docs(governance): propose forensic output conventions (produced_by + process_breakpoint)

audits/promoted/AF-20260614-forensic-output-authorship.yaml

@@ -0,0 +1,37 @@
+audit_finding:
+  version: "1.0"
+  id: "AF-20260614-forensic-output-authorship"
+  date: "2026-06-14"
+  auditor: "external-contributor (@aiox-master in a consumer project)"
+  source_session: "Comparative audit of a hardened consumer fork against aiox-core v5.2.9"
+
+  context:
+    project: "a production consumer project (anonymized)"
+    epic: "framework learning-loop audit"
+    triggered_by: "The framework's evolution-pipeline needs provenance and recurring-breakpoint signals it cannot currently capture"
+
+  finding:
+    summary: "Agent-produced artifacts carry no authorship/provenance metadata, and failed gates record the symptom rather than WHERE in the process the error was born — so recurring issues get re-fixed at the symptom and the loop never learns."
+    evidence:
+      - "No produced_by header on stories/gates/reports → cannot tell which agent/model produced an artifact"
+      - "Gate findings describe the defect, not the process step that produced it (story / handoff / input / environment)"
+      - "Without a breakpoint tag, the same root cause recurs and is patched again at the symptom"
+    impact_observed:
+      blast_radius: "medium"
+      affected_artifacts: ["stories", "quality gates", "reports", "the evolution-pipeline's own inputs"]
+      cost: "Lost provenance for model-tiering decisions; recurring defects re-fixed instead of mechanized"
+
+  framework_candidate: true
+  framework_candidate_rationale: |
+    Project-agnostic. Any agentic framework with multiple models and quality gates
+    benefits from provenance (produced_by) and root-cause tagging (process_breakpoint).
+    Both are presence-of-field predicates — mechanizable as advisory, no judgement needed.
+
+  references:
+    - kind: pattern
+      path: "Forensic authorship + root-cause tagging"
+
+  proposed_disposition:
+    - "Add produced_by frontmatter convention to agent-generated artifacts"
+    - "Add process_breakpoint to every non-PASS gate finding"
+    - "Adopt 'a recurring breakpoint is fixed at the process, not the symptom' as a closing rule"

governance/proposals/PROP-20260614-forensic-output-authorship.yaml

@@ -0,0 +1,58 @@
+framework_proposal:
+  version: "1.0"
+  id: "PROP-20260614-forensic-output-authorship"
+  date: "2026-06-14"
+  proposer: "external-contributor (consumer-project audit)"
+  source_finding: "AF-20260614-forensic-output-authorship"
+
+  target:
+    layer: "L2"
+    artifact_type: "rule"
+    artifact_path: ".claude/rules/output-conventions.md (NEW) + produced_by frontmatter in story/gate templates"
+    operation: "create"
+
+  generalization:
+    pattern_name: "Forensic authorship + root-cause tagging"
+    when_to_apply: |
+      Any project where multiple agents/models produce artifacts (stories, gates,
+      reports, docs), and where failed gates should teach the PROCESS instead of
+      being re-fixed at the symptom. Two conventions:
+        1. produced_by — every agent-generated artifact carries
+           `produced_by: { agent, model, date }` in its frontmatter/header.
+        2. process_breakpoint — every gate with a non-PASS verdict tags, per issue,
+           WHERE in the process the error was born (incomplete story? context-less
+           handoff? unvalidated input? environment?). "Fix the process, not the symptom."
+    examples_from_other_domains:
+      - "Multi-model pipelines: knowing which model produced an artifact is the input to model-tiering / validation sampling"
+      - "Quality gates: tagging where a defect was born reveals recurring process breakpoints to mechanize"
+      - "Audit trails: 'who/which model produced what' is forensic provenance for any agentic system"
+
+  migration_path:
+    breaking_change: false
+    affected_consumers: []
+    rollout_plan: |
+      Additive. Adds a documentation rule and two optional frontmatter fields.
+      No existing behavior changes; both conventions are advisory (report-only).
+
+  deprecation_plan:
+    deprecates: []
+    sunset_window: "n/a"
+
+  cost_benefit:
+    cost: "~1 documentation rule + two optional frontmatter fields"
+    benefit: |
+      Directly feeds the framework's own evolution-pipeline: produced_by gives the
+      provenance the pipeline needs (who/which model), and process_breakpoint surfaces
+      recurring failure points that become promotion candidates. Forensic, cheap, advisory.
+    risk: "low — additive; advisory by default; predicate is presence-of-frontmatter-field (100% checkable)"
+
+  approval:
+    eliel_decision: "PENDING"
+    eliel_decision_at: null
+    eliel_decision_rationale: ""
+    revision_request: []
+
+  implementation:
+    pr_url: ""
+    merged_at: ""
+    distributed_to: []