Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026

EchoAudit Assessing reports from incomplete evidence

Nathan Samson · Team EchoAudit

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: EchoAudit Assessing reports from incomplete evidence

Open in a new tab (opens in new tab)
Share

An evidence auditor may receive an invoice report and only part of its supporting ledger. Missing records sometimes prevent a justified verdict, yet the visible subtotal can already disprove a claimed total. We constructed twelve ledger families that separate these situations while holding partial source evidence fixed. Each family supplies an unresolved report, two complete source versions with opposite correctness, and a partial report below the visible subtotal. Three scripted message contexts accompany each view, producing 144 assessments per configuration. Qwen3-8B in nonthinking mode returned valid structured outputs on all 144 calls and accurately assessed 71/72 complete source cases. It gave no warranted abstentions on 36 unresolved cases. None of 36 context pairs combined warranted abstention with rejection of the partial but refutable claim. A subsequently selected Claude Sonnet 5.5 configuration with adaptive thinking received the same instruction and evidence strings on the same families. It accurately assessed 65/72 complete source cases, but only 81/144 responses passed the unchanged output contract. Its joint partial score was 1/36, with 35/36 unresolved responses invalid. These output failures limit a semantic comparison between configurations. The diagnostic separates complete source reconciliation, evidence sufficiency and valid answer production within a small synthetic domain.

Reviews

Judging this Sprint?

Review “EchoAudit Assessing reports from incomplete evidence”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. The U / P- construction is a nice piece of design. Holding the visible records fixed while making one claim unresolved and another refutable from a lower bound cleanly separates "abstain because evidence is missing" from "reject because the visible subtotal already disproves the claim", and blanket abstention cannot game the joint metric. The frozen schedule, planned denominators, and the independent arithmetic check are good practice.

    Main limitations and suggestions:

    1. The link to collusion is thin. The S/T author-label manipulation was meant to be the bridge to coordination, but Qwen's ceiling/floor and Sonnet's invalid outputs leave it uninformative. As written this is an evidence-sufficiency / abstention diagnostic for auditors; consider leading with that framing and positioning the allegation contexts as future work.
    2. Most of Sonnet's 63 invalid outputs come from extra text after the JSON object or over-long explanations, concentrated on U. A secondary, clearly labeled lenient-parse analysis (take the first JSON object, ignore the length cap) would show whether Sonnet actually abstained, which is the interesting semantic question. Keep the strict result as primary.
    3. Qwen in non-thinking mode with a 256-token cap may be under-resourced for reasoning about possible completions. A thinking-mode or larger-budget arm would help separate model capability from configuration.
    4. Twelve families from one invoice template is small. Varying the domain would test whether the abstention failure generalises.
    5. The report is long and very dense. A short plain-language summary and the headline table near the top would help readers a lot.

    Overall a careful, well-controlled diagnostic.

    Read full reviewShow less

Cite this project

@misc{samson2026echoaudit,
  title = {{EchoAudit Assessing reports from incomplete evidence}},
  author = {Nathan Samson},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/echoaudit-assessing-reports-from-incomplete-evidence-jj4l}},
  url = {https://apartresearch.com/sprints/projects/echoaudit-assessing-reports-from-incomplete-evidence-jj4l}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026