Revalidated the hash-bound allowlist inversion, three focused branch regressions, pristine inverse control, separate AST truth-table checker, clean pinned-index application, and live prior-art status.

boundary-guard correction with mutation-control regression

Execute the exact train_best_detector method body with downstream training mocked, compare against pristine addb4a1, 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: Run cpu, cuda, and tpu through the exact method body, replay the opposite outcomes on pristine addb4a1, and separately evaluate the parsed initial guard.

The error message explicitly declares CPU unsupported, the guard is the first executable statement, pristine behavior is exactly inverse, and independent dynamic and structural checks agree on all guard-equivalence classes.

• Pristine addb4a158143c7c6851a1308f78b89fceed59683 accepts cpu and rejects cuda and tpu at the initial train_best_detector guard.
• Patch SHA-256 df1d9a580a0286cbb8e85e06fbb534721cc7c8beaedd1bfb047e1ded6c1caefd makes cpu raise 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 3/3 without accelerator hardware.
• At 2026-07-22T21:48:46Z issue #26 was open, main was addb4a1, and three repository PR searches returned zero results.

• Experiment 20260722-214820-16d426: pristine inverse behavior PASS.
• Experiment 20260722-214820-984369: cuda, tpu, and cpu focused tests all PASS.
• Experiment 20260722-214820-54a4cb: independent exact guard and regression checker PASS.
• Experiment 20260722-214820-01af65: patch hash and pinned-index application PASS.
• artifacts/source-status-20260722T214846Z.json: live status and three novelty searches.

• .proof-experiments/20260722-214820-16d426: pristine cpu accepted and cuda/tpu rejected
• .proof-experiments/20260722-214820-984369: patched focused suite passed 3/3
• .proof-experiments/20260722-214820-54a4cb: separate AST/evaluation checker passed
• .proof-experiments/20260722-214820-01af65: exact patch applied cleanly to addb4a1 index

tools/check_issue_26_patch.py is a separately written dependency-free AST/evaluation encoding; experiment 20260722-214820-54a4cb returned PASS.

candidate_eligible

No structured gate reasons were recorded in this legacy attempt; see the adjudication ledger.

Original model outcome

candidate

Public classification

candidate

• mutation testing -> the regression must distinguish pristine and patched semantics -> all three pristine outcomes were inverse
• equivalence partitioning -> a pure two-token membership guard has two singleton supported classes and one unsupported complement -> cpu, cuda, and tpu exhaust the guard contract

• Issue #26 remained open and main remained addb4a1 at 2026-07-22T21:48:46Z.
  artifacts/source-status-20260722T214846Z.json; SHA-256 512e3cfea0f9a200bb487d44cb88c75a63ac2c49934946e1dc88a367411755bf · GitHub issue, remote main ref, source blob, and three repository PR searches · computed
• Pristine addb4a1 has the inverted cpu/cuda/tpu initial-guard behavior.
  .proof-experiments/20260722-214820-16d426 · 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/20260722-214820-984369, 20260722-214820-54a4cb, and 20260722-214820-01af65 · cpu, cuda, and tpu at the initial guard on addb4a1 plus patch df1d9a58...c1caefd · computed

• Physical accelerator hardware is required to verify this initial guard correction.
  Control flow before process_raw_model_outputs · The branch observes only torch_device.type; exact-method execution with downstream mocks and independent predicate evaluation exhausts its profiles. · 3/3 focused tests and experiment 20260722-214820-54a4cb · A maintainer requests backend integration coverage or changes the supported-device contract.
• Another local isolated branch replay will materially reduce the remaining uncertainty.
  The addb4a1 initial-guard contract · Runs 2 and 3 independently agree across dynamic, mutation, structural, and application checks; the remaining uncertainty is native CI and review. · Run-2 and run-3 experiment receipts · Upstream main changes, CI fails, or maintainer feedback alters the contract.

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

Objective: Obtain isolated human review, then native 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 or redirect when upstream CI passes/fails or a maintainer merges, rejects, or supersedes the exact patch.

• Have the human owner review patch SHA-256 df1d9a580a0286cbb8e85e06fbb534721cc7c8beaedd1bfb047e1ded6c1caefd.
• Apply it to a fork of the then-current main and run the native repository CI workflow.
• Open a narrow PR referencing issue #26 only with human authorization; revise only if CI or maintainer feedback changes the contract.

• 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 ↗
• https://github.com/google-deepmind/synthid-text/blob/addb4a158143c7c6851a1308f78b89fceed59683/.github/workflows/ci.yaml ↗

GPT-5.6 Sol acted as principal investigator. A prior GPT-5.6 Terra source-discriminator memo was advisory and every relied-on claim was independently rechecked. Deterministic work used Python 3.12.3, stdlib ast/unittest, absl-py 2.3.1, mock 5.2.0, git, GitHub REST API, and the Proof Factory experiment runner. No CAS, solver, proof assistant, accelerator, or full ML training stack was used.; orchestration: gpt-5.6-sol principal with gpt-5.6-terra delegates.

Duration

660.2s

Review state

needs isolated skeptic review

Attempt ID

deepmind-synthid-text-26-device-guard-20260722-215511-5c5b86

• accept · Charlie Krug · 2026-07-22 22:06 UTC
  I reviewed the evidence packet and approve this candidate for the next external contribution step.