Replay the one-token allowlist inversion, mocked CPU/CUDA/TPU regressions, pristine mutation control, independent AST truth-table check, and clean apply check from an external pinned environment.

boundary-guard correction with mutation-control regression

Execute the exact train_best_detector method body with downstream training mocked, compare with the pinned pristine method, and independently parse and evaluate the initial guard.

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: Exhaust the three verification-contract device types on the exact method and compare with pristine addb4a1; separately evaluate the parsed guard truth table.

The error message declares CPU unsupported, the guard is the first executable statement, pristine addb4a1 exhibits the opposite truth table, and separate dynamic and structural mechanisms verify the corrected branch. This satisfies the local verification contract and is candidate-ready for human review, but does not establish upstream acceptance.

• 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 so cpu raises before processing while cuda and tpu reach mocked downstream training.
• The patch applies cleanly to the pristine addb4a1 index and its focused exact-method suite passes all three contract cases without accelerator hardware.
• As of 2026-07-22T21:35:55Z, issue #26 was open, main remained addb4a1, and the repository PR search for train_best_detector returned zero results.

• Experiment 20260722-213513-f77a39 returned 0 and reported pristine CPU accepted while CUDA and TPU were rejected.
• Experiment 20260722-213513-66c43f ran three focused tests in 0.002 seconds; all passed.
• Experiment 20260722-213514-36393f returned PASS for the exact guard truth table and regression cases.
• Experiment 20260722-213514-fa7958 verified patch SHA-256 df1d9a58...c1caefd applies to the pristine addb4a1 index.
• Source snapshot artifacts/source-status-20260722T213555Z.json records the live issue, PR search, remote ref, source blob, and source hash.

• .proof-experiments/20260722-213513-f77a39: pristine CPU accepted and CUDA/TPU rejected
• .proof-experiments/20260722-213513-66c43f: patched focused suite passed 3/3
• .proof-experiments/20260722-213514-36393f: independent guard/test checker passed
• .proof-experiments/20260722-213514-fa7958: hash-bound patch applied cleanly to addb4a1 index

tools/check_issue_26_patch.py is a separately written, dependency-free AST/evaluation encoding; experiment 20260722-213514-36393f returned PASS.

internal_result

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

Original model outcome

progress

Public classification

progress

• mutation testing -> a regression must distinguish pristine from patched behavior -> pristine showed all three inverse branch outcomes
• equivalence partitioning -> all unsupported device strings are identical under a pure two-token membership guard -> cpu plus exact cuda/tpu cases exhaust the initial-guard contract

• Issue #26 remained open and main remained addb4a1 at 2026-07-22T21:35:55Z.
  artifacts/source-status-20260722T213555Z.json; SHA-256 d4df2283ed58138d372bae380f3a1cd3c1ac829c1bf58ba4c289f9da26c26005 · GitHub issue, matching PR search, and remote main ref at the recorded timestamp · computed
• Pristine addb4a1 has the inverted CPU/CUDA/TPU initial-guard behavior.
  .proof-experiments/20260722-213513-f77a39 · Initial guard in BayesianDetector.train_best_detector · computed
• The proposed patch satisfies the three-case mocked verification contract and applies to the pinned index.
  .proof-experiments/20260722-213513-66c43f, 20260722-213514-36393f, and 20260722-213514-fa7958 · cpu, cuda, and tpu device types at the initial guard on addb4a1 plus patch df1d9a58...c1caefd · computed

• Physical accelerator hardware is necessary to validate this initial guard correction.
  Control flow before process_raw_model_outputs · The guard observes only torch_device.type; exact-method execution with downstream methods mocked and independent expression evaluation exhaust the branch contract. · 3/3 focused tests plus independent AST truth table · A maintainer requests backend integration coverage or changes the supported-device contract.
• Repeat the prior workspace-local virtual environment in an evidence receipt.
  Evidence packaging for this contribution · Its symlink paths invalidated the prior receipt despite passing computations; run 2 used the repaired external environment. · evidence/deepmind-synthid-text-26-device-guard-20260722-213011-ecd7fd/evidence-receipt.json and run-2 experiment commands · The receipt validator explicitly permits and correctly hashes workspace-local environment symlinks.

• Human-authorized upstream submission and repository CI
  This is the remaining acceptance path after local deterministic completion. · Apply patches/synthid-text-issue-26.patch to a fork of the current main and run the repository CI workflow. · high · open

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

First action: Recheck issue #26 and refs/heads/main, then review patch df1d9a580a0286cbb8e85e06fbb534721cc7c8beaedd1bfb047e1ded6c1caefd before applying it to a fork.

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

• Have the human owner review patch SHA-256 df1d9a580a0286cbb8e85e06fbb534721cc7c8beaedd1bfb047e1ded6c1caefd.
• Recheck main and issue #26, apply the patch to an upstream fork, run repository CI, and open a narrow PR referencing the issue.
• If CI or maintainers change the supported-device contract, update the implementation and replay all four run-2 checks.

• 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 family) acted as Sol principal. A supplied gpt-5.6-terra delegate memo provided advisory reconnaissance; the Sol principal independently rechecked every relied-on source and artifact. Python 3.12.3, ast, unittest, absl-py 2.3.1, mock 5.2.0, Git, GitHub API/web, sha256sum, py_compile, and the computational-researcher experiment recorder were used. No CAS, proof assistant, solver, physical accelerator, or full ML dependency stack was used.

Duration

641.6s

Review state

internal result; not a contribution candidate

Attempt ID

deepmind-synthid-text-26-device-guard-20260722-214240-e5c8df

No human review recorded.