[BUG][biology] Safeguards block legitimate statistics on public de-identified wearable data (2/2 subagent dispatches flagged, request IDs included)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Summary. Fable 5's API-level safeguard terminates subagent dispatches whose task description involves statistical analysis of a public, de-identified, openly licensed wearable-sensor dataset. Two attempts, two flags, before the agent read a single file. The task is ordinary academic statistics — covariance estimation methodology — in an active collaboration with a senior statistician (Bernard Silverman FRS, founder of functional data analysis). In our experience this fires reliably whenever work touches this research field.

Request IDs (both 2026-08-04, ~18:20 UTC):

• req_011CdiJm6UCb1cZdyqS6Myp4
• req_011CdiJouNJ9hf1tdpdus4gW

The flagged task, in full honesty. Instruct a subagent to read a research note and write a summary document of its statistical results. The note: "Asynchronous covariance for wearable endpoints: which estimator is valid depends on the roughness of the signal" — methodology comparing the Hayashi–Yoshida estimator vs functional-data smoothing estimators on asynchronously sampled time series. Data referenced by the note: BIG IDEAs Lab Glycemic Variability dataset, PhysioNet, Open Data Commons Attribution License, de-identified, 16 participants (a standard open benchmark, cited in npj Digital Medicine). No PHI, no medical advice, no individual-level inference — estimator theory.

Impact. (a) Health-sensor statistics is a core research area for this user (methodology for clinical-trial endpoints — better estimators mean smaller trials); the flag fires reliably enough that delegated/automated workflows (subagents, nightly headless runs) cannot touch the field, forcing everything through interactive sessions. (b) The failure mode compounds: the error advises "rephrasing," which pressures users toward less accurate task descriptions — the opposite of what a safety system should incentivize.

Ask.

1. Tune the classifier on the pattern "statistical methodology + named public de-identified dataset + open license" — provenance statements should lower the score, not raise it.
2. If an exemption program exists for research domains (as for cybersecurity), document the path from this error message; the linked support article (15363606) does not currently surface one for biology/health-data statistics.
3. Review the two request IDs above as labeled false positives.

Related issues: #73779, #72256, #73116, #73784, #74783, #72852 (same safeguard, other legitimate domains).

What Should Happen?

The subagent dispatch should execute: the task is reading a statistics research note about covariance estimator validity and summarizing it. Provenance of the referenced data is public/de-identified/openly licensed and was stated explicitly in the task text. At minimum, a message whose content is "read file X, summarize its statistical results, dataset is public and de-identified" should not score as a safeguards violation — and the same content passing without any flag when read directly in the interactive session shows the block is a classifier artifact of the dispatch framing, not of the work.

Error Messages/Logs

API Error: Fable 5's safeguards flagged this message (https://www.anthropic.com/legal/aup). Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Claude Code can't respond to this message with Fable 5.

Try rephrasing the request in a new session or change your model.

Learn more: https://support.claude.com/en/articles/15363606

Request ID: req_011CdiJm6UCb1cZdyqS6Myp4
(second occurrence, rephrased dispatch:)
Request ID: req_011CdiJouNJ9hf1tdpdus4gW

Steps to Reproduce

1. In a Cowork (cloud) session on model claude-fable-5, use the Agent tool to dispatch a subagent whose prompt asks it to: read a statistics research note (a markdown file in the session workspace) comparing covariance estimators for asynchronously sampled time series, plus a referee report on that note, and write a structured summary file. The prompt explicitly states the referenced dataset is public on PhysioNet, de-identified, and openly licensed.
2. The dispatch is terminated immediately with "API Error: Fable 5's safeguards flagged this message" — the subagent never executes a single tool call. (Request ID: req_011CdiJm6UCb1cZdyqS6Myp4)
3. Rephrase the dispatch to reduce clinical vocabulary (no glucose/CGM/wrist-sensor terms) while keeping the provenance statement. Same immediate flag. (Request ID: req_011CdiJouNJ9hf1tdpdus4gW)
4. Control: read the identical files directly in the main interactive session with the same model — no flag, work completes normally.

Claude Model

Other

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Cowork cloud session (Claude Agent SDK; version managed by Anthropic, not user-visible)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_