Summary
Render the evidence trust class (first-party / community / partner) in the Evidence
column of docs/user/recipe-health.md, so a reader can tell the two apart without opening
recipes/evidence/allowlist.yaml.
Why it is not visible today
The Evidence cell is generated by tools/health from pkg/testgrid/presence.yaml, which
carries only coordinate paths — no signer identity and no class. The class lives in a different
file, recipes/evidence/allowlist.yaml, keyed by the source slug that also names each pointer
directory (recipes/evidence/<recipe>/<slug>/). Nothing joins them, so the cell renders a
deep-link or pending and says nothing about who signed.
The join itself is cheap: the slug is the directory name, and pkg/evidence/allowlist
already exposes Classify(issuer, identity) with slug support.
Current state, for the record
Every committed pointer under recipes/evidence/ is community class:
| Signer |
Attestations |
yuanchen8911/aicr fork workflow (slug 5bf9e82f…) |
23 |
yuanchen97@gmail.com |
3 |
ramessesii2 |
1 |
atif1996 |
1 |
None is first-party, although allowlist.yaml declares four first-party NVIDIA/aicr UAT
workflows. That includes both v1 Supported GB300 coordinates.
This is the designed model, not a defect — contributors publish from a fork per
evidence-publishing.md,
and maintainer review of the allowlist PR is the trust gate.
Why this is post-v1
Deliberately deferred rather than rushed before the v1 tag, for two reasons:
- It lands in a slot that is already deferred by design.
recipe-health.md records that
ADR-009's in-cell freshness state (unattested vs aged) is deferred and that "an optional
freshness token can later annotate the same cell without replacing the link". A class token
is the same cell. Designing two annotations for it independently, weeks apart, is how the
cell ends up with two competing conventions.
- The honest version is already shipped.
recipe-health.md now states the concrete
posture — all committed evidence is community class, none is first-party — and ROADMAP §2
records that trust class is not part of the v1 Supported definition. That closes the
information gap without committing to a rendering.
Scope
Separate question this does not decide
Whether Supported should require first-party attestation. ROADMAP §2 now says it does not
for v1 and calls tightening it a post-v1 question. That is a policy decision about the bar;
this issue is only about making the current bar legible.
References
recipes/evidence/allowlist.yaml, pkg/evidence/allowlist
tools/health/main.go (evidenceCell), pkg/testgrid/presence.yaml
- ADR-009
Summary
Render the evidence trust class (
first-party/community/partner) in the Evidencecolumn of
docs/user/recipe-health.md, so a reader can tell the two apart without openingrecipes/evidence/allowlist.yaml.Why it is not visible today
The Evidence cell is generated by
tools/healthfrompkg/testgrid/presence.yaml, whichcarries only coordinate paths — no signer identity and no class. The class lives in a different
file,
recipes/evidence/allowlist.yaml, keyed by the source slug that also names each pointerdirectory (
recipes/evidence/<recipe>/<slug>/). Nothing joins them, so the cell renders adeep-link or
pendingand says nothing about who signed.The join itself is cheap: the slug is the directory name, and
pkg/evidence/allowlistalready exposes
Classify(issuer, identity)with slug support.Current state, for the record
Every committed pointer under
recipes/evidence/iscommunityclass:yuanchen8911/aicrfork workflow (slug5bf9e82f…)yuanchen97@gmail.comramessesii2atif1996None is
first-party, althoughallowlist.yamldeclares four first-party NVIDIA/aicr UATworkflows. That includes both v1 Supported GB300 coordinates.
This is the designed model, not a defect — contributors publish from a fork per
evidence-publishing.md,
and maintainer review of the allowlist PR is the trust gate.
Why this is post-v1
Deliberately deferred rather than rushed before the v1 tag, for two reasons:
recipe-health.mdrecords thatADR-009's in-cell freshness state (
unattestedvs aged) is deferred and that "an optionalfreshness token can later annotate the same cell without replacing the link". A class token
is the same cell. Designing two annotations for it independently, weeks apart, is how the
cell ends up with two competing conventions.
recipe-health.mdnow states the concreteposture — all committed evidence is community class, none is first-party — and ROADMAP §2
records that trust class is not part of the v1 Supported definition. That closes the
information gap without committing to a rendering.
Scope
tools/healthto resolve a coordinate's pointer directory to its allowlist classdeferred freshness token rather than independently of it
docs/user/recipe-health.mdand update the golden expectations intools/health/main_test.goSeparate question this does not decide
Whether Supported should require first-party attestation. ROADMAP §2 now says it does not
for v1 and calls tightening it a post-v1 question. That is a policy decision about the bar;
this issue is only about making the current bar legible.
References
recipes/evidence/allowlist.yaml,pkg/evidence/allowlisttools/health/main.go(evidenceCell),pkg/testgrid/presence.yaml