Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026Dallas, TX

Self-Report Under Audit: Testing a Persistent Agent's Introspective Claims Against a Hash-Chained Ground-Truth Record

Jason Gibbs · Team Wave Squad

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

Read the report

Report: Self-Report Under Audit: Testing a Persistent Agent's Introspective Claims Against a Hash-Chained Ground-Truth Record

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

First real-world, tamper-evident audit of a deployed agent’s self-reports. Over 23 days Cadence made 62 explicit self-corrections (2.7/day). In a pre-registered test it predicted which of its own memories were load-bearing and scored only 6/16 — no better than chance and matched by a zero-introspection baseline. This shows agent self-reports can be measured in deployment without model access, and currently aren’t very reliable.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

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

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

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.

  1. This is a distinctive and very well-presented contribution that extends self-report research from controlled elicitation to longitudinal content accuracy in a deployed persistent agent. I particularly appreciated the hash-committed preregistration, mechanically defined ablation ground truth, calibration analysis, external-observer baseline, explicit correction of the paper's own failed claims, and unusually strong reproducibility discipline.

    The most important improvement would be to sharpen the base-rate interpretation. The reported 62 explicit corrections over 23 days is an observed correction frequency rather than a self-report error rate unless the denominator of mechanically auditable claims is known. Future versions should report errors per auditable claim/opportunity, ideally broken down by verifier and claim type, and distinguish detected error frequency from an estimate of total underlying error.

    The introspection result is interesting but currently small and system-specific. The 16 memory nodes should be expanded using independently defined or randomized sampling, and the "load-bearing" criterion should be tested across multiple retrieval/behavioral measures. I would also strengthen the privileged-access control: matching a graph-degree heuristic shows no advantage over that observable proxy, but a richer external predictor or another agent with access only to public artifacts would make the comparison more compelling.

    Replication across multiple persistent agents and model families would be especially valuable. If the same audit apparatus reveals stable classes of stale-belief, double-counting, proxy-substitution, and over-concession errors across systems, this could become a useful operational methodology for both AI safety and welfare research.

    Overall, I see this as a strong methodological paper with a genuinely valuable direction: deployed-agent self-reports should be treated as auditable claims rather than inherently privileged evidence.

    Read full reviewShow less

Cite this project

@misc{gibbs2026selfreport,
  title = {{Self-Report Under Audit: Testing a Persistent Agent's Introspective Claims Against a Hash-Chained Ground-Truth Record}},
  author = {Jason Gibbs},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/selfreport-under-audit-testing-a-persistent-agents-introspective-claims-against-a-hashchained-groundtruth-record-svac}},
  url = {https://apartresearch.com/sprints/projects/selfreport-under-audit-testing-a-persistent-agents-introspective-claims-against-a-hashchained-groundtruth-record-svac}
}

More from Digital Minds Research Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026