[v0.8 Epic] Agent-rated Evidence Readiness and Release Eligibility

[Stahl-G]

Summary

Implement v0.8 as Agent-rated Evidence Readiness and Release Eligibility.

This is not a claim that MABW can automatically prove truth or authorize publication. The v0.8 boundary is:

Semantic reviewer agents rate evidence and release risks using structured rubrics. Deterministic Python validators, policy packs, and control artifacts convert those ratings into warnings, blockers, allowed-use labels, and human-review requirements.

Why this matters

v0.7.x can already produce workflow traceability: material claims are linked to Claim Ledger entries and sources, and finalize produces reader-facing delivery artifacts. But recent real-use review exposed a gap:

• a run can be workflow/audit pass while still not being ready for formal research or external release;
• weak-but-present sources can pass traceability checks;
• mixed metric scopes can look cited but remain non-comparable;
• legal/policy/company-event claims can rely on media without official text;
• institutional branding/confidential labels can create release risk even when facts are sourced.

v0.8 should make these risks explicit, machine-readable, and blocking where the selected mode requires it.

Target control surfaces

Add four run-scoped artifacts:

output/intermediate/semantic_review_report.json
output/intermediate/evidence_report.json
output/intermediate/release_readiness_report.json
output/intermediate/human_approval_ledger.json

semantic_review_report.json

Agent-written. Records rubric-based semantic judgments: source authority, source-to-claim support, metric scope, official-source coverage, and branding/authorization risk.

evidence_report.json

Python-written. Converts semantic findings and Claim Ledger metadata into mode-aware evidence status, warnings, blockers, and verification paths.

release_readiness_report.json

Python-written. Aggregates workflow, audit, quality gates, evidence report, finalize report, approval record, branding/use-boundary, and package hygiene into release status.

human_approval_ledger.json

Python-written from explicit human/reviewer commands. Records required reviewer status for formal_release_candidate mode.

Sub-issues / execution sequence

1. [v0.8] Add mode registry and release_mode plumbing #90 — Add mode registry and release_mode plumbing.
2. [v0.8] Add Claim Evidence Schema v3 and semantic review contract #91 — Add Claim Evidence Schema v3 and semantic review contract.
3. [v0.8] Add semantic reviewer agent pack for evidence and release risk rating #92 — Add semantic reviewer agent pack for evidence and release risk rating.
4. [v0.8] Add evidence_report control surface and policy-pack blocker mapping #93 — Add evidence_report control surface and policy-pack blocker mapping.
5. [v0.8] Add release_readiness_report, branding gate, approval gate, and package hygiene checks #94 — Add release_readiness_report, branding gate, approval gate, and package hygiene checks.
6. [v0.8] Add human approval record for release-candidate gating #95 — Add human approval record for release-candidate gating.
7. [v0.8] Add synthetic regression cases for evidence/release blockers #96 — Add synthetic regression cases for evidence/release blockers.
8. [v0.8] Update docs, roadmap, and release notes for agent-rated evidence/release boundary #97 — Update docs, roadmap, and release notes for agent-rated evidence/release boundary.

Expected end-state example

A traceable but not release-ready run should be able to produce:

workflow_status: pass
semantic_review_status: completed
evidence_status: conditional
release_status: blocked
release_mode: research_review
allowed_use:
  - internal discussion draft
  - analyst review input
not_allowed_use:
  - formal external publication
  - institution-branded distribution
  - investment recommendation
  - legal or compliance conclusion
blocking_reasons:
  - unauthorized institutional branding
  - mixed metric scope
  - official legal/policy source missing
  - company latest-official check missing

Acceptance criteria

• release_mode is a first-class run concept.
• Semantic reviewer outputs are schema-validated and cannot mutate content artifacts.
• semantic_review_report.json exists and validates.
• evidence_report.json exists and validates.
• release_readiness_report.json exists and validates.
• human_approval_ledger.json exists and validates where required.
• Policy packs map semantic ratings to mode-specific blockers.
• Internal draft mode can allow conditional evidence without formal release.
• Formal release candidate mode requires stronger evidence and human approval.
• Synthetic fixtures prove workflow_status: pass can coexist with release_status: blocked.
• README and v0.8 release notes clearly state non-goals: no truth certification, no legal/investment advice, no automatic publication authorization.

Non-goals

• No automatic semantic truth proof.
• No automatic source recrawl.
• No live market-data fetching as part of release readiness.
• No runtime-specific artifact schema forks.
• No claim that output quality improved without evaluation evidence.
• No replacement of analyst, legal, compliance, or authorized publisher approval.

[Stahl-G]

Triage after v0.8.2: keep open, but do not force this entire epic into v0.8.3-v0.8.5.

The current public roadmap deliberately keeps v0.8 focused on measurement, fast-rerun, role topology, and evaluation. It also says no full mode-registry expansion before minimal recipe/pack surfaces are proven.

Recommended handling:

• v0.8.3 may borrow the minimal Claim Evidence Schema subset from [v0.8] Add Claim Evidence Schema v3 and semantic review contract #91 if it supports Claim Freeze without pulling in semantic review/release readiness.
• v0.8.5 may cover public-safe synthetic/evaluation fixtures where they support 080 measurement.
• Full mode registry, semantic reviewer pack, evidence report, release readiness, human approval ledger, and release docs ([v0.8] Add mode registry and release_mode plumbing #90-[v0.8] Update docs, roadmap, and release notes for agent-rated evidence/release boundary #97) should remain later work unless separately re-scoped.