that a relationship holds "regardless of memory collection." This can mean an overall relationship after accounting for collection, or a relationship that holds within every collection. Those readings need different evidence. We retain the ambiguity rather than force one reference answer (example p. 49).

Detailed reviews had lower observed agreement with our reference answers in all 12 primary comparisons. But 1,775 of 3,348 assigned outputs were missing or unusable under the fixed rules. Allowing for these unknown answers leaves both directions of difference possible in every comparison. We therefore cannot determine which review approach is more accurate from these data (p. 6).

  1. When does a repair finish the original task?

In an earlier exercise, code rewrote 57 sentences to describe the analysis that had actually been run. All 57 became supported, but only three still referred to the test and conditions in the original request. This is why we now check the original request as well as the final sentence.

We tested actual analysis and repair on 11 new requests. All 11 initial analyses already met their requirements, so their 297 follow-up branches test whether review preserves correct work. A separate 162-branch test starts from six preset mistakes. Selected repairs recalculated the requested coefficient, restored a required population filter, or used training data alone to calculate scaling parameters. The full results also retain unsuccessful, incomplete, and technically ineligible attempts (p. 7; examples pp. 49-50).

The paper now includes a step-by-step checking example, complete results, source locations, the historical-memory audit, and a fuller discussion of related work. The anonymous supplement recomputes tables from saved scoring data; complete rescoring and numerical-check packages remain in the author archive. The earlier degradation was not reproduced. Our reference answers were prepared with AI assistance, and independent expert review remains outstanding.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.