fix: Quorum self-validation doc fixes — critic count, cost, Fixer history

Found and fixed via thorough Quorum self-validation of README, SPEC, CHANGELOG, SEC02:

README:
- Fix inconsistent thorough-depth cost ($2.00+ → $1.50+ in comparison table)

SPEC:
- §1: Update implemented critic count to 5 (Cross-Artifact Consistency is
  also implemented, not just the 4 main critics)
- §1: Correct planned count to 4 (Architecture, Delegation, Style, Tester)
- §2.5: Remove stale version annotation "(v0.5.1)" from depth table header
- §8/§10: Align remaining-critics count with updated §1

CHANGELOG:
- [0.5.0] roadmap: Correct Fixer as not-yet-done in that release; proposal
  mode shipped as NEW in [0.5.1], re-validation loops in [0.5.3]

Code (pre-existing changes, now committed):
- relationships.py: Remove unused base_dir param from load_manifest()
- pipeline.py: Update two call sites accordingly
- utils.py: Correct docstring (compact_match regex, no DOTALL)
- test_relationships.py: Update tests to match new load_manifest() signature

Tests: 796 passed, 0 failed.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: Akkari

README.md

@@ -54,7 +54,7 @@ You've got options. You could ask your agent to self-review. You could eyeball i
 | LLM spends tokens on obvious problems | **Multi-layer pre-screen catches issues first** — 10 deterministic checks + DevSkim + Ruff + Bandit + PSScriptAnalyzer — before LLM runs |
 | Reviews only one file at a time | **Batch validation** — run across a whole directory, or by `--pattern "*.md"`. One command, one verdict per file |
 | Each file judged in isolation | **Cross-artifact consistency** — I check whether your files actually agree with each other via a relationships manifest |
-| Same effort whether it's a quick sanity check or a full audit | I scale: **quick** ($0.15), **standard** ($0.50), **thorough** ($2.00+) |
+| Same effort whether it's a quick sanity check or a full audit | I scale: **quick** ($0.15), **standard** ($0.50), **thorough** ($1.50+) |
 | Each review starts from zero | I **learn patterns over time** — recurring findings auto-promote to mandatory checks via `known_issues.json` |
 
 Single agent, multi-agent swarm, hundred-step pipeline — doesn't matter how it was built. If it produced an output, I can tell you whether it holds up.

SPEC.md

@@ -9,7 +9,7 @@
 
 ## 1. Overview
 
-Quorum is a quality assurance framework with a nine-agent target architecture. Currently, 4 critics are implemented (Correctness, Completeness, Security, Code Hygiene), with 5 additional critics planned. The framework is designed to rigorously evaluate multi-agent systems, configurations, research, code, and operational procedures against domain-specific rubrics. It combines:
+Quorum is a quality assurance framework with a nine-agent target architecture. Currently, 5 critics are implemented (Correctness, Completeness, Security, Code Hygiene, and Cross-Artifact Consistency), with 4 additional critics planned. The framework is designed to rigorously evaluate multi-agent systems, configurations, research, code, and operational procedures against domain-specific rubrics. It combines:
 
 - **Parallel specialized critics** (9 agents with distinct expertise)
 - **Grounded evidence requirement** (every critique must cite tool-verified proof)
@@ -100,7 +100,7 @@ High-frequency patterns automatically promote to mandatory checks in future runs
 
 Three execution profiles balance rigor, speed, and cost:
 
-| Depth | Critics (v0.5.1) | Fix Loops | Runtime | Use Case |
+| Depth | Critics | Fix Loops | Runtime | Use Case |
 |-------|---------|-----------|---------|----------|
 | **quick** | Correctness, Completeness | 0 | 5-10 min | Fast feedback; low stakes |
 | **standard** | + Security, Code Hygiene | 0 | 15-30 min | Most work; default |
@@ -416,7 +416,7 @@ Status as of v0.5.0 (reference implementation):
 - [x] Fixer agent — proposal mode + re-validation loops (Phase 1.5; proposes and applies text replacements for CRITICAL/HIGH, then re-runs critics to verify)
 - [x] Python code rubric (25 criteria, PC-001–PC-025, auto-detects on .py files)
 - [x] Learning memory system (known_issues.json frequency tracking + mandatory check promotion — shipped in v0.5.3)
-- [ ] Remaining 5 critics (Architecture, Delegation, Style, Tester, full Fixer)
+- [ ] Remaining 4 critics — Architecture, Delegation, Style, Tester
 - [ ] Trust/monitoring system (per-critic accuracy tracking)
 
 See IMPLEMENTATION.md for a reference walkthrough.
@@ -441,7 +441,7 @@ Quorum is built on these peer-reviewed papers:
 
 ### Current Limitations (v3.0)
 
-- Only **4 of 9 critics are implemented** (Architecture, Delegation, Style, Tester are specified but not built)
+- Only **5 of 9 critics are implemented** (Architecture, Delegation, Style, Tester are specified but not built)
 - Rubric panel is **static** (doesn't specialize per artifact type dynamically)
 - **No critic-to-critic debate** (relies on Aggregator to resolve conflicts)
 - Learning is **frequency-based** only (no semantic deduplication of patterns yet)

docs/CHANGELOG.md

@@ -247,7 +247,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - [x] Parallel critic execution (Milestone #8) ← **DONE**
 - [x] Parallel batch validation (Milestone #8b) ← **DONE**
 - [x] Python code rubric — 25 criteria, auto-detect on .py ← **DONE**
-- [x] Fix loops / Fixer agent (Milestone #5b) ← **DONE** (proposals; re-validation loops future)
+- [ ] Fix loops / Fixer agent (Milestone #5b) ← deferred; proposal mode implemented as NEW in [0.5.1], re-validation loops in [0.5.3]
 - [ ] **Test suite + CI** (Milestone #13) — pytest smoke tests, import checks, GitHub Actions. Credibility gate: no tests = not production-ready. (Grok + GPT 5.4)
 - [ ] **Claim discipline audit** — GPT 5.4 scored 2.5/5 on both README and SPEC. Fix: README ("production-grade"), GitHub org/repo descriptions ("9 critics" + "learning memory"), model table ("not enough" without benchmarks), SPEC §1 framing (nine-agent → target architecture), research citations (Tomasev/ToolMaker = "engineering interpretation of," not "validated by")
 - [ ] **Pre-screen expansion** (Milestone #15) — Wire actual SAST tools: `ruff check` (S1xx+), `bandit`, `PSScriptAnalyzer`. Current 10 regex checks become universal fallback. Frameworks reference 80+ rules from research corpus — integrate them.

reference-implementation/quorum/pipeline.py

@@ -233,7 +233,7 @@ def _run_phase2(
 
     # Manifest paths are relative to the manifest's directory, not the target's
     manifest_base = relationships_path.parent.resolve()
-    relationships = load_manifest(relationships_path, base_dir=manifest_base)
+    relationships = load_manifest(relationships_path)
     resolved = resolve_relationships(relationships, base_dir=manifest_base)
 
     # Collect Phase 1 findings (NOT verdicts) as context
@@ -292,7 +292,7 @@ def _update_manifest(
     if cross_result_list and relationships_path is not None:
         try:
             from quorum.relationships import load_manifest
-            rels = load_manifest(relationships_path, base_dir=relationships_path.parent.resolve())
+            rels = load_manifest(relationships_path)
             manifest_data["relationships_count"] = len(rels)
         except Exception:
             pass

reference-implementation/quorum/relationships.py

@@ -59,13 +59,12 @@ class ResolvedRelationship(BaseModel):
     role_b_exists: bool = True
 
 
-def load_manifest(manifest_path: Path, base_dir: Path | None = None) -> list[Relationship]:
+def load_manifest(manifest_path: Path) -> list[Relationship]:
     """
     Load and validate a quorum-relationships.yaml manifest.
 
     Args:
         manifest_path: Path to the YAML manifest
-        base_dir: Base directory for resolving relative paths (defaults to manifest's parent)
 
     Returns:
         List of validated Relationship objects

reference-implementation/quorum/utils.py

@@ -15,8 +15,8 @@ def extract_json_from_response(response_text: str) -> str:
     Extract JSON content from a response that may be wrapped in markdown fences.
 
     This function handles common patterns where LLMs wrap JSON responses in
-    markdown code blocks, which can cause JSON parsing to fail. It uses a more
-    robust approach that handles nested backticks in JSON strings.
+    markdown code blocks, which can cause JSON parsing to fail. It does not
+    handle nested backticks in JSON string values.
 
     Args:
         response_text: Raw response text that may contain fenced JSON
@@ -41,7 +41,7 @@ def extract_json_from_response(response_text: str) -> str:
         return text
 
     # Handle compact fences (no newlines): ```json{"key": "value"}```
-    compact_match = re.match(r'```(?:json|JSON)?\s*(.+?)```$', text, re.DOTALL)
+    compact_match = re.match(r'```(?:json|JSON)?\s*(.+?)```$', text)
     if compact_match and '\n' not in text:
         return compact_match.group(1).strip()
 

reference-implementation/tests/test_relationships.py

@@ -23,12 +23,12 @@
 
 class TestLoadManifest:
     def test_load_valid_manifest(self):
-        rels = load_manifest(FIXTURES / "quorum-relationships.yaml", base_dir=FIXTURES)
+        rels = load_manifest(FIXTURES / "quorum-relationships.yaml")
         assert len(rels) == 2
         assert all(isinstance(r, Relationship) for r in rels)
 
     def test_relationship_types(self):
-        rels = load_manifest(FIXTURES / "quorum-relationships.yaml", base_dir=FIXTURES)
+        rels = load_manifest(FIXTURES / "quorum-relationships.yaml")
         types = {r.type for r in rels}
         assert "documents" in types
         assert "implements" in types
@@ -100,14 +100,14 @@ def test_invalid_type_rejected(self):
 
 class TestResolveRelationships:
     def test_resolve_existing_files(self):
-        rels = load_manifest(FIXTURES / "quorum-relationships.yaml", base_dir=FIXTURES)
+        rels = load_manifest(FIXTURES / "quorum-relationships.yaml")
         resolved = resolve_relationships(rels, base_dir=FIXTURES)
         assert len(resolved) == 2
         for r in resolved:
             assert isinstance(r, ResolvedRelationship)
 
     def test_existing_file_content_loaded(self):
-        rels = load_manifest(FIXTURES / "quorum-relationships.yaml", base_dir=FIXTURES)
+        rels = load_manifest(FIXTURES / "quorum-relationships.yaml")
         resolved = resolve_relationships(rels, base_dir=FIXTURES)
         # doc-a.md should be loaded
         docs_rel = next(r for r in resolved if r.relationship.type == "documents")