Skip to content

GLM runtime investigation evidence

This bundle supports the dated finding. The summary is the authoritative source for displayed results; coverage lists the unrun gates. Three matched DCP1 175K/32K overlaps passed after the DCP2 parent crashed. This is bounded mitigation evidence, not kernel attribution, a reliability rate, performance ranking or replacement qualification. The exact baseline is restored.

Artifact manifest binds every retained byte; manifest source reproduces it. Native stability artifacts retain failures, usage, response IDs and bounded outputs as originally collected. The initial parent and 512-cap runs predate the result-retention fixes; see friction and correlation.

Sanitization receipts bind private source hashes to public copies. Generic paths, ports and GPU identities are reconstruction placeholders. Container IDs are consistently redacted with one-way tokens; the evidence inspector marks their provenance as sanitized. Recipe/registry digests inside native evidence identify the original private configuration; recalculate them when loading a local copy. The endpoint placeholders must not be used as proof that a recipe is running.

No graph or social-copy package is required: these are diagnostics without comparable performance populations. All files are text; raw evidence stays below 1 MiB per file and 5 MiB for this bundle.

Source registry

Workload manifest

Run plan

Configuration and identity

Raw run evidence

Failures and friction

Restoration

Decision summary

Independent final source and publication review accepted the reviewed bytes for exact PR-head CI. Its manifest hash deliberately records the bundle before insertion of this review packet.