Skip to content

Qualification evidence summary

The named recipe passed bounded qualification. Deployment, routing and client assignments are outside this public record.

Completed retained results:

  • Historical r4 ordinary-summary comparison: 432 requests across two arms. At about 7.4K input, MTP decode p50 was 67.3 versus 25.4 tokens/s at C1, and 40.0 versus 19.5 at C2. This is a workload-specific observation.
  • Fresh r4 direct image evidence: 12/12 deterministic cases passed and the ordered eight-image probe passed 2/2. Video remains excluded.
  • Fresh r4 agentic evidence: 18/18 full-suite cases and 6/6 pilot cases passed.
  • Fresh r4 C1 context evidence: all 36 cases passed with stop finishes. The 18 near-full cases measured 245596–245626 input tokens, which is at least 99.8046875% of the 245760 bucket; each used a 16384-token output reserve. These quality requests overlapped coding work, so their latency is excluded from performance comparisons.
  • Fresh r4 SWE fixed campaign floor: all five tasks completed independent grading; four resolved and Django10097 remained unresolved. All five agent submissions reached Submitted, with no infrastructure failure. This is a five-task floor, not a broad coding-benchmark population estimate.
  • Fresh r4 C2 concurrent context evidence: two distinct-prefix long-context requests both passed exact-answer scoring and overlapped for 260.115 seconds. During that overlap, retained samples reported running=2 and peak used KV of 458520; the two actual prompt counts plus their output reserves totalled
  • This is observed pair evidence, not a hypothetical full-residency or general capacity claim.
  • R5 at .94 memory allocation was rejected by the required GPU reserve policy; r4 at .90 was restored. This is not an OOM claim.

Limits carried forward:

  • Strict controlled-output diagnostic: 4/4 failures, performance-ineligible.
  • Historical MTP C16 had an effective active cap of 9; no-spec had 16. C1–C8 are the matched portion of the historical comparison.
  • Historical r4 performance activation differs from the fresh final r4 activation. No metric is carried over as final-activation performance proof.
  • Final stability, capacity, and resource-window evidence passed; the fresh final population is separate from the historical comparison.
  • The C2 amendment's receipt digest follows its source canonical-JSON contract; the redaction ledger separately binds original and staged physical file-byte SHA-256 values and does not substitute one digest type for the other.

redaction-ledger.json records original and derived SHA-256 values. The staged JSON preserves native schemas where present but is a redacted derivative, not byte-identical raw native evidence.