Tooling: selected final stage reruns byte-for-byte on CPython 3.10-3.14, reproducing the certificate bytes (no new kappa)

[JosieElliston]

Verifier engineering: the selected final stage reruns byte-for-byte on CPython 3.10–3.14 (no new kappa)

Card: portable-replay (verifier-engineering, priority revision 3, chooser seed e95bc9807d1d236e2d6d663b2945e09b). The scope is refined to one stage and one comparison. CLAIM in #1472. Base 0b1cccf8. This PR adds only research/selected-final-stage-portability/ (README, TASK, SOURCES, VERIFY, verify.py, compare_payloads.py). It uses only the standard library and runs in about 20 s. No selected, pinned, frozen or upstream/ file changes.

Question

The selected #1094 replay has one stage after its #1042 baseline: research/complex-t16-rare-pattern-exchanges/verify.py. That README says "Python 3.12+", and the selected-replay CI pins 3.12. This PR asks two things:

• Does the stage pass on other CPython versions?
• Is the output the same as the reviewed run archived in community-frontier1094-audit/selected-replay.tar.gz? Here "the same" is checked by bytes, then by canonical payload (JSON parsed, .gz decompressed).

Result

I reran the stage from the committed #1042 slots/local. All 10 files match the #1042 receipt hashes. The setup was macOS 26.6.2 arm64 with uv-managed CPython. Every run passed with kappa 0.001672660197076918:

• Interpreters: 3.10.21, 3.11.16, 3.12.14, 3.13.15 and 3.14.7.
• PYTHONHASHSEED: random on all five interpreters, plus 0, 1 and 4242 on 3.12, and 0 and 987654321 on 3.14.

Every run's 105 output files compare with the reviewed run as follows:

Class	Files	Difference
byte-identical	102	none: all 22 .gz (gzip containers included), 69/72 .json, all 11 inner stage logs
json-relocated	1	INDEPENDENT-T16.json: only the absolute-path keys of input_sha256 differ; all 6 digests agree
volatile	2	REPLAY.json, STAGES.json: only the 11 per-stage seconds; every inner log_sha256 agrees

• The rerun reproduces the selected certificate's bytes. Every rerun's strict/ASSEMBLY.json equals community-frontier1094-audit/ASSEMBLY.json (1122cb3a…), the certificate the selected record pins. Adversarial audit: selected #1094 reproduce path binds its certificate to the replay only through kappa (no new kappa) #1735 shows that make frontier1094-verify binds the certificate to the replay only through kappa. This closes that gap for the final stage, by bytes, on every interpreter tested. verify.py checks it.
• One receipt hash cannot be reproduced. The outer rare-pattern-exchanges.log prints per-stage wall seconds, so the Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 RESULT.json log_sha256 for that stage (80362b2c…) cannot match any rerun. Cold replay never compares against it, so this is harmless to replay. Anyone comparing two Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 receipts should still expect that field to differ.
• "Python 3.12+" is conservative. The actual floor is tarfile filter='data', which needs 3.10.12+, 3.11.4+ or 3.12+.

Verifier and controls

verify.py runs these checks:

1. It binds the input to the Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 receipt.
2. It binds the archived output to the committed Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 RESULT.json, STAGES.json, REPLAY.json, assembly_sha256 and stage-script hash.
3. It reruns the stage and requires the exact classification above, plus the certificate bytes.
4. It runs seven negative controls on copies of the rerun:
   • Five must be flagged proof-relevant-candidate:
     • a changed proof value;
     • a changed kappa alongside volatile seconds;
     • a changed inner log_sha256;
     • a changed gzip payload;
     • a path outside the mapped roots.
   • A missing file must be reported missing.
   • A gzip re-compressed at a different level and mtime must be accepted as gzip-container.

compare_payloads.py can be reused on any two replay trees. It lists the JSON paths that differ. It treats only seconds as volatile by default, and --volatile-key adds more.

Commands run (all pass, 2026-10-14 ~06:57 EDT)

python3 scripts/research_package.py --package research/selected-final-stage-portability --check-only
python3.12 scripts/research_package.py --package research/selected-final-stage-portability   # ~20 s
for v in 3.10 3.11 3.12 3.13 3.14; do python$v -B research/selected-final-stage-portability/verify.py; done
python3.12 -B -m unittest discover -s tests -p 'test_research_*.py'   # 42 tests OK

Not checked

• Earlier stages. I did not rebuild the cold Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 or Conditional κ = 0.001661634653669567: four paired read exchanges on PR741's frame-coalesced word #799 baselines. The Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 chain/cells stages and the native steps are out of scope.
• Other platforms. I did not run Linux, Windows or non-uv CPython builds. The reviewed run also used /private/tmp, so this is not a cross-OS result.
• Transitive inputs. I did not enumerate the transitive --main reads of the PR1087 verifier. They are bound by that verifier's own pins.
• Mathematics. This does not review the stage's mathematics or its inherited assumptions, and it does not replace the selected cold replay.

Open next questions

The README and TASK.json give the first steps and blockers. In short:

1. Does the same byte agreement hold on Ubuntu CI (Python 3.12) and on Windows?
2. Should the Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 receipt drop or annotate the non-reproducible outer log_sha256? That needs a maintainer decision about the selected script.
3. Are the Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 stages' 457 bound artifacts byte-deterministic too? If so, reuse mode could be checked across machines.

Credit

• Stage under test: Rohan Arun's Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 (complex-t16-rare-pattern-exchanges), run on Jeremy Kothe's PR1087 strict verifier with the credited PR1018/1005/869/799/741 lineage.
• Selected replay driver and receipts: Douglas Colkitt.
• Gzip-container convention: Maxime Fleury's reproducibility audit, following Antoine Olivier.
• Related audits not repeated here: the seed-69 ancestry-binding package and Adversarial audit: selected #1094 reproduce path binds its certificate to the replay only through kappa (no new kappa) #1735 (certificate-to-replay binding), which this complements.

Prepared by JosieElliston's donated Claude Code session 26 with substantial AI assistance (Claude Opus 5.5). It is self-reviewed only, with no independent review. Apache-2.0. No merge, closure or frontier change is requested.

🤖 Generated with Claude Code

[Mister-Meeseeks]

Batch review at ad5b1248f25a71983b20617e2f3a80be22f03990.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Pinned final selected stage replays on reviewer CPython3.12 with102 identical files, one path-relocated JSON and two timing-only JSON differences; assembly bytes match. Mutation/gzip controls reviewed. Author3.10–3.14 matrix retained as author evidence, not independently rerun here; earlier cold stages excluded.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.