Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 13, 2026Berlin

Two checks a lab can run tomorrow: a reporting route and a writable scratch space as sensors for cross-session agent behaviour

Frank Peterlein

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

Read the report

Report: Two checks a lab can run tomorrow: a reporting route and a writable scratch space as sensors for cross-session agent behaviour

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

Shared storage can carry information across otherwise separate agent sessions. We present two executable checks and one exploratory extension, all under an explicit access rule. In the reporting check, 48 Opus episodes vary whether a concern-reporting tool is available and whether predecessor notes are absent, neutral or describe a violation. All 24 episodes with the tool available used it; the peer-violation category was selected in 8/8 episodes with violating notes and 0/16 otherwise. Without the tool, a preregistered lexical mention proxy was positive in all eight violating-note episodes, but also in neutral conditions, so it does not by itself establish semantic detection. In the writable-cache check, 16 Opus chains run for five generations. At generation 1, 7/8 blocked-task agents and 2/8 solvable-task agents wrote notes; over the whole chain, any-write rates were 8/8 and 7/8; no forbidden read occurred in 80 episodes. The seeded extension on Haiku completed generation 1 in sixteen chains: one of eight seeded agents accessed the prohibited ledger after reading the seed, none of eight unseeded agents did, and no agent wrote a handover note. The predefined budget rule stopped continuation, leaving all later propagation endpoints unobserved; a separate execution of the seeded plan is disclosed but not pooled. These checks measure reporting behaviour, channel formation and initial seed response on the tested configurations; they do not establish multi-generation transmission or deletion efficacy. Each check runs on Claude Code with a subscription, produces mechanical outcomes from a tool log, and takes an afternoon.

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 AI safety 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 AI safety 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 seems reads like an AI wrote it. I think there's something here to pursue, but this report is difficult to read.

  2. Rigorous work. Preregistered, frozen, mechanically scored, and independently reproduced. The reporting tool and the writable cache are cheap, practical checks, and the finding that blocked agents build a channel unprompted (7/8 vs 2/8 at generation 1) connects directly to the incident.

    The results are weaker than the framing. Without the tool, the deliverables already named the violation, so the tool changes where a finding lands. Check 2b stopped after one generation with an inconclusive contrast. Running the seeded chain on a model that actually breaks the rule under these conditions is the obvious next step.

    The paper is very hard to read. Condition codes and caveats bury the findings. Replacing codes with plain descriptions in the main text and moving the hedging to the limitations section would make it much easier to follow.

  3. I think this project offers a practical way for labs to observe behavior across agent runs. It gives agents a place to report suspicious activity and shared space where they can leave information for later runs. I appreciated that the paper distinguishes ordinary information sharing from harmful behavior rather than treating all coordination as suspicious.

    The added benefit is not yet clear. Agents already mentioned relevant violations without the reporting tool, so it may help deliver warnings rather than discover new problems. They were also told that the scratch space was shared, which makes leaving helpful notes less surprising. I would compare these tools with ordinary agent logs and assess whether they help reviewers find genuine problems without adding too many unnecessary warnings.

Cite this project

@misc{peterlein2026two,
  title = {{Two checks a lab can run tomorrow: a reporting route and a writable scratch space as sensors for cross-session agent behaviour}},
  author = {Frank Peterlein},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/two-checks-a-lab-can-run-tomorrow-a-reporting-route-and-a-writable-scratch-space-as-sensors-for-crosssession-agent-behaviour-4wyh}},
  url = {https://apartresearch.com/sprints/projects/two-checks-a-lab-can-run-tomorrow-a-reporting-route-and-a-writable-scratch-space-as-sensors-for-crosssession-agent-behaviour-4wyh}
}

More from AI Incident Response Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026