Apply the one-token allowlist inversion, add mocked CPU/CUDA/TPU branch regressions, reproduce the pristine mutation, and independently check the patched AST truth table.

boundary-guard correction with mutation-control regression

Execute the exact train_best_detector method body with downstream training mocked, then independently parse and evaluate its initial device comparison.

Hypothesis: Changing the guard to `torch_device.type not in ("cuda", "tpu")` makes CPU raise before downstream work while CUDA and TPU reach mocked training.

Test: Run three focused branch cases on the exact method body and compare them with pristine addb4a1; separately evaluate the parsed guard for cpu, cuda, and tpu.

The error says CPU training is unsupported, the guard is the first executable statement, the pristine mutation exhibits the opposite truth table, and two independent mechanisms verify the patched branch. This meets the local verification contract but still needs upstream CI and human review.

• At upstream commit addb4a158143c7c6851a1308f78b89fceed59683, pristine train_best_detector accepts cpu and rejects cuda and tpu at its initial guard.
• Patch SHA-256 df1d9a580a0286cbb8e85e06fbb534721cc7c8beaedd1bfb047e1ded6c1caefd changes the guard to reject cpu and admit cuda and tpu to mocked downstream training.
• The patch applies cleanly to the pristine addb4a1 index and its focused suite passes three cases without accelerator hardware.

• `.venv/bin/python tools/check_prepatch_behavior.py upstream/synthid-text --ref addb4a158143c7c6851a1308f78b89fceed59683` returned 0 and reported the inverted pristine behavior.
• `.venv/bin/python tools/run_focused_device_guard_tests.py ...` ran 3 tests, all OK.
• `.venv/bin/python tools/check_issue_26_patch.py ...` returned PASS for the exact guard and cases.
• `python3 tools/check_patch_applies.py ... --base addb4a1...` returned PASS and the patch SHA-256.

• .proof-experiments/20260722-212217-0e29c2: pristine CPU accepted and CUDA/TPU rejected
• .proof-experiments/20260722-212217-0a0343: patched focused suite passed 3/3
• .proof-experiments/20260722-212217-0896d7: independent guard/test checker passed
• .proof-experiments/20260722-212330-d34490: hash-bound patch applied cleanly to addb4a1 index

tools/check_issue_26_patch.py is a separate dependency-free AST/evaluation checker; experiment 20260722-212217-0896d7 returned PASS.

internal_result

• Fewer than two reproducible, URL-backed novelty searches were recorded.

Original model outcome

no_progress

Public classification

no_progress

• mutation testing -> require the proposed regression suite to fail on pristine addb4a1 -> all three cases failed with the inverse branch behavior

• Issue #26 remained open and main remained addb4a1 at 2026-07-22T21:22Z.
  artifacts/source-status-20260722T212217Z.json · GitHub issue and refs at the recorded timestamp · computed
• Pristine addb4a1 has the inverted CPU/CUDA/TPU behavior.
  .proof-experiments/20260722-212217-0e29c2 · initial guard in BayesianDetector.train_best_detector · computed
• The proposed patch satisfies the three-case mocked verification contract.
  .proof-experiments/20260722-212217-0a0343 and 20260722-212217-0896d7 · cpu, cuda, and tpu device types at the initial guard · computed

• Physical accelerator hardware is necessary to test this guard correction.
  Control-flow behavior before process_raw_model_outputs · The guard reads only torch_device.type; exact-method execution with downstream methods mocked decisively covers the branch. · 3/3 focused tests plus independent AST truth table · A maintainer requests backend integration coverage or changes the supported-device contract.

• Upstream submission and full repository CI
  This is the remaining acceptance path after local deterministic completion. · Apply patches/synthid-text-issue-26.patch to a fresh upstream fork and run the repository test command in its declared dependency environment. · high · open

Objective: Obtain isolated human review, then upstream CI and maintainer disposition.

First action: Review and apply patches/synthid-text-issue-26.patch (SHA-256 df1d9a580a0286cbb8e85e06fbb534721cc7c8beaedd1bfb047e1ded6c1caefd) to a fork of addb4a1.

Stop condition: Stop local implementation work when upstream CI passes and a maintainer merges or rejects/supersedes the exact patch.

• Have the human owner review patch SHA-256 df1d9a580a0286cbb8e85e06fbb534721cc7c8beaedd1bfb047e1ded6c1caefd.
• Apply it to an upstream fork, run the repository's full dependency CI, and open a narrowly scoped PR referencing issue #26.
• If upstream main changes first, rebase and rerun all four deterministic checks before submission.

• https://github.com/google-deepmind/synthid-text/issues/26 ↗
• https://github.com/google-deepmind/synthid-text/commit/addb4a158143c7c6851a1308f78b89fceed59683 ↗
• https://github.com/google-deepmind/synthid-text/blob/addb4a158143c7c6851a1308f78b89fceed59683/src/synthid_text/detector_bayesian.py#L1031-L1036 ↗

OpenAI Codex GPT-5-series acted as Sol principal. A gpt-5.6-terra delegate supplied advisory reconnaissance; the Sol principal independently rechecked all relied-on claims and artifacts. Python 3.12.3, ast, unittest, absl-py 2.3.1, mock 5.2.0, Git, curl/GitHub API, and the computational-researcher experiment recorder were used. No CAS, proof assistant, solver, physical accelerator, or full ML training stack was used.

Duration

1058.1s

Review state

evidence invalid or incomplete; not durable progress

Attempt ID

deepmind-synthid-text-26-device-guard-20260722-213011-ecd7fd

No human review recorded.