Governance

Governance is six primitives, not policies. Industry profiles are editable templates with no legal claim.

Primitive	What it does
Gates	Expressions such as agreement >= 0.90 and shadow_days >= 14 over cv_metric, cv_f1, rows, agreement, shadow_days, fallback_rate, drift, llm_metric, min_class_count. Parsed by a small safe parser: no eval.
Approvals	{groups, users, require: all\|any\|<n>, via: manual \| jira:PROJ \| github:owner/repo}. Satisfied when the ticketer (or pecca approve) reports the required principals.
Evidence	Audit-pack sections that must exist for the version (model_card, confusion_matrix, data_lineage, threshold_rationale, calibration, per_class_report, drift_report, human_oversight_log).
Retention	{evidence: 7y, logs: 2y, signed: true} is stored and written into the audit pack. v0.1 does not delete anything. signed writes a sha256 manifest (plus a detached GPG signature if gpg_key is set).
Schedule	revalidate and drift_review intervals; train_every per call. pecca scheduler run evaluates what is due.
Controls mapping	Free-form strings attached to transitions; rendered in the audit pack. No semantics.

governance:
  extends: default
  transitions:
    shadow->live:
      gates: ["agreement >= 0.90", "shadow_days >= 14"]
      approvals: {groups: [model-risk], require: all, via: jira:MRM}
      evidence: [model_card, confusion_matrix, data_lineage, threshold_rationale]
      controls: ["ISO42001-8.4"]
  retention: {evidence: 7y, logs: 2y, signed: true}
  schedule:
    revalidate: {every: 365d}
    drift_review: {every: 30d, notify: "slack:#ml-governance"}

pecca promote returns a Decision(allowed, failing_gates, pending_approvals, evidence_missing). --force needs PECCA_ALLOW_FORCE=1 and is recorded in the call's history and the audit pack.

Note

Gates assume higher is better for cv_metric. For regression calls (rmse), write lower-is-better gates, e.g. cv_metric <= 5.

Audit pack

pecca audit PATH [--version vN] [--out markdown|pdf|confluence|json] writes one Markdown file per section (summary, model card, data lineage, benchmark leaderboard, confusion matrix, per-class report, calibration, threshold rationale, shadow results, approvals, drift report, controls mapping, environment, manifest) plus charts and a manifest.sha256.

PDF export is untested

--out pdf needs pip install 'pecca[pdf]' plus system libraries (pango). It has not been verified in v0.1; prefer Markdown, JSON or Confluence.