# Independent correction verification result

Actor `research-abi-correction-verifier`, role `verifier`, requested model
`openai/gpt-5.6-sol`, campaign session `run-5bd0e537bb0c4c4f43987ac1`.

## Scope and identities

This was an independent static partial comparison plus held-out stop-boundary ablation. It was not
original-game runtime comparison, live allocator validation, independent replacement, integrated
replay, or an acceptance verdict. The paired source bindings reproduced as engine
`ccd8e02083e8d2e2b3e97976ace2273c8f924dfc02a39e919004eaf3544c50fd` and RE
`6696fd5201e144843617cbf6d78b41b5287ad5dcc9fa1e8aaa861d52b64e72e8` at the contract HEADs.
The 7,898,624-byte input SHA-256 was
`970b7de729956a53094c7eb98aba4270aee98e2fed5daf0d39e290013c90c841`. The dereferenced
373,888-byte GNU objdump 2.38 executable SHA-256 was
`1eaaef2e7f57c4c7f69115c495e2466f5a8c8e5f3bc42221d092382f30f9d4cd`.

## Reproduction and observations

From `/home/alex/sots-re`:

```sh
python3 verify/results/research-completion-abi-correction-verifier/reproduce_run_ee78b8773688ca09f8046e21.py
```

The process exited 1 after eight positive objdump executions (four exact windows and four held-out
windows widened by four bytes). Every subprocess returned zero, emitted nonempty stdout, and emitted
empty stderr. All four exact stdout streams byte-matched the analyst captures. Binary/tool checks,
the independent empty list of pre-`0x00885413` writes to `[EBP-0x20]`, get/create ordering, append
copy ordering, nested empty/nonempty-copy branch markers, and destructor null/non-null branch markers
passed. The generated manifest reports 44 passing checks and one failure.

The failed check is `nested-copy-wide-row-prefix`. Exact stop `0x00779a20` ends with:

```text
779a1f: 74     je 0x779a31
```

The widened stop `0x00779a24` exposes the complete instruction:

```text
779a1f: 74 10  je 0x779a31
```

and then rows at `0x00779a21` and `0x00779a23`; the latter call is itself truncated by that widened
boundary. Thus exact rows are not a byte-identical prefix of widened rows and the analyst's
`nested-copy` window has a truncated terminal instruction. This falsifies the predeclared boundary
prediction. Surprise `s-4f4fb5e8ee42ccff7e68efb5` records and blocks the affected interpretation.

## Provenance defect and residuals

The reproduction helper was inherited from the prior verifier checkpoint and hard-codes prior
session `run-ee78b8773688ca09f8046e21` into its generated manifest. Although the raw commands ran in
this fresh quantum, that manifest does not truthfully identify the current campaign session and is
not acceptable as a current-session evidence record. This report records the actual invocation and
does not relabel the generated manifest.

No success is claimed for the failed full compare. Surviving rows remain partial static
measurements. Residuals include a complete non-truncated nested-copy boundary, fresh analyst and
verifier provenance, live spare/full-capacity fixtures, nonempty live nested-copy state, allocator
safety, induced unwind, original runtime comparison, whole-state/RNG validation, independent
replacement, and integrated replay. No RNG execution occurred; static no-RNG prose was not treated
as runtime state accounting.