# Compass v2 merged BF16 v4 — result and release boundary

Completed 2026-09-10 15:38 Istanbul, independently audited at 15:39.
This report does not modify the original native-adapter benchmark or its scores.

| Metric | Native final adapter | Merged BF16 candidate | Approved gate |
|---|---:|---:|---:|
| Structured first-turn calls | 93/100 | 94/100 | >=93/100 |
| First-turn HTTP/runner errors | 0 | 0 | 0 |
| Strict tool and arguments | 15/89 | 16/89 | >=15/89 |
| Loose/name-only-inclusive match | 43/89 | 41/89 | >=43/89 — FAILED |
| Known anti-pattern | 1 | 1 | <=1 |
| Scenario HTTP/runner errors | 1 timeout | 0 | <=1 |

The candidate passes the mathematical merge, independent weight, native runtime
buffer, finite-weight, save/reload identity, eight-probe FP32 and BF16 greedy/logit
and KV-cache checks. It does NOT pass the complete v4 release contract: loose
matching drops by two cases, from 48.3% to 46.1%. Strict matching rises by one
case, from 16.9% to 18.0%. These small-suite results are not statistical evidence
of broad task performance; no browser actions were executed.

One raw response contains malformed XML (parser reports an invalid token).
This is separately disclosed from the zero HTTP/runner-error count; it was not
repaired or removed from scoring. There were no length-limited responses.
All 194 raw responses and all 100 first-turn / 100 scenario result files are
preserved. Eleven out-of-scope scenarios remain skipped, leaving 89 scored.
All 193 frozen input/code hashes and every recorded scenario reference field
match the original native-adapter run.

First-turn latency sum: 148,464 ms. Scenario latency sum: 175,697 ms.
The run no longer includes the prior 300-second timeout; do not present the
elapsed-time difference as a controlled throughput improvement.

Candidate:
`E:\webbrain-curated-v2-sources\work\curated-v2\compass-v2-model-release-v1\compass-v2-bf16-v4_2-805c995d3797b8a3`

Benchmark:
`E:\webbrain-curated-v2-sources\work\curated-v2\compass-v2-model-release-v1\compass-v2-bf16-benchmark-v4-805c995d3797b8a3`

- Candidate verification SHA256: `a797ba7921c263142d3c0a43204f188e1424b9c51a95639e2682809415564fce`.
- Benchmark completion SHA256: `10039496a9539a2b5aea8a176ce5e87761352da806af2192e702d9420f3b1130`.
- Evidence audit SHA256: `7231b69d180c74bc3f48875c9beb2fc5707628c322d8a79061f8369c0f59de96`.

No model was uploaded and no ONNX export was started. The conditional packager
was not run. The follow-up is paused pending explicit direction. Do not lower
the failed gate, selectively rerun cases, revise expected answers, retrain on
this benchmark or call the candidate fully v4-release-approved.

At 15:41 the exact completed benchmark server was stopped after an idle health,
command-line/ancestry/listener and all-194-responses check. Other servers were
not touched. Shutdown provenance is in server-stop-preflight.json and
server-stopped.json under the benchmark root. All failures and originals remain.