akshay-na/editorial-cro-loop

Editorial CRO loop (editorial-cro)

Singleton execution-boundary critic: after the planning-episode owner (typically cco) has plan v0 — persisted or prompt-only — post–v0 edit round, and user signals execution intent, invoke once. Passes: editorial-cro pass 1 → owner patches v1 → pass 2 → owner patches v2 — before content-lead. See editorial-cro, agent-orchestration.

Purpose

• Pass 1: breadth + structural adversarial on v0 + bundle; full Adversarial dimensions; owner-only bounces.
• Pass 2: residual + freeze compliance + v1-regression scan; frozen IDs not re-litigated; explicit scan for new editorial/regression issues from v1 patches.
• At most one live loop per task_id; second concurrent = contract violation → owner rejects.

Inputs

| Field | Description | |-------|-------------| | Plan draft | Path + hash or prompt-only body (stable excerpt) | | specialist_bundle_refs[] | Specialist bundle refs from owner | | pass_number | 1 or 2 | | ledger_path | ~/ai-brain/session/cursor-<task-id>/critic-ledger.md | | frozen_finding_ids[] | Pass 2 only |

Outputs

• findings[] in subagent-response-protocol envelope.
• Ledger delta rows for owner to append.
• Metrics: editorial-cro.pass.1 | editorial-cro.pass.2.

Ledger schema

Table columns:

| Column | Description | |--------|-------------| | finding_id | Stable e.g. edcro-001 | | pass_raised | 1 or 2 | | category | claim-fact, pii-consent, channel-policy, brand-voice, metric-schema, registry-index, wikilink-graph, repurpose-lineage, coherence, ia, repo-alignment, rollback-process, … | | bounce_target | vp-editorial-architecture, cpo, staff-editor, … or - | | status | open | frozen_accepted | frozen_open_risk | degraded | | evidence_ref | vp-research ref, memory ref, repo path#line, or - if degraded | | resolved_in_plan_section | anchor or - |

Decision rubric

• Domain gap → bounce_target set; owner Tasks specialist; no lateral Task from editorial-cro.
• Coherence / phases → self-resolve; external facts → vp-research; corpus facts → cite repo path, schema/registry lines, or memory/KB.
• No atlassian-pm in content org.

Adversarial dimensions (editorial checklist)

• Claims / PII / consent — sourcing, privacy copy, regulated claims.
• Channel policy / fit — limits, disclosures, platform-sensitive patterns.
• Brand / voice drift vs playbook + corpus norms (evidence).
• Metric / schema / registry / graph — metrics/, _schema, 00-INDEX, wikilinks.
• Repurposing / lineage — parent_id, recycle, cross-channel consistency.
• Coherence / editorial DAG — draft/staging/publish ordering; content vs asset dependencies.
• Rollback / human process — promotion gates, undo semantics, experiment observability.

Pass differentiation

| Pass | Focus | | ---- | ----- | | 1 | Full rubric on v0; editorial bounces within budget. | | 2 | Frozen IDs respected; residual non-frozen; v1 vs v0 regression (new claims, channel scope, index/registry gaps from patches). |

Conditional minimum bar (findings[])

Triggers (any): multi-phase plan; or staging/publish promotion; or metrics/schema/registry/graph edits; or cross-channel repurpose; or high-liability/regulated claims.

When triggered, each pass SHOULD address themes: claims/privacy, channel+voice, corpus automation (schema/registry/index), lineage — combine or mark N/A with reason.

Trivial single-step low-risk draft: no minimum count; stay proportional.

Caps

| Limit | Value | |-------|--------| | Loop instances | 1 / planning episode | | Wall | 420s | | vp-research | ≤ 3 / loop | | Bounces | ≤ 2 / pass (owner) | | Passes | 2 |

Observability

agent-observability: stages editorial-cro.pass.1, editorial-cro.pass.2.

Cross-links

• subagent-response-protocol
• cco