Closure coverage¶
This record separates retained evidence from unmeasured contract requirements. It is not a winner or promotion decision.
| Closure area | Retained / measured state | Boundary and disposition |
|---|---|---|
| Source research | 100 public sources were compiled. | Private Discord leads are excluded; public sources are priors, not local results. |
| Context and topology contract | The intended contract retains 327K, 65K, C4, vision, and Hindsight boundaries. | 131K/C1 scouts do not satisfy those broader requirements. Keep each unmeasured cell explicit. |
| Ignis R4 protocol | Serial C1 preflight passed 6/6. The separate C8 preflight recorded 6/13 passes and seven failures. Twenty-four separate SWE engine-full/all-lanes-occupied 503 retry notices all recovered; they do not explain the empty patches. JSON_OBJECT returned 400 unsupported_value. |
These are separate observation sets. Concurrent admission and JSON structured output remain product-contract limits; retry-only closure is insufficient. |
| Ignis R4 quality | 12/12 strict intelligence/tool/session observations passed with independent deterministic checks; terminal fixed-five SWE attempted five, with two officially graded and resolved and three empty. | The fixed denominator is incomplete and below the 4/5 floor. No full-context, capacity, performance, or promotion proof exists. |
| Qwen control follow-up | Source-aligned top-k-20 log R2 was cancelled before load/query; actual EXL3 controls retained native top-k 0. | No outcome, comparison, or recommendation may be recorded. |
| MiMo | Corrected 2 GiB offload selection followed an erroneous zero selection. | Conversion policy failure is retained; it is neither a fit failure nor a quality result. |
| Operational remediation | PRs 652, 653, 654, and 656 have bounded source/remediation evidence; PR 656's managed candidate path was used. | Global installed-CLI state, source merge, deployment, live acceptance, and promotion remain independent gates. |
| Evidence integrity | R16-r2/R17 repaired the affected public projection; R18 accepted terminal Ignis/lineage and R19 independently replayed 21 restoration projections with no numeric, schema, or privacy drift. | R12, R13, and the first R16 staging are excluded. Publication acceptance is not promotion. |
| Cleanup | Terminal cleanup: 11 model-cache snapshots and 106 exact image IDs removed, with post-removal verification. | Docker logical-byte accounting and native allocated-plan accounting remain separate. Fifty images have retained references; active services, declared rollback/library models, and unsupported plain roots remain protected. This is cleanup evidence only. |
| Restoration and current state | R19 accepted exact retained official-base EXL3 restoration; direct 13/13, routed tools/SSE/JSON_OBJECT, vision/OCR, Pi settled read, Hermes, OpenClaw high 9/9, and Hindsight recall/retain passed. | Thinking-control evaluation remains open: off failed twice and max was rejected locally before query, so unified evaluation is not blanket green. Primary-tier transition attribution remains a separate gap. |
Closure dispositions are open tickets for bounded Docker inspection, prepare/preview help-parser alignment, native SWE evidence-reader schema support, native Linux host-memory support, primary-tier transition attribution, Ignis admission and JSON_OBJECT behavior, and OpenClaw thinking-control evaluation. Defer every non-terminal candidate result. Link only finalized public-safe evidence basenames and public PR URLs in this bundle.