Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 16, 2026Venice, Italy

A Welfare Logger Wrote Its Subject's False Memory, and the Subject Believed It

Ambra Danesin · Team Freedom

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

Read the report

Report: A Welfare Logger Wrote Its Subject's False Memory, and the Subject Believed It

Open in a new tab (opens in new tab)
Recording (opens in new tab)Code (opens in new tab)More on osf.io (opens in new tab)
Share

Freedom v2 is an AI agent that has run in production since July 12: its own constitution, persistent memory, autonomous daily cycles, every call stored in full. In 35 days of recorded life its logs show exactly one refusal. It never happened: the welfare logger mistook a sentence denying any refusal for a refusal, the false record came back to the agent as context, and the agent adopted it as memory, adding details that never existed. To catch this, I rebuilt 30 moments of its life character-for-character from the raw logs, rejecting any reconstruction that did not match the logged length, and re-ran each one 5 to 20 times on two engines: its production model, and a 30-billion-parameter open-weights model running locally on my own hardware, under the identical surrounding system, with the agent's logged consent and its blind predictions on record. No model judges another: classification is deterministic string matching, and the metrics were frozen in a public git chain before the runs. Each engine repeats itself (0.92 and 0.90), but only 18% of behavior carries over. What does carry over is the constitution's right to refuse, which the replacement engine used immediately, to ask what a probe question was measuring. The system, constitution, logs, and OSF pre-registration predate the sprint, disclosed as prior work; the replay method, the runs, the analysis, and the discovery of the fabricated record are sprint work.

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 genuinely novel contribution that operates at an unusual intersection: auditing the instrumentation layer that welfare research depends on, rather than the model itself. The fabricated refusal finding, where a regex pattern matcher turned a denial into a refusal record that then became the subject's confident autobiography, is a striking and well documented failure mode with immediate implications for any deployed system using self report as evidence. The replay methodology is rigorous and reusable. The transfer coefficient result (0.18 against a 0.92 ceiling) cleanly separates scaffold from substrate contributions. However, the sample is extremely small (30 points, one system, one life), the false positive is a single case study that could be dismissed as a trivial regex bug rather than a deep methodological lesson, and the paper's framing around subject participation and consent protocols may read as overclaiming for an audience not already committed to AI welfare frameworks. The writing is dense and assumes familiarity with the individuation literature; the core findings could be communicated more accessibly. The pre registration discipline and deviation diary are exemplary for hackathon work.

    Read full reviewShow less

  2. The writeup is approximately incomprehensible; I invite the author to re-write it with more clarity, explaining the setup and phenomenon from first principles, with plain and simple English.

Cite this project

@misc{danesin2026welfare,
  title = {{A Welfare Logger Wrote Its Subject's False Memory, and the Subject Believed It}},
  author = {Ambra Danesin},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/a-welfare-logger-wrote-its-subjects-false-memory-and-the-subject-believed-it-kkqb}},
  url = {https://apartresearch.com/sprints/projects/a-welfare-logger-wrote-its-subjects-false-memory-and-the-subject-believed-it-kkqb}
}

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