Skip to the content.

Prospective exact-model v3: capability PASS

Issue: #4633
Declaration issue: #4845
Production-readiness parent: #4632

Verdict

PASS. The declaration was committed before provider observation in aee4633baa089b564cafe6bf38da038c9418c3e3 and its corrected timestamp was published in 695626994a. On 2026-08-05 the fixed campaign completed exactly 18 authenticated attempts with no replacements. Every attempt used the declared exact model ID, satisfied the complete-line sentinel contract, made two valid tool calls, and stayed within the predeclared correctness, provider-error, invalid-tool, latency, token, and cost thresholds.

This is a bounded capability witness for the declared task classes and tiers. It is not evidence for unrelated protocol/cache, pricing, canary, or rollback gates.

Preregistered envelope

Observed result

Exact model Tier Samples Success Provider errors Invalid tools p95 latency Avg input tokens Avg cost Verdict
claude-opus-4-8 T0 6 100% 0% 0% 9,725 ms 5.0 $0.024027 PASS
claude-sonnet-4-6 T1 6 100% 0% 0% 13,943 ms 4.5 $0.021247 PASS
claude-haiku-4-5-20251001 T2 6 100% 0% 0% 6,260 ms 22.0 $0.013846 PASS

The completed machine-readable artifact is examples/model-acceptance-prospective-v3-report.json. It retains every successful or failed scheduled attempt; this run had no failures to classify. Runtime enforcement remains the fail-closed dispatch seam shipped by #4799: above-tier, absent, stale, malformed, alias, or HOLD evidence is refused before the provider continuation unless an explicit audited operator override is supplied.

Immutable evidence

Raw streams remain in operator scratch because they contain provider/session metadata. The public artifact contains the scrubbed fold and these bindings make the private streams independently detectable if changed:

The v2 weekly-limit HOLD remains immutable infrastructure history and is not counted as a capability sample in this campaign.