Goal
Every recipe AICR ships must be validated end-to-end: it must resolve from criteria, ship a runtime performance gate anchored in empirical thresholds, and execute against the validator capability its workload kind requires. A recipe must not be passing because it silently skipped a gate.
This is a multi-quarter strategic outcome — the first AICR Initiative. It does not have a single landing PR. It closes when each Epic underneath it closes.
Why this is an Initiative, not an Epic
The "Recipe validation parity" outcome cleanly factors into three independent deliverable streams:
- Overlay coverage — authoring-blocked work on
recipes/overlays/.
- Performance phase — testbed-blocked work on empirical thresholds.
- Validator capability — engineering-blocked work on
validators/*.
Each stream has its own owner shape, blocking dependency, and "done" definition. Grouping all leaf issues directly under a single tracker conflated those streams; the Epic tier separates them so each can be sequenced and reviewed independently.
This Initiative also sets the organizational precedent for Initiative → Epic → Issue in this project:
- Initiative — strategic outcome, no single PR, lives until the outcome is real
- Epic — deliverable chunk with concrete success criteria, may span multiple PRs
- Issue / Enhancement / Bug — single tactical work unit, typically one PR
Success criteria
The Initiative is done when:
recipes/overlays_validation_floor_test.go passes with AICR_VALIDATION_FLOOR_STRICT=1 and zero exemptions tracked to open issues.
- For every accelerator in
pkg/recipe/criteria.go, aicr recipe --accelerator <x> --service <y> --intent <z> resolves a recipe end-to-end.
- Every validator referenced by an overlay's
validation.phases.*.checks actually runs against the workload kind that overlay ships (no silent skipped paths).
Sub-issues (Epics)
New work that fits the scope of one of these Epics should be filed as a standalone issue and attached as a sub-issue of the appropriate Epic. New work that doesn't fit any of the three (e.g., a new gating axis surfaced by a future audit) is filed as a new Epic and attached here.
Explicitly out of scope
These belong to other initiatives, not this one:
- Validation framework / runtime changes (validators package architecture, agent wiring).
- Recipe data model changes (overlay schema, mixin shape).
- Conformance certification work (downstream consumer of validated recipes).
Tracking sources of truth
- Strict-floor:
recipes/overlays_validation_floor_test.go (AICR_VALIDATION_FLOOR_STRICT=1).
- Coverage:
pkg/recipe/criteria.go cross-referenced with recipes/overlays/.
- Validator capability: per-check
validators/<phase>/*.go skip / pass logic.
Goal
Every recipe AICR ships must be validated end-to-end: it must resolve from criteria, ship a runtime performance gate anchored in empirical thresholds, and execute against the validator capability its workload kind requires. A recipe must not be passing because it silently skipped a gate.
This is a multi-quarter strategic outcome — the first AICR Initiative. It does not have a single landing PR. It closes when each Epic underneath it closes.
Why this is an Initiative, not an Epic
The "Recipe validation parity" outcome cleanly factors into three independent deliverable streams:
recipes/overlays/.validators/*.Each stream has its own owner shape, blocking dependency, and "done" definition. Grouping all leaf issues directly under a single tracker conflated those streams; the Epic tier separates them so each can be sequenced and reviewed independently.
This Initiative also sets the organizational precedent for Initiative → Epic → Issue in this project:
Success criteria
The Initiative is done when:
recipes/overlays_validation_floor_test.gopasses withAICR_VALIDATION_FLOOR_STRICT=1and zero exemptions tracked to open issues.pkg/recipe/criteria.go,aicr recipe --accelerator <x> --service <y> --intent <z>resolves a recipe end-to-end.validation.phases.*.checksactually runs against the workload kind that overlay ships (no silentskippedpaths).Sub-issues (Epics)
New work that fits the scope of one of these Epics should be filed as a standalone issue and attached as a sub-issue of the appropriate Epic. New work that doesn't fit any of the three (e.g., a new gating axis surfaced by a future audit) is filed as a new Epic and attached here.
Explicitly out of scope
These belong to other initiatives, not this one:
Tracking sources of truth
recipes/overlays_validation_floor_test.go(AICR_VALIDATION_FLOOR_STRICT=1).pkg/recipe/criteria.gocross-referenced withrecipes/overlays/.validators/<phase>/*.goskip / pass logic.