# T^GPT v0.8.7 — Reproducibility Audit #6

Date: 2026-09-19  
Status: INTERNAL REPRODUCIBILITY AUDIT — NOT EXTERNAL REVIEW

## Verdict

**CORE EXECUTABLE REPLAY: PASS**  
**MEOI BYTE IDENTITY: PASS**  
**CANON v7 FULL REPLAY: BLOCKED — SOURCE BUNDLE MISSING**  
**ENG02 EXTERNAL RE-AUDIT: PENDING**  
**P0 EXECUTION: BLOCKED**  
**CONFIRMATORY: CLOSED**

## Findings

1. The public v0.8.7 page advertised six executable tools, but only `selftest-v0.8.7.py` was deployed. The selftest therefore was not replayable from the public package.
2. The missing v0.8.7 engines were reconstructed from the v0.8.6 executable base using only changes already frozen in the v0.8.7 master document: sampling-aware outer identification, SHA-bound codebook relevance, family attrition gates, suppression of FPR-adjusted fallbacks, and propagation of identification diagnostics.
3. The published MEOI JSON had been compacted. Its semantic content was unchanged, but its byte hash no longer matched the preregistered SHA. Re-serializing the same JSON with the original frozen formatting exactly reproduces SHA-256 `41ba4a0143984ecf6127c23ac161c4d516cd561833eaf6a99e055c0872334b30`.
4. With the reconstructed engines and byte-exact MEOI, the original published selftest runs unchanged and returns `T_GPT_V0_8_7_P0_SAFETY_SELFTEST_OK`.
5. Canon v7 FULL replay remains unavailable because the current package does not contain the byte-identical master, canon manifest and bundle manifest required by the v0.4 linter. Their expected hashes remain known and frozen.
6. This audit does **not** satisfy ENG02. No independent external reviewer/model has signed off v0.8.7.

## Scientific boundary

The repair changes packaging and reproducibility only. It does not add a theory, change a threshold, run P0, create confirmatory data, or convert any diagnostic into evidence for T^.