sots-re/verify/results/research-completion-abi-independent/result.md

Independent ABI verification result (paused)

Session run-6c7a726f5a2849ef0407876f; verifier research-abi-independent; model claim openai/gpt-5.6-sol.

Classification and verdict

Fail / blocked for independent-cross-check pending Astra resolution of surprise s-5c58b6a921997731f5b5ae1f. This is an independent static reproduction plus archived-save cross-check only. It is not original-assisted runtime execution, a partial or full live compare, an independent replacement, or integrated replay. No live allocator-safety or replacement claim was tested.

Identities

• dumps/sots.exe: SHA-256 970b7de729956a53094c7eb98aba4270aee98e2fed5daf0d39e290013c90c841, MD5 9969481c39f4b33a8a21c48b62abee4c (independently recomputed).
• Tool reports GNU objdump (GNU Binutils for Ubuntu) 2.38.
• Engine worktree: HEAD 7741d42fc5e4e761e6449bdaf0e4a61d00036a23, common Git directory /home/alex/sots-engine/.git, source-binding ccd8e02083e8d2e2b3e97976ace2273c8f924dfc02a39e919004eaf3544c50fd.
• RE worktree: HEAD 3bfde5a70d874a723e797a695bbd847fd82c0aa7, common Git directory /home/alex/sots-re/.git, source-binding 6696fd5201e144843617cbf6d78b41b5287ad5dcc9fa1e8aaa861d52b64e72e8.
• Both worktrees had the pre-existing dirty inventories recorded in the checkpoint. This verifier changed neither worktree and touched no lab resource.

Reproduction results

All four handoff commands were executed directly, with no output filter. Each exited zero and emitted nonzero instruction output. Fresh stdout SHA-256 and instruction counts were:

window	instructions	stdout SHA-256	archive comparison
0x00825d40..0x00825e65	120	1c2408cd49cc10383bad9fe06d3287476b103205f4155adf64c8a50ccd5205e8	mismatch
0x0046f8c0..0x0046f8f0	26	373b04cbc836e81ee84145796c86766fcbeb363ac3c5be44dc8419ee91a170ed	exact after line-end whitespace normalization
0x004236a0..0x0042370d	51	1debe85f054a442a09cb84c895e1950487abe996051dde0820e5d687022d9a66	exact after line-end whitespace normalization
0x00422720..0x00422796	51	4e2f8375cc0c6f645a09d2ebedb95e40d0414fb8e667b8c01768357397e8c580	exact after line-end whitespace normalization

The sole semantic-text mismatch is at the first window's terminal instruction. The exact declared command stops at 0x00825e65, one byte into the three-byte instruction beginning at 0x00825e64. Fresh objdump stdout reports only byte c2 while still decoding ret 0x8; the archive reports bytes c2 08 00. Therefore the archive cannot be unfiltered stdout from the declared command under the recorded effective instrument. The mismatch is not dismissed merely because widening the boundary would likely recover the missing bytes.

This failed the predeclared provenance/instrument falsifier. The dedup ABI interpretation is paused; the three matching windows and zero exit statuses are measurements, not a pass.

Independent archived-state check (completed before pause)

python3 verify/save-reader/save_reader.py verify/results/saves/turn3-state.sav --dump --json --strict exited zero. The parser reported 37,847 items, 2,453 frames, zero resyncs, zero hint failures and zero best-effort recoveries (informational positional-tag issues remain). Direct tree inspection, rather than a reported event count, recovered player 1's event storage at inflated offset 120372: EvNxID=4; turn-2 bucket with event id 1; turn-3 bucket with event ids 2 and 3. Event 3 is Research Over Budget, message Research for Waldo Units has gone overbudget., image EVENT_RESEARCH_OVERBUDGET, location 0, action 1, chain id 0, and all three position words are the integer bit pattern 2139095039 (0x7f7fffff). Players 2 and 3 each have EvNxID=0 and an empty event collection. This supports those archived record values only; it is not live execution.

Challenge status, failed predictions, and residuals

The planned held-out NaN-position negative control was not promoted to a result after the instrument provenance surprise. Its prediction remains untested in this quantum. Live empty/short/long description fixtures, allocation failure, allocator compatibility, runtime event construction, full-capacity pre/post memory, RNG state and integrated replay also remain residuals. No RNG counter was treated as state evidence; no runtime workload executed, so RNG behavior is outside this result.

Reproduction commands are the four exact commands printed in objdump-2026-09-09-dedup-helper.txt, followed by the strict save-reader command above. The exact next action belongs to Astra and is recorded in the canonical checkpoint.