- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper asks whether the record an LLM agent harness writes about its own run can serve as evidence for a reader who was not present. It calls a record evidentiary if it can be checked without trusting its writer. The paper makes three contributions.
Static audit. It audits 16 open-source harnesses on 10 source-level properties with an LLM instrument. None of the 32 “absent-reader” cells is present. Hearsay benchmark. Five harnesses run 14 tasks twice each (140 runs, claude-sonnet-5) in a sandbox that independently captures model traffic and process/file events. Three LLM examiners from three vendors read each record under seven material conditions and answer three questions: reconstruction (Q1), fault attribution (Q2), and alteration detection (Q3). Every quoted citation is mechanically labeled by who wrote the text it lands on: narration, corroborated, or independently witnessed. Second writer. This is an append-only, hash-chained ledger fed from the model and environment boundaries. It is reconciled against the native record by a forward query (omissions) and a per-format reverse check (fabrications).
(omissions) and a per-format reverse check (fabrications).
Main results:
Examiners attribute the fault in 74–91% of runs. Only 16–22% of their native-condition citations land on text the harness did not write, and all of that is the two benchmark anchors. The best low-false-positive examiner detects 48.7% of injected alterations at 5.2% FP. On write-time falsifications in aider, the second writer catches 28/28 while a self-computed hash chain catches none. Under paired scoring it reports 98/115 edited copies. Held out on another vendor’s runs, it catches 29/34 fabrications.
The study is preregistered with dated deviations and releases code, records, and alteration pairs.
Critical strengths
- The citation-provenance instrument is the paper’s most reusable idea, and it yields a non-obvious result.
Grading an LLM judge by who authored the text each citation lands on, rather than by its verdict, is a clean, mechanical, model-agnostic measurement. It transfers well beyond agent audit to any LLM-as-judge or LLM-as-auditor setting. It produces a result I would not have predicted: all three examiners overstate how much of their own evidence is corroborated. E1 claims 86.7% where the rule finds 68.8%; E2 and E3 overstate by 55 and 46 points. Examiners with similar fault accuracy rest on very different material: a +27.7-point E1–E2 gap in corroborated share. Combined with 82/100 human agreement on landing labels, this is solid evidence that “the judge was right” and “the judge’s answer was supported” must be measured separately. The underlying thesis, that a self-authored record is testimony and sealing it does not change that, is a structural claim that will survive model scaling.
- Process transparency is exceptional.
The preregistration appendix sets every registered item against what was delivered. It reports the failed predictions: reconstruction below band, alteration above chance, held-out fabrication recall below 90%, and the dissociation criterion failing cross-vendor. Every post-hoc measure is marked, and Table 30 gives each headline a second check. Multi-vendor examiners and graders, a 37-person human panel, and full artifact release make the paper auditable in exactly the sense it argues for. Most of the weaknesses below come from the authors’ own disclosures, which deserves credit regardless of the decision.
Critical weaknesses
W1. The headline dissociation is largely true by construction. The parts that are not rest on post-hoc, in-sample, or paired measures that overstate deployment performance.
(a) By-construction results. Under the native condition, the only text the harness did not write is the two anchors: the failing-test output and the final diff. By definition these say nothing about intermediate steps. So the abstract’s claim that fewer than one in ten citations about “what happened in between” land on harness-independent text (2–9% on Q1) is fixed by the design, as §5.2 itself concedes. The write-time contrast is similar. A hash chain over the harness’s own record passing 28/28 self-authored falsehoods is a definitional property of hash chains. A verbatim diff against a proxy log is close to guaranteed to catch injected fabrications that never reached the model. An informed expert would not bet against either result: they demonstrate the concept rather than test it.
(b) Paired, in-sample scoring. The empirical content that is not by construction is where evidence is thinnest.
The 98/115 is scored by paired difference against a clean twin of the same run. A real auditor never has that twin, and the reverse rules were written on those same pairs. Read stand-alone, the reverse check flags 52 of 95 clean copies, including 26/28 for mini-SWE-agent and 26/28 for cline. On held-out runs it still flags 21/22 and 21/21 clean copies for those two formats. Held-out fabrication recall (29/34, 85.3%) misses the registered 90%. Removals are invisible to the reverse check by design (0/35 held out). The deployment-relevant number is never a headline: unpaired detection vs. unpaired false positives, on held-out records, with rules frozen in advance.
(c) Post-hoc measures carry the headlines. Several were added after results were seen:
the independently witnessed share, which underlies both the “one in ten” claim and the ledger-shift claim; the majority-of-three grading rule, which superseded the registered grader 1; the 10% false-positive threshold; the dissociation criterion, which fails on the other-vendor runs (E1 witnessed share 31.1%, E3 accuracy 69%).
Meanwhile the registered primary outcome for the intervention, the corroborated share, barely moves. It rises only for E1 (+5.3), and only when ledger rows of the agent’s own words count; without them it is −1.4, and E2/E3 move −0.5/−1.6. The abstract’s statement that examiners “rest more of their citations on what the harness did not write” is therefore a post-hoc measure presented as a finding. Human graders were also stricter on native answers (12/25 correct vs. 18/25), so the 74–91% accuracies may run high.
(d) Little evidence of false records in the wild. There is roughly one natural case: a SWE-agent run that filed “Exit due to multiple consecutive command timeouts” as an assistant line the model never produced. The other “wrong clean record” came from the authors’ own run bound failing, which no harness could have seen. Beyond that, 7/140 records lack a last unit after crashes or timeouts. This is the finding that would make the problem pressing, and it is the least developed.
Overall, the paper claims at the “the intervention works” rung, but that evidence is in-sample, paired, from one trust domain, on 14 short tasks capped at 7 minutes, and the write-time test uses a single-turn harness.
W2. The “second author” is a second component, not a second party, so the evidentiary/hearsay framing overclaims.
Both writers, the database, and the proxy are run by the operator, who is explicitly trusted (§4.2, §7, Table 2). In the settings the paper invokes (disputes, investigations, regulatory audits under the EU AI Act’s logging duties), the party whose account is in question is usually the operator, not the harness code. Against a buggy harness, the ledger is a useful consistency check. Against an interested operator, it adds little beyond an operator-signed log. Appendix G notes the operator’s role can drop the append-only trigger or truncate the table. On the main grid, the harness was not confined to the proxy and shared a database role with the witness. So the independence assumption was unenforced for all 140 runs; it was enforced only in the 28-run re-run. “What makes a record evidence is who writes it” is right as a principle, but the delivered system changes which process writes, not which party. The one witness the paper itself names as outside the operator’s control, the model provider’s request log, is never tested.
I would want either a scope correction (from “evidentiary record” to “harness-independent consistency check”) or an experiment with a witness the operator does not control.
Presentation (secondary). The prose is extremely compressed and term-dense: located, corroborated, independently witnessed, narration, event vs. narration reading, clean twin, gate, floor/ceiling. Nearly every main-text sentence carries a qualifier resolved only in a 38-page appendix. The abstract is hard to parse on first reading, and Figure 1 and Table 1 pack in more numbers than a reader can absorb. Restructuring the main text around the registered predictions, which ones held, and the deployment-relevant numbers would help more than additional checks.
Questions for authors
I would raise my rating if Q1 shows stand-alone performance near the paired numbers, and the abstract is restated around the registered outcomes and the citation-provenance instrument.
(Could change soundness and rating.) Please report the reconciliation as a deployed auditor would use it. Freeze per-format rules and allowlists before seeing the data, score unpaired (no clean twin) on held-out records, and report detection by alteration type and false positives per format. If fabrication recall stays at or above 85% with stand-alone false positives at or below 10% on all four formats, my main concern about §6 is resolved. (Could change my significance assessment.) How often are deployed records wrong in the wild? You have 140 + 140 + 31 natural runs with a ledger. Please give a full census of query and reverse-check findings on unaltered records, classified as benign bookkeeping (e.g., the synthetic “action was not executed” observation), role mislabeling (the SWE-agent case), or true omissions after crashes. Can this be run at longer horizons, given that the 7-minute bound is short relative to production agent sessions? (Clarification, threat model.) In which concrete audit or dispute scenario is the harness the adversary while the operator is trusted? If the target is bugs and misfiling, would you reframe the contribution as a consistency check? If the target is an interested operator, what does the ledger add over an operator-signed log, and have you tried reconciling against provider-side request logs? (Clarification.) Detection in §5.3 depends on how subtle the injector was made. It was revised five times until E1’s model family could not pick the edited copy above chance, while another vendor’s judge still could (60.9%, p = 0.012). Can you report detection across injector revisions, so that “catches about half” reads as a curve over edit subtlety rather than a property of the record? (Clarification, §3.) Domain provenance (as-of time, entitlement, and cost as first-class fields) is still written by the harness. By your own definition, it cannot make a record checkable without trusting the writer, so why is it one of the two absent-reader properties? Relatedly, the registered instrument (claude-fable-5) was replaced by claude-sonnet-5, and human agreement on the 33 re-judged cells is 26/33, below your registered 80%. How should readers weight the activity correlation (ρ = 0.55, n = 16)? (Clarification.) Please state recruitment, compensation, and IRB approval or exemption for the 37 annotators; the ethics statement does not cover them.
Suggestion (does not affect my score): the strongest contribution is the citation-provenance instrument plus the self-report overstatement finding, not the 14-task grid. Leading with that, and treating the benchmark as its testbed, would position the paper outside an already crowded agent-benchmark space.
4 comments
Reply to Reviewer 3E3M (2/2): remaining points
W1(a). By-construction results. Agreed on all three. The "one in ten" claim is gone; §4 says the witnessed share "is set by the design". The hash chain passing 28 of 28 is presented in §2 as a demonstration that the one protection in use does not bind the writer, and the write-time test opens §6 as "the easiest test".
W1(b). Condition 1, in our first reply. Omissions remain invisible to the reverse check by design; the query reports them, 43 of 44 held out and 29 of 30 on the third vendor's runs.
W1(c). Post-hoc measures. The dissociation criterion and the 10% threshold are out of the body; the criterion's failures on both other sets of runs are in Appendix F. The majority rule and the witnessed share remain post hoc, and Appendix H says so. For the third vendor's runs we registered five predictions on these measures: four held and the dissociation criterion missed (Tables 21 to 23, 36). §4 says the accuracies "may run high" because human graders were stricter.
W1(d), Q2. Records wrong in the wild. The census you asked for is in Table 32: every finding of both checks on 395 unaltered records, classified by rules fixed in a script. In the three main sets, 33 of 365 would mislead a reader, 31 of them SWE-agent's. We state the base rate as it is: such records are a minority, and nothing in a record shows which they are. The SWE-agent fault is real and unprovoked: on its third-vendor runs, 37 of 412 tool calls the model sent are absent from the saved trajectory. Longer horizons were not run: the bound is reached on few runs and longer tasks need a new task set; §7 states the scope.
W2. A second component, not a second party. We did both things you asked. The experiment: the provider's stored copy of all 708 calls of the third vendor's runs is byte-equal to the ledger, and both checks reach the same findings from either on 84 of 84 runs. This is one provider; §7 says so. The enforcement: separate database roles and a network namespace on those 84 runs. Inside the namespace no outside connection succeeded. The scope correction: §7 says that where no outside copy exists, "the ledger is a consistency check under the operator's control and our claim is the narrower one". Fig. 8 shows the operator's rewriting as a row that only the provider's copy exposes.
Q4. Threat model. A harness that is untrustworthy while its operator is honest is the ordinary case of a faulty harness: it drops, misfiles or truncates what it records, with no intent anywhere. The census documents it, and "consistency check" is the right name for what the ledger does there. Against an interested operator the ledger adds nothing over an operator-signed log unless a copy sits outside the operator.
Q5. Detection against subtlety. We can give two things and not the third. Appendix E has the gate's accuracy at each of five rounds (83.7, 70.9, 64.5, 70.3, 57.5%) and under a second vendor's judge (60.9%). Appendix F has each examiner's detection by size of the altered unit (Tables 15, 16): E1 stays between 44% and 53%. Examiners read only the final revision, so we cannot give their detection at earlier rounds; a sentence claiming cruder changes are caught more often has been removed. §3 now states the five rounds in the body.
Q6. The audit. Provenance is no longer an outsider's property; the count is 0 of 16. The instrument swap is disclosed (Appendix B). A second instrument from another vendor agrees on fifteen of sixteen verdicts and gives partial credit on the sixteenth; neither finds one. Human agreement is 26 of 33, one short of the bar we registered, and the reported property has no human-checked cell; §2 says both. The activity correlation should be given no weight.
Q7. Annotators. The ethics statement now covers them: volunteers known to the authors, unpaid, consenting to anonymised release, no ethics-board review, our assessment of minimal risk reported in the statement.
Presentation. We followed your advice that restructuring would help more than additional checks, and did both. The terms you list are gone from the body or defined once where they first appear; Fig. 1 and Table 1 of the submission are replaced by eight figures and a table of record sets.
Most of the weaknesses you raised came, as you noted, from our own disclosures. We have tried to answer in the same spirit, and we would value your reading of whether the revised §6 now claims only what it shows.
Follow-up before the discussion closes
Dear Reviewer 3E3M,
Thank you again for the most exact review we received, and for stating what would change your assessment. Our reply and the revised manuscript were posted on 11 November.
You named two conditions. On the first, the evaluation now follows your protocol: rules committed before the records existed, one record read alone, results by type and by format. Across five formats, no clean record is flagged (0 of 140; your bar was at most 10%), and 40 of 42 fabrications are reported (your bar was at least 85%). One format, cline, is below the bar at 8 of 10, and we did not repair its rule after seeing the result. On the second, the abstract now leads with the citation-provenance result and reports the registered outcomes.
The discussion period closes in less than 24 hours. Could you let us know whether these results meet your conditions, or what remains open? If they do, we would be grateful if you would update your review to reflect this.
Thank you again for your time!
The Authors
Thank you for the detailed reply. I am raising my rating from 2 to 3.
The frozen-rule test resolves my main concern: 0 of 28 clean records flagged in each of the five formats, fabrication recall at my bar in three of four (cline 8 of 10), and no rule repaired after the result. The abstract now leads with the provenance measurement, as I asked.
For the final version, not conditions for my score:
- The abstract's "flags none of 84 clean records" holds for the reverse check only; the query reports real omissions on 15.
- State in the abstract that re-used genuine text passes the committed check (35 of 35) and that the operator is trusted.
- §6 should give opencode omissions as registered (3 of 8) beside the decoded 8 of 8.
Reply to Reviewer 3E3M (1/2): your two conditions
Thank you for the most exact review we received, and for stating what would change your assessment. Your protocol, rules frozen first, one record read alone, held-out records, results by type and by format, is now the paper's main evaluation: §6 and Fig. 7 are built on it, and the abstract reports its numbers and no longer the paired one. We answer your two conditions here, in your terms, and the remaining points in a second comment. "Items" refer to our first detailed comment to all reviewers; the table by format is in the second.
Condition 1: stand-alone performance, rules frozen first, held out, by type and per format.
Your bar: rules and exemptions frozen before the data, one record read alone, fabrication recall at or above 85%, false positives at or below 10%, per format. The test that meets the freezing without qualification is the one whose rules were committed to the repository before any record existed. It covered three formats; because your condition is per format, we ran the other two during the review, rules file untouched, predictions registered before each.
‡ Run during the review (Appendix G, Table 28). † Read with the database export decoded; see Appendix G.
Where this falls short of your bar, and why.
The held-out records (115 pairs, five formats) meet it in part. Stand-alone and paired readings agree pair for pair on every fabrication and rewrite (Table 25). The unit rules were never written on those records, but only three of the seven exemption classes were named beforehand: with those three, 37 of 115 clean records are flagged; with all seven, 1. Fabrications are 38 of 43 pooled, with mini-swe-agent 7 of 9 and cline 6 of 8 under your bar. §6 says this.
If there is a further element you would want fixed in advance, we would be glad to register it.
Condition 2: the abstract restated around the registered outcomes and the citation-provenance instrument.