Closure friction log¶
This records campaign friction and its disposition. It is not a promotion result; no private operator material is included.
| Failure or limit | Durable disposition | Evidence / status |
|---|---|---|
| RAM admission depended on a nonzero swap threshold despite available RAM. | PR 652 added the zero-swap containment path; it is a source fix, not a global-host setting change. | coder390-tp1-131k-zero-swap-load.json; closed for that admission defect. |
| Quality execution treated a timeout as a quality result. | PR 653, merged at 2bdbbd44e71ddf12c957784f3c9bd1ea118c4539, separates timeout from a scored result. |
quality-timeout-fix-pr-merged.json; source merged, campaign/runtime acceptance remains separate. |
| Large-file acquisition lacked the required Xet path. | PR 654, merged at 59bb05d84cf21b6eec803d721d3c65ce96ef4612, repaired acquisition. |
xet-fix-pr-merged.json; source merged, not a model qualification result. |
| Non-root model-cache access blocked the initial Ignis load. | PR 656, merged at fdd956ed1e4d81d348f08c89f9447a1bfd901638, adds an optional non-root cache mount that was used by the managed candidate path. |
nonroot-cache-mount-merge-reconciliation.json; reviewed source CLI was actually used; global installed CLI remained unchanged. |
| Tabby consumed a schema-incompatible artifact layout. | Schema-aware operator patch 997 is the fix-forward; retain the original identity and the corrected parser evidence. | tabby-qwen3-coder-schema-fix-independent-review.json; do not infer broader compatibility. |
| MiMo first selected zero offload, then the corrected 2 GiB selection. Conversion still failed policy/trust gates. | Keep the corrected selector and classify the conversion result as a policy failure, not a physical-fit result. | mimo-mopd-nvfp4-tp2-32k-failure-decision.json, mimo-routed-offload-r6-failure-decision.json; unresolved candidate. |
| Ignis initially lacked a conversion sidecar and later exposed a health-path 404. | Recipes were repaired; the retained R4 receipt shows bounded readiness only. | ignis-trellis25-tp1-131k-failure-decision.json, ignis-sidecar-r3-health-path-diagnostic.json, ignis-health-r4-native-ready-gates.json; no promotion conclusion. |
C8 protocol preflight passed 6/13 with seven immediate 503s; serial C1 passed 6/6. Separate SWE engine-full retry notices recovered 24 times; JSON_OBJECT returns 400 unsupported_value. |
Open ticket: concurrent admission and structured-output contracts remain distinct. | ignis-health-r4-preflight-native.json, ignis-health-r4-json-object-boundary-response.json; open. |
Observer state, unwrap, and extra -- launch defects made some earlier receipts ambiguous. |
Preserve errata and use the corrected observer/launcher records; do not rewrite earlier evidence. | qwen-exl3-schema-r1-swe5-observer-state-erratum.json, qwen-qad-tp1-131k-kv6-observer-erratum.json; bounded correction only. |
| An invalid SWE paired-image map invalidated that run binding. | Retain the map erratum and exclude the affected run from outcome claims. | deepseek-swe-paired-image-map-erratum.json; open until a valid terminal map exists. |
| R12 public staging over-redacted semantic slash-bearing fields. | R16-r2/R17 repaired the audited projection. R18 independently accepted the final public-native lineage and terminal Ignis records; R12, R13, and the first R16 staging remain excluded. | stock-qwen27-fp8-tp1-bf16kv-131k-xhigh-swe5-native.json, ignis-health-r4-swe5-r2-native.json; publication staging only. |
| Docker process inspection is unbounded. | Open ticket: no broad diagnostic command becomes the supported workflow. | Deferred/watch. |
| CLI help and parser behavior disagree for a scoped prepare/preview path. | Open ticket: use the reviewed managed path only. | stock-qwen27-fp8-tp1-bf16kv-131k-xhigh-swe5-preview-erratum.json; open. |
| Native evidence reader rejects a retained native schema. | Open ticket: manual custody does not convert to reader acceptance. | Deferred/watch. |
| Linux host-memory command is WSL-only. | Open ticket: retain measured native memory evidence without claiming the unsupported command passed. | Deferred/watch. |
| A quiesced primary tier readmitted unexpectedly; actor remains unproven. | Open ticket: re-quiesce before any identity-fenced action. | router-quiesce-drift-analysis.md; open. |
| Routed OpenClaw thinking-off failed twice with reasoning visible; high passed 9/9, max was rejected locally before query, and unified evaluation hardcodes off. | Open ticket: preserve controls and report the selected evaluation mode. | restoration-summary-r19.md; open. |
The terminal Ignis fixed-five SWE run attempted five cases: two were officially graded and resolved, and three were empty. It is incomplete on the fixed denominator and below the 4/5 floor, so it supplies no promotion result. The Qwen source-aligned top-k-20 log R2 was not loaded or queried and supplies no comparison or improvement claim. Terminal cleanup verified 11 snapshot removals and 106 exact image-ID removals; the earlier Ornith removal is included once. Docker logical bytes and native allocated-plan bytes remain separate. The campaign compiled 92 core public sources plus eight public community-addendum sources; Discord material was discovery-only and is excluded.
Native-cache apply was initially unavailable, and the safety audit exposed real configuration formats plus persistent unlocked Hugging Face lock files. The managed cache-removal command now audits decoded YAML and relative dependencies, permits only an explicitly declared read-only non-model root view, and fences actual POSIX locks through deletion. The reviewed source passed 89 focused tests plus independent safety probes and was used for the five native removals. Global installed CLI state remained unchanged. Persistent lock paths were preserved; external writers must be quiescent.