Upstream refresh friction¶
These entries retain causes, corrections and failed attempts. The canonical baseline is closed and restored; none of its success retroactively passes a predecessor. Archive paths and source hashes are enumerated in provenance.
| Event | Immediate disposition / durable correction | Evidence selected or required | Closure boundary |
|---|---|---|---|
| Retained signed GLM template rejected by new boundary contract | Minimal exact-byte legacy-template compatibility path; no package/graph rewrite | Compatibility source patch, commit/build/test provenance and independent frontend review | Code compatibility evidence; not numerical qualification |
| Driver userspace/kernel mismatch after system update | Stop trial, retain failed preview; owner recovery and driver-consistency pre-outage guard | First window, driver recovery/containment/observability receipts and restored direct/routed checks | New driver version is a comparison boundary |
| Preflight collector expected one tool-result observation | Correct exact observation accounting for initial and continuation records; retain rejected wrapper receipt | Closed direct artifact and focused negative checks; revised action/review | Harness contract correction, not model failure |
| Healthy candidate lacked declared versioned-engine monitoring mapping | Withhold metrics and restore; explicit operator mapping correction plus prospective loader check | Second window, monitoring diagnosis/correction and independent action review | Product static-preflight ticket remains open; manual recovery alone does not close product gap |
| SSE canonical package model ID differed from advertised alias | Restore; prospectively pin distinct catalog alias/root and canonical response roles | Third window, source identity diagnosis and canonical collector tests/review | Failed original fixture remains failed; no server-response bug claimed |
| Canonical run | 391 adapter requests plus one smoke passed; exact incumbent restored | Complete native phase and restoration archives, authoritative summary and independent results review | Closed diagnostic baseline; performance qualification remains false |
Private operator data and key references require sanitization. Credential files are never publication inputs. The publisher's static review does not replace review of the selected corpus and staged payload.
The first summary reader failed fixture binding and its rejected output remains archived. Corrected reader/check bytes and HOLD/GO review history remain together; this was evidence processing, not a failed model request.
The product chart CLI rejects the custom native decision schema. Its failed invocation is retained; a minimal offline adapter reuses the unchanged deterministic SVG renderer and binds exact summary metric paths/hashes. No artifact is relabeled as benchmark/v1.