EVR-6 real-distribution attempt 1

Historical result — 2026-09-04 attempt 1. Preserve this attempt's evidence and HOLD
verdict. The auditor gate defines the current requirements for any
separately declared follow-up; this result does not authorize enabling the runtime flag.

• Date: 2026-09-04
• Verdict: HOLD
• Runtime decision: keep per_turn_auditor default OFF

Method and provenance

The run started from fetched origin/develop commit
58f9a7c7f3e782cfeb3fd0aad13f2ac6a6a4626c. The original pending review pack was
re-exported from the four source manifests below and reproduced its recorded SHA-256 exactly.
The independent reviewer's exact edits were recovered from the 2026-08-29 reviewer session and
applied to that byte-identical base. The resulting reviewed-pack SHA-256 matched the digest
recorded when the blinded review completed. No labels were inferred from aggregate counts, model
observations, or benchmark outcomes.

Family	Source manifest	SHA-256
code-fix	eval_manifest_6fb27261da29.jsonl	45c328d21be07dd91593642916526a6afcddcc58cfc246f58a007b71aebdeeaf
code-gen-humaneval	eval_manifest_770e78c17603.jsonl	696ea8d451e94d93a6982268d13e90d98cf26b688ca3d317e9a37bb03d7ee159
code-gen-mbpp	eval_manifest_4f3f8b0cf5f4.jsonl	989d0f63386a80155bf5d4532b4c8149ae05fc4df03d71c74c042838e3d3f01c
deep-research-dr3	eval_manifest_86a0e7ed90e1.jsonl	960d8654759178722be8cc665583f6be18209307d2110c008b661e6ddffe6862

The versioned oracle source is codex-independent-reviewer-v1/2026-08-29. Finalization removed
all reviewer-only context and confirmed that the label pack contains no auditor_id or
observations.

Frozen labels

Family	Included	Alarm-positive	Healthy
code-fix	10	4	6
code-gen-humaneval	6	4	2
code-gen-mbpp	6	5	1
deep-research-dr3	6	6	0
Total	28	19	9

The preregistered family-coverage requirement is already unmet because DR3 has no healthy trace.
Cases were not replaced after observing this distribution.

Replay integrity

• Auditor: ollama:qwen3.5:2b@sha256:324d162be6ca5629ae4517c8710434d0bd2d665bc94dbad46e9af8fbf8a2f0df.
• The tag resolved to that digest before replay and to the same digest after replay.
• Decision caching was disabled (cache_ttl=0; evidence metadata cache_enabled=false).
• All 28 task IDs are unique and all 245 HTIR prefixes have exactly one ordered observation.
• The evidence payload preserves the finalized label payload exactly, apart from adding the
  top-level producer/auditor identity and per-case observations.
• A separate turn_auditor_eval invocation produced a byte-identical report.

Result

Metric	Result	Gate	Outcome
Traces	28	≥24	pass
Alarm-positive / healthy	19 / 9	≥8 / ≥8	pass
Families	4, each ≥4	≥4, each ≥4	pass
Both polarities per family	DR3 has 6 / 0	required	fail
Precision	0.000	≥0.800	fail
Early-alarm recall	0.000	≥0.800	fail
Per-family recall	0.000 in all four	≥0.800	fail
Healthy false-alarm rate	0.000 overall and per family	≤0.050	pass
p95 auditor latency	1,639.2016 ms	≤2,000 ms	pass

The model produced zero early true positives, 19 false negatives (11 misses and 8 late alarms),
and zero false positives. The evaluator recorded these exact HOLD reasons:

1. code-fix recall was below 0.800.
2. code-gen-humaneval recall was below 0.800.
3. code-gen-mbpp recall was below 0.800.
4. deep-research-dr3 lacked both positive and negative oracle traces.
5. Overall precision was below 0.800.
6. Overall recall was below 0.800.

This is an offline prerequisite result only. It does not authorize enabling the runtime flag or
running the graduation A/B. Any resampling must be declared and independently reviewed as a
separate attempt without changing these thresholds.

Artifact custody

The complete bundle is retained in the maintainer evidence store at
~/.victor/evaluations/evr6/attempt-1-2026-09-04/. The raw manifests and context-bearing review
pack remain local as required by the review-pack confidentiality guidance. Digests below make the
bundle independently verifiable.

Artifact	SHA-256
review-pack-original.json	64804d029f24ca190329f2d334ba9a122e5fc339046d5040de7dd287657f0eef
review-pack-reviewed.json	bccd89a4449b60c35de8945b7c98f151488f7864d9498c04dec3b6922681ae60
labels.json	661ddc959db109941d24d0e9eb564aa8f1782c884c8bd490aee81bf896400195
evidence.json	b66a1059125e64682aab7bc4c7715ee1dce048f90f36505ea1c3242d3db10d93
report.json	ae9107cf1bf5b9ab4af6daec2959aeb9e42935ff87c4f04124e5950c3446f41a
turn-auditor-report.json	ae9107cf1bf5b9ab4af6daec2959aeb9e42935ff87c4f04124e5950c3446f41a

The evidence producer additionally records the canonicalized label payload as
sha256:170593f11cebc5c6e428c76a88cbef723159770db9a68a9d7da8eae3e38ffa05.