2026-10-19 Ignis 0.7.2 focused qualification¶
Local outcome: bounded serving checks passed; coding qualification failed. The candidate was removed and the incumbent restored.
Setup: direct online requests to Qwen3.8 Flash Next, Ignis 0.7.2, one 96 GB RTX PRO 6000 Blackwell Max-Q, 131,072 configured context, eight decode lanes, MTP3, and a 52 GiB container limit with no swap. See the sanitized managed recipe. Artifacts had recently been downloaded/read; startup is not a cold-load measurement.
| Check | Local observation | Limit |
|---|---|---|
| C1 protocol | 5/5 named checks passed | Text and tools only |
| C8 tool batch | 8/8 passed | One short, shared-prefix batch |
| Long context | Correct at 100,739 and 103,256 actual prompt tokens | Not full 131,072-token or simultaneous capacity |
| SWE coding | 265 HTTP 200 responses; one graded unresolved patch, four empty ungraded submissions | Run incomplete; no qualification pass |
| Timeout diagnostic | Partial output at a configured two-second deadline; same-lane recovery | Default 2,400-second expiry not exercised |
| Restoration | Direct and routed preflights each passed all five checks | No candidate promotion |
The serving improvements did not translate into successful coding qualification in this fixed five-task run. Artifact manifest · Evidence index.
Decision: no-promotion. Ignis 0.7.2 was exercised as an unpromoted candidate, then removed; the incumbent was restored and passed direct and routed five-check preflights.
The pinned Qwen3.8 Flash Next candidate used Ignis 0.7.2, one RTX PRO 6000 Blackwell Max-Q, 131,072 configured context tokens, the maintainer’s RTX PRO 6000 profile, and eight decode lanes. C1 passed 5/5 named checks; the explicit shared-prefix C8 tool workload returned 8/8; two rerun long-context probes used 100,739 and 103,256 prompt tokens. These are bounded functional observations, not capacity, p99, or full-context claims.
A 112.120-second request stopped at its 16,384-token cap and does not cross the former 120-second default. A separate 2-second diagnostic returned partial length results for nonstream and stream requests, correlated server timeout cancellation, and same-lane recovery. It demonstrates the scaled timeout behavior only; the candidate profile’s configured 2,400-second deadline was not allowed to expire.
SWE evidence is incomplete. The initial five-task run made no requests because its client dependency contract was incompatible; source revision 803e44e2 made the exact dependency pins durable. The corrected run attempted five tasks, received 265 HTTP 200 responses with no observed retries or transport errors, and ended with one officially graded unresolved Django patch plus four empty, ungraded 60-step LimitsExceeded submissions. This is not a fully graded 0/5 score and does not qualify the candidate.
The evidence packet preserves native artifacts, failures, public-copy hash provenance, and restoration receipts. The incumbent restoration recorded expected identities, unchanged router mount/image/hash, candidate containers absent, no swap or OOM, and direct/routed preflight 5/5.
Structured response_format: json_object remains unsupported (HTTP 400), separately from the successful JSON-prompt check. Vision was disabled. Four deterministic microchecks passed three repetitions each; this does not outweigh the coding failures. No matched performance comparison was run, so this campaign makes no causal MTP speedup claim.
The harness dependency fix passed 109 focused tests and independent review. The full source suite stopped at a Connect fixture error after 857 passes and five skips; it did not pass in full. The cache-inspector image issue remains a tracked product gap: the complete pinned cache was verified with a managed explicit inspector, then automatic verification was disabled for the trial and restoration. The restoration’s runtime settings matched the original; its verification setting and recipe labels differed, as recorded in the restoration evidence.