Document Lab decision policy rules

README.ko.md

@@ -92,6 +92,7 @@ Guardrail:
 
 - `portfolio-demo-check`는 committed Studio fixture, README/PPT 수치, portfolio docs, local Studio asset을 검증합니다.
 - `core4-conformance-check`는 Forge manifest/metadata fixture, Runtime result JSON, Lab compare/deployment decision surface, AIGuard `guard_analysis` evidence를 기존 schema 변경 없이 검증합니다.
+- Lab decision surface는 `policy_version`, `triggered_rules`, `policy_summary`를 노출하며, reviewer는 [Lab Decision Policy](docs/policy/lab_decision_policy.md)에서 rule별 의미를 확인할 수 있습니다.
 
 추가 report path:
 

README.md

@@ -202,7 +202,7 @@ Guardrails:
 
 - `portfolio-demo-check` validates committed Studio fixtures, README/PPT metrics, portfolio docs, and local Studio assets without starting workers, queues, databases, or a production SaaS service.
 - `core4-conformance-check` validates bundled Forge manifest/metadata, Runtime result JSON, Lab compare bundle aliases/rendered outputs, Lab deployment decision surface, and AIGuard `guard_analysis` without mutating existing schemas.
-- The Lab decision surface exposes `policy_version`, `triggered_rules`, and `policy_summary` so reviewers can see which local policy rules produced deploy/review/block/unknown outcomes.
+- The Lab decision surface exposes `policy_version`, `triggered_rules`, and `policy_summary` so reviewers can see which local policy rules produced deploy/review/block/unknown outcomes; see [Lab Decision Policy](docs/policy/lab_decision_policy.md).
 
 Additional report paths:
 

docs/pipeline_contract.md

@@ -128,6 +128,9 @@ Deployment decisions include a backward-compatible policy trace:
 - `triggered_rules`: compact rule IDs explaining which Lab/Guard/contract conditions affected the decision.
 - `policy_summary`: reviewer-facing rule effects and descriptions.
 
+The reviewer-facing policy reference is maintained in
+[Lab Decision Policy](policy/lab_decision_policy.md).
+
 The v1 policy keeps existing decision behavior stable while making review/block reasoning explicit. Examples include Guard error evidence mapping to `blocked`, Guard warning evidence mapping to `review_required`, shape mismatch mapping to review, and latency/accuracy trade-off risk mapping to review.
 
 ## SaaS Boundary

docs/policy/lab_decision_policy.md

@@ -0,0 +1,69 @@
+# InferEdgeLab Decision Policy
+
+Policy version: `inferedge-lab-decision-policy-v1`
+
+This document is the reviewer-facing reference for the Lab-owned
+`deployment_decision` surface. It explains why a comparison is marked
+`deployable`, `deployable_with_note`, `review_required`, `blocked`, or
+`unknown` without changing the existing compare, Runtime result, or AIGuard
+contracts.
+
+## Scope
+
+InferEdgeLab owns the final deployment decision. Forge and Runtime provide
+provenance and execution evidence, and AIGuard provides optional deterministic
+diagnosis evidence. AIGuard, EdgeEnv, Orchestrator, and CI artifacts can add
+context, but they do not replace Lab policy ownership.
+
+## Decision Trace Fields
+
+Every Lab deployment decision includes:
+
+- `policy_version`: the policy identifier used to produce the decision.
+- `triggered_rules`: compact rule IDs explaining which evidence paths affected
+  the decision.
+- `policy_summary`: reviewer-facing rule effects and descriptions for each
+  triggered rule.
+
+These fields are additive policy trace metadata. They do not remove or rename
+existing deployment decision fields.
+
+## Rule Table
+
+| Rule | Effect | Description |
+|---|---|---|
+| `guard_error_block` | `blocked` | AIGuard reported error-level diagnosis evidence. |
+| `guard_warning_review` | `review_required` | AIGuard reported warning-level diagnosis evidence. |
+| `guard_skipped_unknown` | `unknown` | AIGuard was skipped, so diagnosis evidence is incomplete. |
+| `guard_unavailable_unknown` | `unknown` | AIGuard evidence is unavailable for this comparison. |
+| `guard_ok_lab_favorable_deployable` | `deployable` | Lab comparison is favorable and AIGuard passed. |
+| `guard_ok_lab_neutral_deployable_note` | `deployable_with_note` | Lab comparison is neutral and AIGuard passed. |
+| `guard_ok_lab_unfavorable_review` | `review_required` | Lab comparison indicates regression or mismatch despite AIGuard passing. |
+| `guard_ok_lab_unknown` | `unknown` | Lab comparison judgement is not recognized by the decision policy. |
+| `guard_status_unrecognized_unknown` | `unknown` | AIGuard status is not recognized by the decision policy. |
+| `shape_mismatch_review` | `review_required` | Input shape mismatch requires explicit deployment review. |
+| `system_mismatch_unfavorable_review` | `review_required` | System mismatch combined with unfavorable Lab judgement requires review. |
+| `system_mismatch_note` | `deployable_with_note` | System mismatch reduces confidence and must be noted in release evidence. |
+| `tradeoff_risk_review` | `review_required` | Latency/accuracy trade-off risk requires deployment review. |
+| `worker_uncompared_unknown` | `unknown` | Worker result has not been compared by Lab yet. |
+| `edgeenv_runtime_regression_review` | `review_required` | EdgeEnv same-condition runtime regression evidence requires review. |
+
+## Reviewer Interpretation
+
+- `blocked` means deployment should not proceed until the blocking evidence is
+  resolved.
+- `review_required` means a human reviewer should inspect the cited Lab,
+  AIGuard, EdgeEnv, or contract evidence before deployment.
+- `deployable_with_note` means deployment can proceed only with the noted
+  evidence retained in the release record.
+- `unknown` means Lab does not have enough compatible evidence to make a
+  confident deployment call.
+
+## Boundaries
+
+- This policy does not mutate `metadata.json`, `manifest.json`, Runtime
+  `result.json`, compare output, or AIGuard `guard_analysis`.
+- This policy does not make AIGuard, EdgeEnv, Orchestrator, or CI the final
+  deployment decision owner.
+- This policy is local-first validation evidence. It is not a production SaaS
+  approval workflow, cloud control plane, or model zoo automation policy.

tests/test_deployment_decision.py

@@ -1,5 +1,8 @@
 from __future__ import annotations
 
+from pathlib import Path
+
+from inferedgelab.services.deployment_decision import POLICY_RULES
 from inferedgelab.services.deployment_decision import POLICY_VERSION
 from inferedgelab.services.deployment_decision import build_deployment_decision
 
@@ -217,3 +220,31 @@ def test_diagnosis_guard_verdict_review_requires_lab_review():
     assert decision["guard_status"] == "warning"
     assert decision["guard_verdict"] == "review_required"
     assert_policy(decision, "guard_warning_review")
+
+
+def test_lab_decision_policy_reference_documents_current_rules():
+    root = Path(__file__).resolve().parents[1]
+    policy_doc = (root / "docs" / "policy" / "lab_decision_policy.md").read_text(
+        encoding="utf-8"
+    )
+    readme = (root / "README.md").read_text(encoding="utf-8")
+    readme_ko = (root / "README.ko.md").read_text(encoding="utf-8")
+    pipeline_contract = (root / "docs" / "pipeline_contract.md").read_text(
+        encoding="utf-8"
+    )
+
+    assert POLICY_VERSION in policy_doc
+    assert "Lab owns the final deployment decision" in policy_doc
+    assert "policy_version" in policy_doc
+    assert "triggered_rules" in policy_doc
+    assert "policy_summary" in policy_doc
+    assert "not mutate `metadata.json`, `manifest.json`, Runtime" in policy_doc
+    assert "not make AIGuard, EdgeEnv, Orchestrator, or CI" in policy_doc
+    assert "production SaaS" in policy_doc
+    for rule, metadata in POLICY_RULES.items():
+        assert f"`{rule}`" in policy_doc
+        assert f"| `{rule}` | `{metadata['effect']}` | {metadata['description']} |" in policy_doc
+
+    assert "docs/policy/lab_decision_policy.md" in readme
+    assert "docs/policy/lab_decision_policy.md" in readme_ko
+    assert "policy/lab_decision_policy.md" in pipeline_contract