calibration-pipeline

name	calibration-pipeline
description	Calibration and review-pipeline tooling guidance. Use when running or maintaining calibration scripts, review aggregation, and related deterministic assets for issue #360. DO NOT USE FOR: changing orchestration ownership or agent review policy.

Calibration Pipeline

Reusable entryway for deterministic tooling, committed assets, and pipeline-metrics reference material for the calibration pipeline introduced by issue #360.

Survival: SMC-09 covers raw aggregate/prosecution-depth snapshots as per-dispatch until selected PR-body metrics are durable. SMC-10 covers the within-worktree per-finding cache; SMC-11 covers within-worktree cross-PR state. Script temp files are per-process scratch only.

When to Use

• When running or maintaining calibration writers, aggregation scripts, or helper libraries under scripts/
• When updating committed calibration data or configuration under assets/
• When a workflow needs the canonical pipeline-metrics schema, verdict mapping, or findings-construction rules without re-reading Code-Conductor

Purpose

• Describe the calibration and review-aggregation domain
• Provide the stable home for pipeline scripts under scripts/
• Provide the stable home for committed pipeline data under assets/
• Metric segmentation (sustain-rate by confidence tier) needed to make pre/post coverage-first sustain_rate readings comparable is deferred to umbrella #761 — see the ## Gotchas entry below

Contents

• scripts/ contains calibration writers, aggregation tools, improvement-issue helpers, and related shared helpers
• assets/ contains committed pipeline configuration and data consumed by those scripts
• references/ contains extracted pipeline-metrics contract material that other skills and agents can load directly
• schemas/phase-containment.schema.json is the validation schema for <!-- phase-containment-{ID} --> ledger entries; the ledger's warn-only enforcement surface is .github/scripts/phase-containment-emission-check.ps1 (hub maintainers only — repo-relative, does not resolve from a consumer repo's CWD), which sweeps code-review, design-challenge, and plan-stress-test surfaces for judge-sustained findings missing a paired ledger block and reports the gap; it never blocks a caller flow

Composite References

• references/metrics-schema.md: canonical ## Pipeline Metrics block shape, top-level field semantics, and the write-calibration-entry.ps1 invocation contract
• references/verdict-mapping.md: judge verdict to metrics-field mapping for main and post-fix review
• references/findings-construction.md: findings-array construction, backward compatibility, malformed-entry handling, and rework-cycle population rules
• references/release-hygiene-credit-emission.md: v4 release-hygiene credit row construction (state-file reading, YAML examples) and CE Gate S2 synthetic-PR test protocol (issue #441 Step 7b)
• references/review-credit-emission.md: v4 review credit row construction via Build-ReviewCreditRow — adapter selection, RunIndex semantics, cross-tool handoff, and CE Gate S3 coverage (issue #441 Step 8b)
• references/conductor-credit-emission.md: Code-Conductor-owned credit-row emission procedures — Pipeline-Entry Credit Harvest (SMC-17), Deferred Port Credit Rows, process-review trigger-absent emission, and the Post-PR Credit Row (D10 category 3)
• references/conductor-metrics-protocol.md: emission timing and ownership boundary, Add-FCLCreditRow accumulator pattern, release-hygiene and review credit row pointers, and dispatch-cost samples — extracted from agents/Code-Conductor.agent.md via D5 compression (issue #774)

Boundary

• This skill owns deterministic tooling and static assets only
• Review orchestration, judgment, and step execution stay with the owning agents and existing methodology skills

Gotchas

Trigger	Gotcha	Fix
Adding review pipeline helpers	Tooling and orchestration responsibilities get mixed together	Keep scripts and assets here, but leave routing and judgment with the owning agents
Reading sustain_rate trends across the #784 coverage-first prosecution change (dated 2026-07-02)	Coverage-first prosecution reports more low-confidence findings that were previously silently omitted, mechanically lowering sustain_rate even though prosecutor quality is unchanged	Treat sustain-rate readings from before vs. after the coverage-first epoch as non-comparable; do not trigger prosecution-depth auto-tuning on a low-confidence-driven sustain-rate dip. Metric segmentation (sustain-rate by confidence tier) is deferred to umbrella #761.