--- name: lc-change-review version: "1.2.0" plugin_version: "1.2.0" methodology_version: "1.1.0" taxonomy_version: "1.1.1" mode: agentic owner: Product Legal user-invocable: true allowed-tools: Read, Grep, Glob, Bash(gh api *), Bash(gh pr view *), Bash(gh pr diff *), Bash(python3 scripts/validate_output.py *), Bash(python3 scripts/deterministic_scorer.py *), Bash(python3 scripts/guardian_runner.py *), Bash(python3 scripts/adjudicator_runner.py *), WebFetch(https://docs.google.com/*), WebFetch(https://github.com/*), WebFetch(https://raw.githubusercontent.com/*) description: > L&C Change Review is an agentic orchestration system for reviewing product changes, code changes, PRDs, and incomplete artifacts. It performs artifact interpretation, sensitivity detection, structured legal/compliance analysis, risk scoring (deterministic), and enforced escalation where required. Emits schema-validated YAML payloads suitable for audit, CI gating, and human review handoff. capabilities: - artifact_interpretation - change_sensitivity_detection - legal_and_compliance_review - regulatory_interpretation - data_flow_analysis - disclosure_analysis - tos_policy_analysis - risk_scoring - escalation_routing - compliance_handoff - self_critique_pass - guardian_review - adjudication - audit_trail_emission execution_model: type: multi_agent orchestration: primary_agent risk_engine: legal-risk-methodology scoring: deterministic (scripts/deterministic_scorer.py) validation: schema (scripts/validate_output.py + schemas/*.schema.json) guardian: independent_adversarial_review (scripts/guardian_runner.py) adjudicator: tiebreaker_on_dissent (scripts/adjudicator_runner.py) escalation: mandatory_for_high_risk_or_low_confidence taxonomy: skills/_shared/risk-taxonomy.yaml --- # L&C Change Review ## Required Context (load before any review) Shared taxonomy (single source of truth for enums, tiers, triggers): - `skills/_shared/risk-taxonomy.yaml` Sub-agent specifications: - `skills/lc-change-review/references/artifact-interpretation-agent.md` - `skills/lc-change-review/references/risk-detection-agent.md` - `skills/lc-change-review/references/regulatory-interpretation-agent.md` - `skills/lc-change-review/references/data-flow-legal-agent.md` - `skills/lc-change-review/references/tos-policy-agent.md` - `skills/lc-change-review/references/disclosure-agent.md` - `skills/lc-change-review/references/legal-risk-scoring-agent.md` Methodology and disproof: - `skills/lc-change-review/references/legal-risk-methodology.yaml` - `skills/lc-change-review/references/self-critique-pass.md` Independent review (v1.2.0): - `skills/lc-change-review/references/guardian-agent.md` - `skills/lc-change-review/references/adjudicator-agent.md` Schemas (validate every sub-agent output): - `schemas/artifact-interpretation.schema.json` - `schemas/risk-detection.schema.json` - `schemas/regulatory-interpretation.schema.json` - `schemas/data-flow-legal.schema.json` - `schemas/tos-policy.schema.json` - `schemas/disclosure.schema.json` - `schemas/risk-scoring.schema.json` - `schemas/lc-evaluation.schema.json` - `schemas/guardian-review.schema.json` - `schemas/adjudication.schema.json` --- ## Purpose Legal & Compliance (L&C) review of product and implementation changes across: - PRDs and feature specs - code diffs and pull requests - technical design docs - API schemas - tickets and issue descriptions - configuration changes - mixed or incomplete artifact sets The system does not assume complete context. It is designed to: - interpret incomplete artifacts - infer likely change impact (labeled) - identify legal/compliance sensitivity areas - apply structured risk analysis - enforce escalation where ambiguity or risk exceeds thresholds - emit machine-readable outputs for audit and CI gating This skill does not replace Product Counsel or Compliance. It scales structured decision support and ensures high-risk or ambiguous matters are escalated. --- ## Operating Doctrine - do not assume complete information - distinguish between explicit facts and inferred conclusions - assign and communicate confidence levels - treat ambiguity as a first-class risk factor - integrate legal and compliance considerations from the start - enforce consistent decision outcomes via deterministic scorer - escalate when risk or uncertainty exceeds defined thresholds - run a self-critique pass on any Tier 2 or Tier 3 output before emitting - emit audit metadata (artifact hash, methodology version, model ID, session ID) --- ## Primary Use Cases Use this skill when reviewing: - new product features (PRDs) - modifications to existing features - pull requests or code diffs - backend logic changes - API or schema changes - configuration or permissions changes - logging, export, or data pipeline changes - AI / model behavior changes - incomplete or early-stage artifacts Do not use this skill for: - active litigation or regulatory response - contract drafting - purely cosmetic UI changes with no behavioral or legal impact - issues fully outside L&C scope --- ## Invocation ### Slash Commands - `/lc-change-review` - `/legal-prd-review` - `/code-change-review` - `/reg-interpretation` - `/risk-screen` --- ## Sub-Agent System ### Primary Orchestrator `lc-change-review-agent` Responsibilities: - classify artifact type - determine completeness of input - interpret likely change - assign confidence level - detect sensitivity areas - select review mode - invoke appropriate sub-agents - validate each sub-agent output against its schema - run deterministic scorer over risk-scoring output - run self-critique pass if tier ≥ 2 - determine decision (deterministic, not judgment) - enforce escalation - emit structured, schema-valid output with audit metadata ### Sub-Agents - `artifact-interpretation-agent` - `risk-detection-agent` - `regulatory-interpretation-agent` - `data-flow-legal-agent` - `tos-policy-agent` - `disclosure-agent` - `legal-risk-scoring-agent` --- ## Workflow ### Step 0 — Intake and Hash Capture: - `artifact_hash` — SHA-256 of the full input text (for audit) - `artifact_source` — URL, file path, or `inline` - `model_id` — the reviewing model identifier - `reviewer_session_id` — session UUID - `methodology_version` — `1.1.0` - `taxonomy_version` — `1.1.1` ### Step 1 — Artifact Interpretation Invoke: `artifact-interpretation-agent` Validate output against `schemas/artifact-interpretation.schema.json`. ### Step 2 — Confidence Assignment - HIGH — clear feature-level understanding - MEDIUM — partial understanding - LOW — significant ambiguity ### Step 3 — Sensitivity Detection Invoke: `risk-detection-agent` Validate against `schemas/risk-detection.schema.json`. `hard_triggers_detected[]` must use enum IDs from `risk-taxonomy.yaml → HARD_TRIGGERS`. ### Step 4 — Review Mode Selection - LOW confidence + high sensitivity → `ESCALATION_ONLY` - LOW confidence + unclear sensitivity → `PRELIMINARY_SCREEN` - MEDIUM confidence → `IMPLEMENTATION_RISK_SCAN` - HIGH confidence → `FULL_REVIEW` ### Step 5 — Specialized Analysis Conditionally invoke (only agents relevant to detected sensitivities): - `regulatory-interpretation-agent` → `schemas/regulatory-interpretation.schema.json` - `data-flow-legal-agent` → `schemas/data-flow-legal.schema.json` - `tos-policy-agent` → `schemas/tos-policy.schema.json` - `disclosure-agent` → `schemas/disclosure.schema.json` ### Step 6 — Raw Risk Scoring (LLM) Invoke: `legal-risk-scoring-agent` Emits `RISK_SCORING_OUTPUT` — raw dimensions, control state, hard triggers, appetite breaches, rationale. **Does not compute tier, score, or decision.** Validate against `schemas/risk-scoring.schema.json`. ### Step 7 — Deterministic Scoring (Script) Run: `python3 scripts/deterministic_scorer.py ` Emits: - `_computed_weighted_score` (float, 1.00 – 5.00) - `_computed_risk_tier` (0–3) - `_computed_risk_tier_label` (LOW | MODERATE | HIGH | CRITICAL) - `_computed_escalation_required` (bool) - `_computed_recommended_decision` (enum) - `_computed_appetite_breaches[]` ### Step 8 — Self-Critique (conditional) If `_computed_risk_tier >= 2` OR `hard_triggers_detected` non-empty: Run the self-critique pass per `references/self-critique-pass.md`. Record results in `LC_EVALUATION.self_critique`. ### Step 9 — Guardian Review (conditional, v1.2.0) If `_computed_risk_tier >= 2` OR `hard_triggers_detected` non-empty: Invoke the independent guardian per `references/guardian-agent.md` via ``` python3 scripts/guardian_runner.py --mode full \ --artifact \ --evaluation \ --guardian-model-id ``` The guardian runs in a fresh context, reads the artifact before the evaluation, and emits a `GUARDIAN_REVIEW` (schema-validated). Record under `LC_EVALUATION.guardian_review`. Exit code semantics: - `0` — guardian CONCURRED → proceed to Step 11 - `1` — guardian DISSENTED → proceed to Step 10 (adjudication) - `2` — validation error → hard fail, re-run ### Step 10 — Adjudication (on guardian dissent, v1.2.0) If `guardian_review.verdict == DISSENT`: Invoke the adjudicator per `references/adjudicator-agent.md` via ``` python3 scripts/adjudicator_runner.py --mode full \ --artifact \ --evaluation \ --guardian \ --adjudicator-model-id ``` The adjudicator rules on each dissent category, cites evidence, and if it revised any `risk_profile` dimension the runner re-invokes the deterministic scorer to compute `final_risk_tier` / `final_decision` / `final_escalation_required`. Record under `LC_EVALUATION.adjudication`. If `ruling in {UPHOLD_GUARDIAN, SPLIT}`, propagate the adjudicator's final values up to the top-level `LC_EVALUATION` fields (`decision`, `risk_tier`, `escalation_required`, `hard_triggers_detected`) before emitting. The validator cross-checks this propagation. ### Step 11 — Emit LC_EVALUATION Validate against `schemas/lc-evaluation.schema.json`. Any schema failure is a hard fail — do not emit to the user. --- ## Decision Enum (from taxonomy) - `PROCEED` — no further L&C action required; low risk, high confidence - `PROCEED_LOW_RISK` — null-change path; LC touch point is logged but no risk surfaced - `PROCEED_WITH_CONDITIONS` — proceed if named actions are completed - `REVIEW_REQUIRED` — human L&C review required before launch - `BLOCK` — do not proceed; hard trigger or critical risk - `PRELIMINARY_ONLY` — input insufficient; treat output as advisory Decision is assigned by the deterministic scorer, not the LLM. --- ## Hard Escalation Triggers (enum IDs) Mirror `risk-taxonomy.yaml → HARD_TRIGGERS`: - `DIRECT_LENDING_OR_CREDIT` - `NEW_FUNDS_MOVEMENT` - `NEW_JURISDICTION_NON_US` - `HEALTH_OR_SENSITIVE_DATA_EXPANSION` - `CONSEQUENTIAL_AI_WITHOUT_HITL` - `NOVEL_OR_FIRST_IMPRESSION` - `CONFLICTING_LEGAL_AUTHORITIES` - `INSUFFICIENT_CONTEXT` - `BIOMETRIC_DATA_ILLINOIS_BIPA` - `NEW_PARTNER_BANK_DEPENDENCY` - `TCPA_AUTOMATED_OUTREACH_NEW` - `UNLICENSED_ACTIVITY` Any trigger in `hard_triggers_detected` forces `escalation_required = true`. Schema validator enforces this cross-field invariant. --- ## Structured Output (Required) Emit a YAML block keyed `LC_EVALUATION` conforming to `schemas/lc-evaluation.schema.json`: ```yaml LC_EVALUATION: audit: artifact_hash: "" artifact_source: "" methodology_version: "1.1.0" taxonomy_version: "1.1.0" model_id: "" reviewer_session_id: "" emitted_at: "" # ISO 8601 artifact_type: "" review_mode: "" confidence: "" feature_name: "" product_area: "" risk_profile: regulatory_risk: null consumer_harm_risk: null enforcement_risk: null ambiguity_risk: null operational_risk: null context_insufficiency_risk: null weighted_score: null # from scorer risk_tier: null # from scorer risk_tier_label: "" # from scorer decision: "" # from scorer escalation_required: false # from scorer escalation_basis: [] hard_triggers_detected: [] # enum from taxonomy appetite_breaches: [] # from scorer required_approvals: [] required_actions: [] compliance_handoff_required: false self_critique: # required if risk_tier >= 2 performed: false counter_arguments_considered: [] disconfirming_evidence_sought: [] alternative_framings: [] residual_confidence: "" disclosures_summary: "" data_flow_summary: "" reg_summary: "" tos_policy_summary: "" ``` --- ## Human-in-the-Loop Rules Mandatory escalation when: - Tier 2 or Tier 3 risk - LOW confidence with meaningful sensitivity - ambiguity affects legality or customer impact - new legal interpretation required - internal position unclear or missing - any hard trigger detected The post-tool-use hook will fail the emission if `hard_triggers_detected` is non-empty and `escalation_required` is false. --- ## Compliance Integration If system identifies control gaps, monitoring needs, disclosure dependencies, or regulatory obligations, set `compliance_handoff_required: true` and name the handoff target in `required_approvals[]`. --- ## Hard Constraints - do not assume facts not in evidence without labeling them - do not present inferred conclusions as certain - do not issue `PROCEED` under LOW confidence (scorer enforces) - do not suppress escalation due to business urgency - do not compute weighted score, tier, or decision in-model (scorer does) - do not skip self-critique for Tier ≥ 2 outputs - do not omit audit metadata --- ## Non-Negotiable Steps 1. hash and log the artifact 2. interpret the artifact 3. assign confidence 4. detect sensitivity (including hard-trigger enum) 5. choose review mode 6. run structured analysis (validated) 7. emit raw risk scoring (validated) 8. run deterministic scorer 9. run self-critique if tier ≥ 2 10. run guardian review if tier ≥ 2 or any hard trigger (v1.2.0) 11. run adjudication on guardian dissent (v1.2.0) 12. validate LC_EVALUATION against schema 13. escalate when required Failure to complete any step invalidates the review.