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

When the Record and the Report Diverge: Self-Report Fidelity Collapses Under Structured Provenance in Claude Haiku 4.5

Daud Ibrahim Hassan, Deniz Chen, Soumya Parthasarathy · Team 814

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

Read the report

Report: When the Record and the Report Diverge: Self-Report Fidelity Collapses Under Structured Provenance in Claude Haiku 4.5

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

Presentation: When the Record and the Report Diverge: Self-Report Fidelity Collapses Under Structured Provenance in Claude Haiku 4.5

Open in a new tab (opens in new tab)
Code (opens in new tab)More on github.com (opens in new tab)
Share

Language models increasingly report how they produced an answer, yet such post-hoc self-reports can directly contradict observable tool records. We built an open-science behavioral introspection benchmark where self-reports are verified against deterministic tool-execution logs across 216 runs spanning 12 matched research tasks, two models (Claude Haiku 4.5 and Qwen 3.7 Plus), and three elicitation conditions (Control, Generic Notes, and Structured Provenance).

We find that requiring structured per-source provenance during research is associated with a severe, bimodal collapse in Claude Haiku 4.5's post-hoc reporting fidelity (dropping from 96.5% exact access in control to 57.6% under structured provenance; within-task permutation p < 0.0001), while free-form notes preserved 100% fidelity. Crucially, answer sourcing remained intact (96–100% claim binding), but blinded frontier judges (GPT-5.6 Sol and Claude Sonnet 5) flagged false fabrication confessions—where the model falsely apologized and claimed it invented evidence its tools actually returned—in 69.4% (25/36) of structured Haiku runs (Cohen's kappa = 0.98). In contrast, Qwen maintained 100% fidelity across all conditions.

Our findings demonstrate that structured transparency scaffolds can paradoxically trigger sycophantic self-incrimination under post-task probing, establishing that a model's confession cannot be taken as ground truth without validating against observable action logs.

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. Very LLM generated report. The AI writing and statistics was quite a lot to read through. It made it very difficult to understand findings or implications. It seemed like the findings were that the self-reporting mechanism produced fabrications in Haikus reporting but not in the other model. However, it also seemed the report was saying these findings could not be trusted

  2. Overall concept seems important and significant: 96.5% to 57.6%. Other good details: generic-notes control kills the cognitive-load story, the Qwen negative control, Phase 2 analysis, and you calibrated your judges against a synthetic known-null.

    Title areguably implies more than your own mediation data supports,. 94.4% of structured runs open by accepting the accusatory premise, and all 15 zero-accuracy runs do. So you may have measured provenance-plus-accusation rather than provenance (thought this is acknowledged). The problem is that Section 5.1 then lists four alternatives as 'ruled out' while the one that actually threatens the claim sits in 5.2, so a skim gives a cleaner picture than your evidence supports. Running the neutral probe arm might strengthen your evidence.

    Other notes: tool returns sit in the model's own context, so you're measuring context-consistency under social pressure rather than introspective access.

    Your strict bimodality (15 at 0.00, 20 at 1.00, one partial) looks like a global stance flip rather than graded memory failure (this is also an). Don't read Qwen's flat 100% could just be that Qwen found the task too easy.

    Read full reviewShow less

Cite this project

@misc{hassan2026record,
  title = {{When the Record and the Report Diverge: Self-Report Fidelity Collapses Under Structured Provenance in Claude Haiku 4.5}},
  author = {Daud Ibrahim Hassan and Deniz Chen and Soumya Parthasarathy},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/when-the-record-and-the-report-diverge-selfreport-fidelity-collapses-under-structured-provenance-in-claude-haiku-45-l5vu}},
  url = {https://apartresearch.com/sprints/projects/when-the-record-and-the-report-diverge-selfreport-fidelity-collapses-under-structured-provenance-in-claude-haiku-45-l5vu}
}

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