All notable changes to Vibe Science are documented here.
',i)+3;console.log(m.computeBoundaryHash(t.slice(end)))})"and update the hash= value below in a reviewed commit. -->
Trigger: v7 implementation from the TRACE spec, followed by repeated adversarial review and paranoid debugging until convergence.
- automatic ingestion of claim lifecycle events into
claim_events - automatic ingestion of serendipity seeds with
source_claim_id - automatic ingestion of Reviewer 2 artifacts into
r2_reviews - citation extraction for DOI / PMID / arXiv with bounded sync verification
- plugin-level source validity gates:
L0blocksUNRESOLVED/RETRACTED,D1blocks anything notVERIFIED - FTS5/BM25 retrieval closure through
memory_fts, curated indexing, and tiered fallback
- schema migration chain expanded to
schema_version = 4 - benchmark runner now writes artifacts and records
benchmark_runs - smoke test and TRACE readiness gate with A/B comparison
- strict integrity persistence via
INTEGRITY_OK/INTEGRITY_DEGRADED
- project version bumped to
7.0.0 - plugin manifests, setup output, skill metadata, templates, and active docs aligned to
TRACE - README and ARCHITECTURE now describe TRACE as the current release instead of NEXUS
104/104tests green in__test_e2e.mjsnode evals/smoke-trace.mjsgreennode scripts/v7-readiness.mjsgreen
[6.0.49] — 2026-03-21 — Baseline stabilization: test fixes, SSOT alignment, version label correction (Adversarial Review R1)
Trigger: Adversarial review Round 1 — full SSOT matrix built by cross-referencing README.md, ARCHITECTURE.md, CLAUDE.md, schema.sql, __test_e2e.mjs, package.json, benchmark-reporter.js, and CHANGELOG.md. Found 2 test failures, 1 table count drift, and 2 version label mismatches.
- __test_e2e.mjs line 599: Removed
onnxruntime-nodefromrequiredDepsarray - WHY:
onnxruntime-nodeis NOT a direct dependency in package.json — it is pulled in transitively by@huggingface/transformers. The B5 test was checking for its presence in package.json and failing. B7 already tests the import with graceful fallback. The test was wrong, not the code.
- __test_e2e.mjs line 743: Added
'CHANGELOG.md'toAUTHOR_FILESexclusion set - WHY: The forbidden-names scan caught
th3vib3coderin CHANGELOG.md (repo URLs in historical entries). CHANGELOG.md is append-only historical documentation that legitimately references the repository URL. It was already excluded for README.md, CITATION.cff, LICENSE, and NOTICE — CHANGELOG.md was the only historical document missing from the exclusion list.
- README.md: 5 occurrences updated (lines 51, 129, 131, 220, 326, 395)
- ARCHITECTURE.md: 2 occurrences updated (lines 41, 101–116, 152) +
benchmark_runsadded to table listing - plugin/db/schema.sql: Comment updated (line 212): "12 tables" → "13 tables"
- WHY: The
benchmark_runstable was added to schema.sql and __test_e2e.mjs (which expects 13 tables and passes), but the documentation still said "12 tables". The 13th table isbenchmark_runsfor eval/benchmark result tracking.memory_embeddingsis still NOT counted (optional fallback, documented as such).
- plugin/lib/benchmark-reporter.js line 5:
"v6.1 NEXUS"→"v6.0 NEXUS" - plugin/db/schema.sql line 258:
"v6.1: Benchmark & Eval Tracking"→"v6.0: Benchmark & Eval Tracking" - WHY: The project version is 6.0.0 (package.json, plugin.json, README badge). These two files had "v6.1" in their header comments, creating the false impression that a v6.1 release exists. The benchmark infrastructure shipped as part of v6.0.x patches.
- CHANGELOG.md historical entries referencing "12 tables" (R52, R53) are left as-is — they were correct at the time of writing, before benchmark_runs was added.
- 53/53 pass, 0 fail (was 51 pass, 2 fail before)
Trigger: Round 68 paranoid deep debug — 6 parallel agents audited hooks.json+settings.json, all 12 schemas, root docs (CLAUDE.md, ARCHITECTURE.md), pre-compact.js, skills/vibe/, .claude/. Found 3 bugs in CLAUDE.md + gates.md; 6 pre-compact.js findings triaged as design choices.
- CLAUDE.md LAW 7 (line 31): Removed stale
TREE-STATE.jsonreference — file doesn't exist, v6.0 uses SQLite DB for tree state. Now reads "resumable from STATE.md alone (database enriches but is not required)." - CLAUDE.md FILE STRUCTURE (lines 46-56):
schemas/,protocols/,assets/were listed under.vibe-science/but they live at project root. Split section into "Runtime state" (.vibe-science/) and "Static definitions" (project root). Removed non-existentTREE-STATE.jsonentry. - gates/gates.md:269:
(exists in TREE-STATE.json)→(exists in tree state)— file doesn't exist. - gates/gates.md:570:
(RQ.md, TREE-STATE.json planned branches)→(RQ.md, planned tree branches)— same stale reference.
- pre-compact.js:78:
R2_REVIEWEDin terminal state exclusion — INTENTIONAL (same pattern as stop.js:150, identifies claims needing R2 attention). - pre-compact.js:79-83: Cross-session claim resolution query — INTENTIONAL (compaction needs project-wide claim state, unlike stop.js which checks current session only for LAW 4).
- pre-compact.js:72: DISTINCT without ORDER BY — NOT A BUG (each claim_id has one CREATED event).
- pre-compact.js:120-124: Seed snapshot omits
discriminating_test— BY DESIGN (summary only; full data remains in DB). - ARCHITECTURE.md R3 dimensions: Not named — STYLE CHOICE (count "6 dimensions" is correct; judge-rubric.yaml is source of truth).
Trigger: Round 67 paranoid deep debug — 6 parallel agents audited plugin/lib/ (8 files), ARCHITECTURE.md, examples/, codebase-wide stale action type sweep, blueprint-vs-schema.sql comparison, worker-embed.js. Found 1 bug; stale action types fully purged.
- ARCHITECTURE.md:62: "20+ action types classified" → "16 action types classified" to match spine-entry.schema.json enum (exactly 16 values).
- plugin/lib/ (all 8 files): No bugs, no stale values, no unused exports.
- Stale action type sweep: FEATURE_EXTRACTION, R2_REVIEW, WEB_SEARCH, FILE_WRITE, CODE_WRITE, DATA_INSPECT, DOCUMENTATION — zero remaining occurrences outside archive/ and CHANGELOG.md.
- Blueprint vs schema.sql: Column names, types, defaults, FKs, indexes all match. IF NOT EXISTS difference is intentional (idempotent schema.sql). vec_memories commented-out is documented design.
- worker-embed.js: Column references match schema.sql, model name correct, error handling sound.
Trigger: Round 66 paranoid deep debug — 6 parallel agents audited session-start.js, prompt-submit.js, pre-tool-use.js, stop.js+subagent-stop.js, post-tool-use.js (full), all 21 protocols. Found 4 bugs.
- plugin/db/schema.sql:25: Comment listed
FEATURE_EXTRACTION→EXTRACTto match spine-entry.schema.json. - plugin/db/schema.sql:72:
j0_dimensionscomment listed only 2 of 6 R3 dimensions with ellipsis (same bug fixed in blueprint R62, missed here). Now lists all 6. - plugin/lib/db.js:219: JSDoc
FEATURE_EXTRACTION→EXTRACT. - plugin/scripts/post-tool-use.js:1743: Removed dead
CALIBRATEfrom phaseExpectedActions CALIBRATION phase —classifyAction()never returns it. - plugin/scripts/pre-tool-use.js:54: Tightened LAW 9 NOT_APPLICABLE regex from
/confounder.*not.?applicable/(greedy.*could match across unrelated text) to/confounder\S*\s*[:=]?\s*not.?applicable/(requires field-like context).
- prompt-submit.js:66:
reviewer2vs canonicalR2— DESIGN CHOICE. Internal string consistentlyreviewer2across permission-engine, prompt-submit, stop hooks. Documentation saysR2as shorthand. - subagent-stop.js:65-69: Salvagente query uses LIKE instead of FK — DESIGN GAP (no
source_claim_idcolumn in DB table). Tracked but not a bug per current schema.
Trigger: Round 65 paranoid deep debug — 6 parallel agents audited commands/, references/, .claude-plugin/, package.json, action_type cross-ref, README line counts. Found 4 real bugs (1 critical).
- CRITICAL: plugin/scripts/post-tool-use.js:1735-1744:
phaseExpectedActionsused 9 invalid action type names (SEARCH, FILE_READ, DATA_INSPECT, FEATURE_EXTRACTION, CODE_WRITE, EVALUATION, VISUALIZATION, DOCUMENTATION, CONFIGURATION) thatclassifyAction()never returns. Drift detection was silently broken — always reporting drift for valid actions. Rewrote to use only spine-entry.schema.json enum values. - plugin/scripts/post-tool-use.js:1073-1089: JSDoc listed old blueprint vocabulary instead of schema enum. Replaced with accurate 16-type enum documentation.
- skills/vibe/references/research-spine.md:32-45: Action type table listed 12 of 16 types. Added 4 missing: CALIBRATION, CONFORMAL_PREDICT, TOOL_USE, COMPACT_SNAPSHOT. Also fixed field names
inputs→input,outputs→outputto match schema (lines 22-23), and stale exampleWEB_SEARCH→LITERATURE_SEARCHin auto-logging section (line 151). - package-lock.json:10: Root package license said "MIT" while package.json says "Apache-2.0". Synced to Apache-2.0.
Trigger: Round 64 paranoid deep debug — 6 parallel agents audited hooks/, schemas/, plugin/lib/, examples/, JSON validity, stale string sweep. Found 3 real bugs (1 stale threshold, 2 schema-vs-doc action type mismatches).
- skills/vibe/SKILL.md:514: "batch review at 3" → "batch review at 5" (last remaining stale BATCH threshold)
- protocols/loop-otae.md:341: Spine action_type list used
FEATURE_EXTRACTION/R2_REVIEW(don't exist in schema). Replaced with full 16-type enum from spine-entry.schema.json: EXTRACT, REVIEW, + 4 missing types (CALIBRATE, CONFORMAL_PREDICT, TOOL_USE, COMPACT_SNAPSHOT). - blueprints/v6.0-NEXUS-BLUEPRINT.md:906: DB schema comment listed
FEATURE_EXTRACTION→EXTRACTto match schema.
- spine-entry.schema.json: CALIBRATE + CALIBRATION dual enum values — already documented in [6.0.24] as legacy/unused, not harmful.
- serendipity-seed.schema.json: source_claim_id IS required for SALVAGED_FROM_R2 (line 138). Agent misread. FALSE POSITIVE.
- vigilance-check.schema.json: Escalation logic at orchestrator level, not schema level — DESIGN CHOICE.
Trigger: Round 62 paranoid deep debug — 6 parallel agents audited references/, blueprints/, scripts/agents, root SKILL.md 700-1369, plugin/db, line counts. Found 2 real bugs.
- blueprints/v6.0-NEXUS-BLUEPRINT.md:953: DB schema comment listed only 2 of 6 R3 dimensions with ellipsis. Now lists all 6: specificity, counter_evidence_search, confounder_analysis, falsification_demand, independence, escalation.
- plugin/db/literature-registry.json:124,160: pubmed and pmc entries had undefined category
"biomedical"(not in canonical 12). Changed to"medicine_clinical".
[6.0.42] — 2026-03-05 — 5 stale BATCH/terminology fixes across protocols, examples, SKILL.md, UPGRADE_PLAN (R61)
Trigger: Round 61 paranoid deep debug — 6 parallel agents audited 21 protocols, plugin/lib JS, examples, cross-file enums, config metadata, root SKILL.md. Found 5 real bugs across 4 files.
- protocols/reviewer2-ensemble.md:58: "accumulates 3 minor findings" → "accumulates 5 unreviewed claims"
- examples/walkthrough-literature-review.md:66: "minor claims" → "unreviewed claims, BATCH threshold is 5"
- examples/walkthrough-literature-review.md:93-100: Off-by-one BATCH trigger (4 claims triggered >= 5 gate). Added C-005, count now 5.
- UPGRADE_PLAN_V2.md:207:
minor_findings_pending→unreviewed_claims_pending - SKILL.md:1050: "batch review at 3" → "batch review at 5"
Trigger: Round 60 paranoid deep debug — 6 parallel agents audited hook scripts, 12 JSON schemas, commands, .claude/ config, README.md, ARCHITECTURE.md. Found 3 real bugs, 3 false positives.
- plugin/scripts/post-tool-use.js:873: Salvagente comment "Blueprint v5.0" → "Blueprint v6.0"
- README.md:296: SKILL.md line count "528" → "527"
- README.md:341,371: Root SKILL.md line count "1,363" → "1,369"
- claim-promotion.schema.json:42 "v5.0 lock": Provenance tag (lock introduced in v5.0, like "v5.0-geometric" for formula)
- ARCHITECTURE.md:339 v4.0 gate count "26": Historical snapshot of v3.5→v4.0 transition, not a current-state claim
- R2 verdict vocabulary (PASS/FAIL vs ACCEPT/REJECT across schemas): Intentional — different gate stages use different verdict vocabularies
Trigger: Round 59 paranoid deep debug — 6 parallel agents audited R3 dimension completeness (clean), plugin lib modules (clean), file counts (clean), rarely-audited protocols (clean), gates.md stale values, SKILL.md versioning.
- gates/gates.md:458: "all 25 existing gates" → "the 25 pre-v5.5 gates" (was stale v5.0 count in v5.5 DQ section intro)
- gates/gates.md:671: Schema Enforcement Summary header "(v5.0)" → "(v6.0)" (section covers v5.0+v5.5 gates)
- skills/vibe/SKILL.md:6
v6.0.1: Intentional patch version from v6.0.1 best-practices upgrade (has CHANGELOG entry [6.0.1], archive blueprint, consistent across plugin+archive copies)
Trigger: Round 58 paranoid deep debug — 6 parallel agents audited R3 dimension completeness, CITATION/README, examples, YAML configs, assets templates, and internal links (84 verified). Found 1 remaining stale R3 dimension name and root templates.md significantly behind plugin version.
- skills/vibe/references/reviewer2-ensemble.md:204: "Constructiveness" → "Escalation" (matches judge-rubric.yaml)
- TREE-STATE version:
"4.0.0"→"6.0.0" - TREE-STATE: Added missing
"cycle": 1field - STATE.md template: Added v6.0 gate fields (dq_gates, new_gates: DD0/DC0/L-1, observer_alerts, spine_entries)
- tree_health: Added missing
exploration_ratio: 0.0(LAW 8) - CLAIM-LEDGER template: Added missing
R2 INLINEfield (v5.5+) - Added 3 v6.0 templates: Handoff Document, Instinct Entry, Pattern Report
claude-code.yaml:18—claude-haiku-4-5is a comment showing model family name, not the deployment IDassets/templates.mdqueries_deduped field — present in root (full reference), absent in plugin (minimal) — design choicefault-taxonomy.yamlstructural split — root has domain-specific faults, plugin has meta-faults only — intentional
Trigger: Round 57 paranoid deep debug — 6 parallel agents audited JSON schemas, hook scripts, protocols, top-level docs, config/settings, and skills/vibe/references. Found R3 Judge dimension names in 4 docs didn't match the actual rubric YAML.
- Source of truth:
assets/judge-rubric.yaml+protocols/judge-agent.mddefine 6 dimensions: Specificity, Counter-Evidence Search, Confounder Analysis, Falsification Demand, Independence, Escalation - 4 files had wrong names ("Depth, Constructiveness, Consistency" instead of "Confounder Analysis, Falsification Demand, Escalation"):
.claude/rules/roles.md:49SKILL.md:251skills/vibe/references/constitution.md:118skills/vibe/AGENTS.md:152-153
schemas/claim-promotion.schema.json:179—v5.0-geometricis the formula's own version identifier (introduced in v5.0 IUDEX), not the project version. Formula unchanged in v6.0.CLAUDE.md.vibe-science/paths — these describe the structure created in user projects, not this repo's layoutskills/vibe/references/brainstorm-engine.md:253"0-15 scale" — correctly refers to brainstorm hypothesis scoring (5 dims × 0-3), not serendipity (0-20)
Trigger: Round 56 paranoid deep debug — 6 parallel audit agents covered commands/, examples/, YAML frontmatter, table cross-refs, JS hooks, and protocol gate definitions. Found 4 remaining stale
minor findingsreferences in command docs.
- commands/loop.md line 48:
Minor findings pending: 2→unreviewed_claims_pending: 2 - commands/loop.md line 97:
Minor findings pending: 3/3→unreviewed_claims_pending: 5/5 - commands/loop.md line 159:
Minor findings pending: 0→unreviewed_claims_pending: 0 - commands/reviewer2.md line 26:
minor findings→unreviewed claims
- ARCHITECTURE.md TOC anchor links: GFM slug generation strips em-dashes correctly, links valid
- plugin/scripts/subagent-stop.js SQL LIKE: claim IDs are system-generated
C-NNN, no wildcard risk
[6.0.36] — 2026-03-05 — Rename minor_findings_pending → unreviewed_claims_pending across 6 files (R55)
Trigger: Round 55 paranoid deep debug — 6 parallel audit agents covered internal links, examples, README/config, AGENTS.md, CHANGELOG, and JS hook logic. Found the variable name
minor_findings_pendingwas never renamed tounreviewed_claims_pendingdespite R53 changing the terminology everywhere else.
- examples/walkthrough-literature-review.md lines 97, 100: Variable name in prose
- assets/templates.md line 20: STATE.md template YAML key
- skills/vibe/assets/templates.md line 20: Condensed STATE.md template YAML key
- protocols/loop-otae.md line 40: OBSERVE phase reference
- skills/vibe/references/loop-otae.md line 40: Condensed OBSERVE phase reference
- WHY: R53-R54 changed the BATCH trigger description from "3 minor findings" to "5 unreviewed claims" but never renamed the underlying STATE.md frontmatter key or the protocol references to it. JS code (stop.js) already uses
unreviewed_claims. This completes the rename.
- ARCHITECTURE.md TOC anchors: All 7 links are CORRECT — GFM strips em-dashes from headings, so
## v6.0 — NEXUS→ slug#v60--nexus - agent-teams.md Researcher model "Sonnet": DESIGN CHOICE for TEAM mode cost optimization. AGENTS.md (Opus) is the canonical definition; TEAM roster intentionally suggests Sonnet as budget alternative
- subagent-stop.js LIKE wildcards: LOW RISK — claim IDs are system-generated (
C-001format), will never contain%or_ - CHANGELOG 6.0.31 references [6.0.24]: Historical append-only doc, attribution error doesn't affect system behavior
Trigger: Round 54 paranoid deep debug — audited commands/, schemas/, all reference cards, and remaining protocols. Found 3 real bugs + 2 false positives (brainstorm 0-15 scale is correct for hypothesis scoring; plugin/schemas/ doesn't exist because hooks don't use schemas).
minor_findings_pending >= 3→unreviewed_claims_pending >= 5- WHY: R53 fixed lines 230/237 in this file but missed the R2 trigger table at line 282
minor_findings_pending >= 3→minor_findings_pending >= 5- WHY: Example walkthrough still showed old threshold
- Replaced additive formula
E×0.30 + R×0.25 + C×0.20 + K×0.15 + D×0.10with v5.0 IUDEX formulaE * D * (R_eff * C_eff * K_eff)^(1/3) - WHY: v5.0 replaced the additive formula with a hybrid (hard veto on E,D + geometric mean on R,C,K). ARCHITECTURE.md still showed the obsolete v4.5 version
- brainstorm-engine.md "0-15 scale": CORRECT — brainstorm hypothesis scoring uses 5 dimensions × 0-3 = max 15 points. This is NOT the serendipity scale (0-20). Agent initially flagged this; investigated and confirmed correct
- plugin/schemas/ empty: NOT A BUG — plugin/schemas/ doesn't exist because plugin JS hooks don't reference schemas at all. Schemas live in schemas/ and skills/vibe/assets/schemas/ only
Trigger: Round 53 paranoid deep debug — audited ALL plugin JS code (17 files), cross-referenced ALL gate definitions repo-wide, verified CLAUDE.md F: vs D: consistency, and deep-audited skills/vibe/ directory. R52 fixed the BATCH threshold in protocols/reviewer2-ensemble.md but missed 6 occurrences in 4 other active files.
- skills/vibe/SKILL.md line 308:
"3 minor findings accumulated"→"5 unreviewed claims accumulated"(condensed skill mode table) - skills/vibe/references/loop-otae.md line 265:
"minor_findings_pending >= 3"→"unreviewed_claims_pending >= 5"(condensed loop reference R2 trigger table) - SKILL.md line 601:
"BATCH: 3 minor findings accumulated"→"BATCH: 5 unreviewed claims accumulated"(ASCII art diagram) - SKILL.md line 679:
"3 minor findings accumulated"→"5 unreviewed claims accumulated"(R2 mode table) - commands/loop.md line 177:
"3 minor findings"→"5 unreviewed claims"(loop R2 trigger table) - commands/reviewer2.md line 16:
"3 minor findings accumulated"→"5 unreviewed claims accumulated"(reviewer2 auto-invocation list) - WHY: R52 correctly established >= 5 as the canonical BATCH threshold (matching the formal mode table in protocols/reviewer2-ensemble.md line 33 and the reference card line 55), but only fixed 2 of 8 total stale references. These 6 occurrences in the root SKILL.md, condensed skill, commands, and condensed references were missed because the R52 audit focused on protocols/ only. The
protocols/reviewer2-ensemble.mdline 58 ("v5.0 BATCH mode accumulates 3 minor findings") is correctly left as-is — it's historical context describing old v5.0 behavior.
- Plugin JS code (17 files): All canonical values correct. Gate IDs, serendipity thresholds (10/15), temporal decay (-0.02/week), DB schema (12 tables), hook count (7), version (6.0.0) — all match
- KNOWN_ROLES array: Has 6 entries — CORRECT (6 permission roles ≠ 7 agent types; these are separate classifications per CLAUDE.md line 41)
- Gate cross-reference (entire repo): All 32 gates correctly defined. Zero phantom gates (G7, S0, DQ0, L3, L-2). Zero stale "34" or "27" remaining
- CLAUDE.md: F: and D: byte-identical (8118 bytes, 95 lines). All canonical values correct
- skills/vibe/ package: All canonical values correct except BATCH threshold (now fixed)
Trigger: Round 52 paranoid deep debug — deep-audited root SKILL.md (~1,370 lines), condensed SKILL.md (~528 lines), 36 reference cards, and 21 protocols. Found 6 bugs + 2 gaps across 5 files.
- SKILL.md line 329:
"All 27 v5.0 gates"→"All 25 v5.0 gates" - WHY: This was the ROOT CAUSE of the persistent "34 gates" error seen in earlier rounds. The math: 27+7=34 (wrong) vs 25+7=32 (correct). v5.0 had 25 gates (G0-G6=7, L0-L2=3, D0-D2=3, T0-T3=4, B0=1, S1-S5=5, V0=1, J0=1 = 25). The stale "27" propagated through any reader who computed total gates from this line.
- SKILL.md line 754:
"scored 13/15"→"scored 13/15 (v4.0 scale; current scale is 0-20)" - WHY: The CRISPR case study predates v5.0 when the serendipity scale was changed from 0-15 to 0-20. The "13/15" is historically accurate but confusing without annotation — readers would think the current scale is /15 when it's actually 0-20 (7-component scoring).
Fixed — protocols/agent-teams.md Line 129: Serendipity Alert "score >= 8" Should Be ">= 10" (MEDIUM)
- protocols/agent-teams.md line 129:
"score >= 8"→"score >= 10" - WHY: Score 8 falls in the FILE band (5-9), not the QUEUE band (10-14). The canonical thresholds are: NOISE (0-4), FILE (5-9), QUEUE (10-14), INTERRUPT (15-20). Creating an alert file at score 8 is wrong — alerts correspond to QUEUE threshold (>= 10). This value was likely never updated when the serendipity scale was recalibrated from 0-15 to 0-20.
- Lines 110, 133:
"3 minor findings accumulated"→"5 unreviewed claims accumulated" - WHY: The formal BATCH mode table (line 33) defines the trigger as ">= 5 unreviewed claims". The reference card also says ">= 5". But the explanatory text at line 110 and the "When to Invoke" table at line 133 still said "3" — the old v5.0 threshold before INLINE mode reduced BATCH frequency. Line 58 stays as-is because it explicitly describes "v5.0 BATCH mode" (historical context). The authoritative source is the formal mode table.
- ARCHITECTURE.md line 430: Updated "Quality Gates (12)" section with all 32 current gates
- WHY: This section was frozen at the v3.5 gate count (12), listing only G0-G5, L0-L2, D0-D2. Current system has 32 gates across 11 categories: G0-G6, L-1+L0-L2, D0-D2, T0-T3, B0, S1-S5, DQ1-DQ4, DD0, DC0, V0, J0. Any reader of ARCHITECTURE.md would get a completely wrong picture of the gate system.
- Line 230: Added
L-1: Literature pre-checkbefore L0 in Literature Gates section - Line 237: Added
T3: Tree healthafter T2 in Tree Gates section - WHY: The gate listing in the EVALUATE phase was incomplete. L-1 (Literature Pre-Check, added in v5.5) and T3 (Tree Health Check, added in v4.0) were defined elsewhere in the document but missing from the gate summary sections. An agent reading only this section would not know these gates exist.
- SKILL.md line 175-176: Added: confidence range (0.3-0.9), temporal decay formula (exp(-0.02 × weeks), half-life ~34.7 weeks), lifecycle stages (nascent→developing→established→proven), archival threshold (< 0.2)
- WHY: LAW 12 INSTINCT was described qualitatively but lacked the canonical parameters defined in CLAUDE.md and the instinct scanner implementation. Without these values, the instinct system has no numeric boundaries.
- SKILL.md line 12: Added v6.0.0 NEXUS changelog entry
- WHY: Frontmatter changelog listed versions v4.0.0 through v5.5.0 but was missing v6.0.0 despite
version: "6.0.0"being set in line 6. Any tool reading the changelog field would not know what v6.0.0 changed.
- Condensed SKILL.md: Zero stale values, zero domain contamination across ~528 lines
- Reference cards: 35/36 clean. 1 design-choice flag (brainstorm-engine.md collision threshold >= 8 vs canonical queue >= 10 — intentionally lower bar during brainstorming)
- Protocols: 14/21 fully clean. Issues fixed in this commit cover all actionable findings
Trigger: Round 51 paranoid deep debug — audited all internal markdown links across 88 .md files, cross-checked CLAUDE.md consistency between F: and D: drives, validated all 12 JSON schema files (3 copies each), and verified hooks configuration parity between settings.json and hooks.json.
- CLAUDE.md line 41:
`AGENTS.md`→`skills/vibe/AGENTS.md` - WHY: CLAUDE.md references
AGENTS.mdwithout a path prefix, implying it exists at the repo root. But noAGENTS.mdexists at root — the active copy lives atskills/vibe/AGENTS.md(with archived versions inarchive/). An agent reading CLAUDE.md and attempting to readAGENTS.mdfrom the repo root would get a file-not-found error. The reference needs the full relative path.
- Internal markdown links: All 13 links across 88 .md files resolve to existing targets. Zero broken links
- JSON Schemas (12 files): All valid JSON, all
$idmatch filenames, all$refresolve, all enum values consistent across 3 copies (root schemas/, skills/vibe/assets/schemas/, archive/). Informational:CALIBRATEis a legacy enum value in spine-entry.schema.json never emitted by JS code (onlyCALIBRATIONis used) — not harmful, just unused - Hooks configuration: Both
.claude/settings.jsonandhooks/hooks.jsondefine identical 7 hooks with matching timeouts, matchers, and script references. All 7 scripts exist. All lib module imports resolve. Blocking behavior matches CLAUDE.md documentation - CLAUDE.md F: vs D: consistency: Both files byte-identical after git sync. Gate count (32), schema count (12), hook count (7), law count (12 including INSTINCT) all correct and consistent
[6.0.31] — 2026-03-05 — Fix literature-registry stale total_databases and revert wrong category count (R50)
Trigger: Round 50 paranoid deep debug — deep-audited README.md, ARCHITECTURE.md, CITATION.cff, NOTICE, marketplace.json, plugin.json, and BEHAVIOR-LOGBOOK references. Found 2 bugs: stale
total_databasesmetadata field and a wrong category count introduced by a previous erroneous "fix" in [6.0.24].
- plugin/db/literature-registry.json line 6:
"total_databases": 108→"total_databases": 102 - WHY: The
total_databasesmetadata field was never updated when databases were removed or reorganized.Object.keys(d.databases).lengthreturns 102, matching the count stated everywhere else in the documentation. The stale 108 value would confuse any code or documentation that reads this field for validation or display purposes.
- ARCHITECTURE.md lines 132, 159:
13 categories→12 categories - WHY: The [6.0.24] "fix" changed "12 categories" to "13 categories", but this was WRONG. The actual
d.categories.lengthis 12. The [6.0.24] WHY text listed "biomedical" as a category, but no such category exists in literature-registry.json — the actual categories are: multidisciplinary, biology_life_sciences, chemistry_pharmacology, physics_math_astronomy, materials_photonics_engineering, computer_science_ai, medicine_clinical, social_sciences_humanities, earth_sciences_environment, preprints_domain_specific, regional, grey_literature_special = 12. This is a cautionary example of a "fix" that introduced a bug by miscounting.
- CHANGELOG.md [6.0.24] section: Added revert annotation and correction note to the erroneous "12→13 categories" entry
- WHY: The CHANGELOG is append-only and historical, so the wrong entry was not deleted but annotated with "REVERTED in [6.0.31]" and the correct count documented. This preserves the audit trail showing how the error was introduced and corrected.
- README.md: 99.7% accurate. Minor line count variance (±1-6 lines from recent edits) and 3 minor omissions (legacy archive dir, root assets/ extra files, v6.0.1 blueprint) — cosmetic, not bugs
- CITATION.cff: All values correct (12 schemas, 32 gates, 7 hooks, v6.0.0)
- NOTICE: Correct (copyright 2026, Apache-2.0, correct repo URL)
- marketplace.json + plugin.json: Both correct (v6.0.0, valid paths, correct metadata)
- BEHAVIOR-LOGBOOK references: Runtime files (.vibe-science/ contents) correctly not present in repo — they are created by
/initcommand at runtime
Trigger: Round 49 paranoid deep debug — audited commands/ and agents/ directories, cross-referenced protocol ↔ reference card mappings, verified archive directory integrity, and checked F: ↔ D: drive consistency.
- commands/init.md lines 102-108: Table column header
Template sourcewith values liketemplates/STATE.md→Section in templates.mdwith values like## STATE.md Template - WHY: The templates live in a single file
skills/vibe/assets/templates.mdwith section headers like## STATE.md Template. The table's "Template source" column usedtemplates/STATE.mdformat, which looks like a file path to a standalone file that doesn't exist. An agent following the table literally would attempt to readtemplates/STATE.mdand fail. Line 100 of the same file correctly states the templates location, but the table contradicted it with path-like references.
- F: ↔ D: drive consistency: All 5 critical files (CLAUDE.md, roles.md, enforcement.md, settings.json, hooks.json) byte-identical across both drives after git sync
- Archive directory (552 files): Well-organized across 9 version directories. All README.md archive references point to existing files. No orphaned active files that should be archived
- Protocol ↔ reference card cross-refs: 20 of 21 protocols have reference cards (95%). Missing: agent-teams.md — intentional omission (TEAM mode is experimental, multi-agent-config.md reference already covers role configuration)
- Python script thresholds: All 6 scripts (dq_gate.py, gate_check.py, spine_entry.py, sync_check.py, tree_health.py, observer.py) — thresholds match gates/gates.md exactly
- agents/ directory: Only reviewer2.md present — BY DESIGN (other roles are behavioral dispositions, not spawnable subagents)
- 32 gate count re-verified: G(7) + L(4) + D(3) + T(4) + B(1) + S(5) + DQ(4) + DD(1) + DC(1) + V(1) + J(1) = 32, all 8 schema-enforced gates confirmed
Trigger: Round 48 paranoid deep debug — audited package.json dependencies vs actual imports across all plugin JavaScript files. Found
onnxruntime-nodelisted as a dependency but never imported or referenced anywhere in the codebase.
- package.json line 13: Removed
"onnxruntime-node": "^1.21.0"from dependencies - WHY: Cross-referencing all
require()and dynamicimport()statements across 17 plugin JS files (7 hooks + 8 lib modules + 2 utilities) found that onlybetter-sqlite3(3 files) and@huggingface/transformers(1 file) are actually used.onnxruntime-nodehas zero references — it was a historical artifact from early development when ONNX runtime was considered for model inference, but the implementation settled on@huggingface/transformersinstead. The unused dependency adds native compilation overhead (onnxruntime-noderequires platform-specific binaries) and installation time for zero functionality.
- Stale number sweep (9 categories): All previously fixed numbers (32 gates, 12 schemas, ~7,800 LOC, /15 triage, 0-3 per dimension, 102 databases, 7 hooks, 12 Laws) remain correct across all active documentation
- Blueprints directory: v6.0-NEXUS-BLUEPRINT.md — all canonical values correct, all referenced paths exist
- Orphaned files: No orphaned .js files in plugin/. Two .gitignore'd planning files (UPGRADE_PLAN_V2.md, CHANGELOG_V2.md) are properly excluded from git — historical artifacts, not bugs
- LICENSE: Apache 2.0, complete and properly formatted (192 lines)
- Cross-file path references: All path references in .md and .js files resolve to existing files. All 13 primary directories verified. Dual-config hook strategy (dev vs plugin mode) confirmed synchronized
Trigger: Round 44 paranoid deep debug — audited
.claude/rules/files against CLAUDE.md. Found version label mismatch between CLAUDE.md file reference and enforcement.md's own header.
- CLAUDE.md line 42:
v5.0 structural enforcement→v6.0 structural enforcement - WHY: The file
.claude/rules/enforcement.mdidentifies itself as# v6.0 Structural Enforcementin its header (correctly — it describes v6.0 mechanisms). But CLAUDE.md's file reference still labeled it asv5.0, a leftover from when the enforcement section was inline under the## v5.0 STRUCTURAL ENFORCEMENTheader (which marked when the mechanisms were introduced). When roles and enforcement were extracted to.claude/rules/, the version label in the reference wasn't updated to match the file's own header.
- 12 JSON Schema files: All valid JSON Schema (draft 2020-12), all 8 gate mappings correct, no stale references
- Plugin metadata: marketplace.json, plugin.json, package.json — versions (6.0.0), URLs, counts all consistent
- Round 41-42 fixes re-audited: No regressions introduced — all 10 fixes verified in-place
- "CRYSTALLIZE" references: Confirmed as phase name (LAW 10, MCTS cycle), not folder — NOT a bug
- "(v5.0 scale: 0-20)" label: Confirmed as versioning provenance tag (v5.0 introduced the 0-20 scale) — NOT a bug
Trigger: Round 42 paranoid deep debug — cross-referenced canonical numbers across ALL active files. Found 3 more inconsistencies from R41's fixes not propagated to all locations.
- SKILL.md lines 453-459:
(0-5)per dimension and/25total →(0-3)per dimension and/15total - WHY: The canonical protocol (protocols/brainstorm-engine.md:221) defines 5 dimensions scored 0-3 each, max 15. The root SKILL.md (biology-instance) still had the pre-canonical (0-5) scale from early drafts, totaling /25. Same class of bug as the skills/vibe/SKILL.md fix in [6.0.26], but in the biology-instance copy.
- README.md lines 115, 305:
~7,100 LOC→~7,800 LOC - WHY: ARCHITECTURE.md's line count table was updated to ~7,800+ in [6.0.26] after verifying all 19 component files with
wc -l. README.md referenced the same total but was not updated in the same round. Both should match.
- CHANGELOG.md line 66:
108-database registry→102-database registry - WHY: The [6.0.24] entry marked literature-registry.json as "verified CLEAN" with "108-database registry", but the actual database count (verified by JSON parsing in R41) is 102. The 108 count was from a pre-[6.0.7] version; [6.0.7] fixed the count from 108→102 in README.md, but the [6.0.24] verification claim still used the old number.
[6.0.26] — 2026-03-05 — Documentation accuracy audit R41: schema count, URL, algorithm, category count, scoring scale, line counts
Trigger: Round 41 paranoid deep debug — full audit of all remaining documentation files (CITATION.cff, NOTICE, ARCHITECTURE.md table, skills/vibe/SKILL.md). Cross-verified every factual claim against source code and data files. 7 bugs found: 4 factual errors propagated from earlier versions, 1 wrong URL, 1 stale line count table with 10+ wrong entries, 1 scoring scale error.
- CITATION.cff line 17:
9 read-only schema files→12 read-only schema files - WHY: Same root cause as SKILL.md fix in [6.0.25] — the canonical schema count is 12 (9 v5.0 + 3 v5.5). The CITATION.cff abstract was written during v5.0 when there were 9 schemas, and was never updated when v5.5 added 3 more.
- NOTICE line 13:
https://github.com/vibe-science-contributors/vibe-science→https://github.com/th3vib3coder/vibe-science - WHY: NOTICE used a placeholder organization URL that doesn't exist. CITATION.cff line 62 has the correct repository URL. Anyone following the NOTICE citation link would get a 404.
- ARCHITECTURE.md line 125:
a deterministic SHA-256 hash vector→a deterministic character-code hash vector - WHY: The actual
simpleEmbedding()function in worker-embed.js usestext.charCodeAt(i) / 255accumulation into a Float32Array, NOT SHA-256 hashing. SHA-256 produces a 256-bit digest (not a 384-dim float vector). The description would mislead anyone trying to understand or reproduce the fallback behavior.
- ARCHITECTURE.md lines 132, 159:
12 categories→13 categories— THIS FIX WAS WRONG. Actual category count is 12, not 13. The WHY below miscounted by including a phantom "biomedical" category that does not exist in literature-registry.json (the actual categories are: multidisciplinary, biology_life_sciences, chemistry_pharmacology, physics_math_astronomy, materials_photonics_engineering, computer_science_ai, medicine_clinical, social_sciences_humanities, earth_sciences_environment, preprints_domain_specific, regional, grey_literature_special = 12). Corrected back to 12 in [6.0.31]. - Original (wrong) WHY: Actual unique categories in literature-registry.json: biology_life_sciences, biomedical, chemistry_pharmacology, computer_science_ai, earth_sciences_environment, grey_literature_special, materials_photonics_engineering, medicine_clinical, multidisciplinary, physics_math_astronomy, preprints_domain_specific, regional, social_sciences_humanities = 13. The count was likely 12 before a category was added.
- ARCHITECTURE.md lines 141-160: Updated all 19 component line counts to match actual
wc -lvalues - WHY: Files were edited during 40 rounds of debug, but the table was never updated. Major discrepancies: post-tool-use.js (1,482→1,765), gate-engine.js (630→471), db.js (~500→668), stop.js (171→258), literature-registry.json (~800→952). Total: ~7,100+ → ~7,800+. Stale line counts would mislead anyone estimating plugin complexity.
- skills/vibe/SKILL.md line 159:
/25→/15 - WHY: The TRIAGE scoring system uses 5 dimensions scored 0-3 each, maximum 15. Canonical definition at protocols/brainstorm-engine.md:221 says "5 dimensions (0-3 each, max 15)". The /25 implies a 0-5 scale per dimension, which doesn't match the protocol. Same class of bug as the brainstorm-engine.md fix in [6.0.25].
[6.0.25] — 2026-03-05 — Cross-directory consistency audit R40: SKILL.md schema count, brainstorm scoring scale
Trigger: Round 40 paranoid deep debug — full cross-directory consistency audit comparing
protocols/vsskills/vibe/references/(20 common files) andassets/vsskills/vibe/assets/(6 common files). Verified that content divergence between directories is BY DESIGN (protocols/ = full implementation docs with TEAM mode + v5.5 sections; references/ = condensed domain-agnostic reference cards). Confirmed 13 same-size files are content-identical (only LF vs CRLF line endings). Two real bugs found.
- SKILL.md line 330:
All 9 JSON schemas: unchanged (read-only)→All 9 v5.0 JSON schemas: unchanged (3 new schemas added in v5.5: data-quality-gate, finding-validation, spine-entry; total: 12) - WHY: Line 329 explicitly mentions gate additions from v5.5, but line 330 still said "9 schemas" without mentioning the 3 new schemas added in v5.5. The canonical schema count is 12 (9 original v5.0 + 3 new v5.5). A reader comparing these adjacent lines would think schemas weren't updated, contradicting the actual file count in
schemas/.
- skills/vibe/references/brainstorm-engine.md line 253:
0-20 scale→0-15 scale - WHY: The hypothesis scoring system uses 5 dimensions scored 0-3 each, for a maximum of 15 points. The canonical definition at
protocols/brainstorm-engine.md:263correctly says "5 dimensions (0-3 each, max 15)". The Near-Tie Rule example in the references/ copy said "0-20 scale", which is wrong — the protocols/ copy correctly says "0-15 scale" at line 295. This would cause agents using the references/ version to miscalculate the 10% score difference threshold.
- 13 files content-identical (protocols/ vs references/): analysis-orchestrator, audit-reproducibility, auto-experiment, blind-first-pass, circuit-breaker, data-extraction, judge-agent, schema-validation, search-protocol, seeded-fault-injection, tree-search, vlm-gate, writeup-engine ✓
- 6 files with expected larger protocols/ (TEAM mode, v5.5 sections): brainstorm-engine (+59 lines), reviewer2-ensemble (+383 lines, completely different structure), evidence-engine (+25 lines), experiment-manager (+26 lines), loop-otae (+27 lines), serendipity-engine (+13 lines) ✓
- 1 expected biology-vs-generic difference: knowledge-base.md (CRISPR DOIs vs generic placeholders) ✓
- 6 asset files with expected domain differences: fault-taxonomy.yaml, judge-rubric.yaml, metric-parser.md, node-schema.md, stage-prompts.md, templates.md ✓
[6.0.24] — 2026-03-05 — Schema index fix + full non-JS audit R39: schema.sql, 6 Python scripts, configs
Trigger: Round 39 paranoid deep debug — audited all 16 remaining non-JavaScript files: schema.sql (DB schema), hooks.json (plugin hooks), __test_e2e.mjs (E2E tests), plugin.json, package.json, marketplace.json, domain-config-template.json, literature-registry.json, claude-code.yaml, and 6 Python scripts (gate_check.py, sync_check.py, tree_health.py, observer.py, spine_entry.py, dq_gate.py). One real bug found: index name typo + column mismatch between schema.sql and worker-embed.js.
- schema.sql line 221:
idx_memembed_project ON memory_embeddings(project_path)→idx_membed_project ON memory_embeddings(project_path, created_at) - WHY: Two problems: (1) Typo in index name —
idx_memembed_project(double 'm') vsidx_membed_projectin worker-embed.js (correct spelling). Since SQLite treats these as different indexes, both would be created, causing redundancy. (2) Column mismatch — schema.sql had a single-column index on(project_path), while worker-embed.js correctly created a composite index on(project_path, created_at). The composite index is strictly better: it covers both project filtering AND temporal ordering (used by vector search queries). Now both schema.sql and worker-embed.js create the same index with the same name.
- hooks.json: All 7 hooks present, correct script paths via
${CLAUDE_PLUGIN_ROOT}, correct matchers ✓ - __test_e2e.mjs: B1-B7 test suite. 12 EXPECTED_TABLES correct (memory_embeddings is optional fallback, excluded). Index check covers 14 of ~17 ✓
- plugin.json: Plugin manifest, version 6.0.0, Apache-2.0 ✓
- package.json: Dependencies correct (better-sqlite3, @huggingface/transformers, onnxruntime-node), engines >=18.0.0 ✓
- marketplace.json: "32 quality gates" matches canonical number ✓
- domain-config-template.json: Multi-domain template with CRISPR/photonics/particle-physics examples (known exception) ✓
- literature-registry.json: 102-database registry, reference data only ✓
- claude-code.yaml: 7 agent types match AGENTS.md, model tiers logical, web search rule present ✓
- gate_check.py: Lightweight JSON Schema validator, correct type/required/properties/items/min/max/enum handling ✓
- sync_check.py: Number extraction with tolerance, percentage conversion fallback, correct skip patterns ✓
- tree_health.py: T3 gate checks — good ratio, exploration ratio (LAW 8), stale branches, branch diversity ✓
- observer.py: Orphan detection, desync check, design drift, naming consistency. Dual stage format check covers patterns ✓
- spine_entry.py: Entry creation with argparse validation, VALID_TYPES enforced. CALIBRATE/CALIBRATION intentional distinct types ✓
- dq_gate.py: DQ1-DQ4 checks with configurable thresholds, YAML/JSON config fallback, Bessel's correction in CV calculation ✓
[6.0.23] — 2026-03-05 — Per-claim error handling + Windows path fix R38: subagent-stop.js, r2-calibration.js, worker-embed.js
Trigger: Round 38 paranoid deep debug — full JavaScript code audit of all 16 plugin/ files (8 hook scripts + 8 lib modules). Launched Explore agents to audit each file, then manually verified every finding against actual code. Most agent findings were FALSE POSITIVES (e.g., better-sqlite3
.all()null check — it always returns array; vecSearch uncaught error — already wrapped in try/catch; db null in fallbackBuildContext — caller already guards). Three REAL bugs survived verification.
- subagent-stop.js lines 63-77: Wrapped individual seed-check query in try/catch inside the for-loop
- WHY: If one claim's seed query failed (e.g.,
serendipity_seedstable missing or corrupted row), the entire for-loop threw an unhandled exception. The outer catch (line 98) caught it but exited with code 0 (allow), meaning the Salvagente Rule was silently bypassed for ALL killed claims — not just the one that failed. Now each failed query treats that individual claim as "missing seed" and continues checking the rest.
- r2-calibration.js
updateSeedStatuseslines 208-216: Wrapped individualstmt.run()in try/catch inside the for-loop - WHY: Same pattern as subagent-stop.js. If
stmt.run()threw for one seed update (e.g., seed_id not found, constraint violation), all subsequent seed updates in the batch were lost. Now each failed update is silently skipped and remaining seeds are still processed.
- worker-embed.js line 26: Added
import { fileURLToPath } from 'node:url' - worker-embed.js line 56:
path.dirname(new URL(import.meta.url).pathname)→path.dirname(fileURLToPath(import.meta.url)) - WHY: On Windows with Node < 21 (where
import.meta.dirnameis unavailable),new URL(import.meta.url).pathnamereturns/F:/path/...with a leading forward slash.fs.existsSync()at line 217 fails because/F:/pathis not a valid Windows path.fileURLToPath()fromnode:urlcorrectly strips thefile://protocol and handles Windows drive letters, returningF:\path\....
- session-start.js:
fallbackBuildContext(db)null — caller at line 282 guardsif (db && dbAvailable)✓ - stop.js:
.all()null check — better-sqlite3.all()always returns array ✓ - pre-compact.js: Cross-session query — intentional (different semantics from stop.js) ✓
- db.js:
JSON.parse(null)in upsertPattern — catch block handles gracefully ✓ - context-builder.js: vecSearch uncaught — lines 78-87 already have try/catch ✓
- vec-search.js: Keyword splitting
C-001— regex[^a-z0-9\s-]preserves hyphens ✓ - r2-calibration.js: Decay docstring "2% per week" —
e^(-0.02) ≈ 0.98, accurate for small rates ✓ - r2-calibration.js: Silent JSON parse failure — explicit graceful degradation by design ✓
- worker-embed.js:
loadEmbeddingModel()not awaited — intentional non-blocking (line 470 comment) ✓ - worker-embed.js: Race condition
modelReady/embeddingPipeline— JS single-threaded, assignment order safe ✓ - prompt-submit.js, pre-tool-use.js, setup.js: CLEAN — no issues found ✓
- narrative-engine.js, pattern-extractor.js, permission-engine.js: CLEAN — no issues found ✓
Trigger: Round 37 paranoid deep debug — JavaScript code logic audit of gate-engine.js and post-tool-use.js. Found critical claim ID format mismatch: ALL templates/protocols use
C-001format (with hyphen, 643 occurrences across 109 files), butextractClaimId()regexes only matchedC001(without hyphen). This caused gate checks and Salvagente rule to silently skip all claims using the canonicalC-xxxformat. Also found asymmetric case sensitivity: prerequisite gate check only matchedCLAIM-LEDGER(uppercase), while Salvagente check matched both cases.
- gate-engine.js line 74:
extractClaimIdregex/\bC(\d{3})\b/→/\bC-?(\d{3})\b/— now matches bothC001andC-001 - gate-engine.js line 98:
getRequiredGatesForClaimregex/^C(\d)\d{2}$/→/^C-?(\d)\d{2}$/— now matches both formats for tier extraction - post-tool-use.js line 645:
extractClaimIdregex/\bC(\d{3})\b/→/\bC-?(\d{3})\b/— same fix, kept in sync with gate-engine.js - WHY: Templates universally use
C-001format (with hyphen). The old regexes requiredC001(no hyphen), causing all gate prerequisite checks, tier-based gate routing, and Salvagente rule enforcement to silently fail. Claims were written to the ledger without gate validation becauseextractClaimIdreturned null for the canonical format.
- post-tool-use.js line 361:
filePath.includes('CLAIM-LEDGER')→filePath.includes('CLAIM-LEDGER') || filePath.includes('claim-ledger') - WHY: The prerequisite gate check (line 361) only matched uppercase
CLAIM-LEDGER, while the Salvagente check (line 384) already matched both cases. A lowercaseclaim-ledger.mdpath would bypass prerequisite gates but still trigger Salvagente — asymmetric enforcement.
Trigger: Round 35 paranoid deep debug — full canonical number consistency scan across all live files. Verified: 12 Laws, 32 gates (8 schema-enforced), 12 schema files, 7 lifecycle hooks, 36 reference documents, 7 agent roles (AGENTS.md), 6 permission roles (Permission Engine), 21 protocols, 12 SQLite tables, 6 Python scripts, 7 skill asset files. Found terminological ambiguity: CLAUDE.md said "6 agent types" while README.md uses "7 agent roles" — these refer to different classification systems (Permission Engine vs AGENTS.md sub-agent definitions).
- Line 41:
all 6 agent types→all 6 permission roleswith explanatory note distinguishing from the 7 AGENTS.md agent types - WHY: CLAUDE.md used "agent types" for the 6 Permission Engine roles (researcher, R2, serendipity, experimenter, lead, judge), which could be confused with the 7 AGENTS.md sub-agent definitions (researcher, r2-deep, r2-inline, observer, explorer, r3-judge, instinct-scanner). These are intentionally different classification systems: Permission Engine controls access control in TEAM mode, AGENTS.md defines model selection and disposition for sub-agent spawning.
Trigger: Round 34 paranoid deep debug — full repo-wide grep sweep across all live files (excluding archive/, CHANGELOG.md, SKILL.md, fault-taxonomy.yaml, domain-config-template.json). Found
AnnData schema contractin ARCHITECTURE.md. Historical/provenance references in ARCHITECTURE.md (lines 25, 168, 374 — CRISPR research origin story) and blueprints/ are intentional context, not domain assumptions.
- Line 447:
AnnData schema contract→data schema contract - WHY: The Data Extraction protocol summary references a domain-agnostic schema contract.
AnnDatais a single-cell-specific data structure; the contract applies to any structured data format.
Trigger: Round 32-33 paranoid deep debug — scanned 16 references-only files (all PASS), agent-teams.md (PASS), plugin/ directory, hooks/, .claude-plugin/, root-level files, schemas/, and assets/. Found CRISPR-specific confounder list in templates.md and
anndataas first tool example in skill-router.md.
- Line 278: Confounder harness template
[list: n_mm, affinity, PAM, region, guide RE, etc.]→[list: age, treatment, batch, site, collection_method, etc.] - WHY: The confounder harness template is domain-agnostic. Example confounders should demonstrate the concept without assuming a specific research domain.
n_mm(mismatches),PAM, andguide REare CRISPR-specific terms.
- Line 37: Tool examples
(e.g., anndata, polars, pandas)→(e.g., polars, pandas, dask) - WHY:
anndatais a single-cell-specific Python library. The tool example list should reference domain-agnostic data tools.
[6.0.18] — 2026-03-04 — Domain generalization R31: audit-reproducibility, handoff-protocol, pattern-extraction
Trigger: Round 31 paranoid deep debug — full content-diff of protocols/ (21 files) vs skills/vibe/references/ (36 files). Found
batch_keyin both copies of audit-reproducibility.md, CRISPR examples in references/handoff-protocol.md and references/pattern-extraction.md (3 occurrences). Also confirmed: the two directories are intentionally different versions (protocols/ = full operational specs, references/ = condensed plugin references), not copies that drifted.
Fixed — Domain Generalization (protocols/audit-reproducibility.md + references/audit-reproducibility.md)
- Line 71:
batch_key | source_id | source_id | same→group_label | treatment | treatment | same - WHY:
batch_keyis a scRNA-seq/scVI-specific parameter name. The audit parameter diff table should use domain-neutral names.
- Lines 117-129: Replaced CRISPR-specific handoff example: "batch-corrected DE signature in CRISPR perturbation data" → "batch-corrected integration signature in treatment-response data", "library size" → "sample size"
- WHY: Handoff examples between R2 and Researcher should demonstrate the protocol mechanics without assuming a specific research domain.
- Line 49: JSON example
"CRISPR off-target effects single-cell"→"optimal transport data integration methods" - Line 63: Actionable output example replaced CRISPR query with domain-neutral query
- Line 208: SessionStart pattern example
"CRISPR off-target"→"optimal transport integration" - WHY: Pattern extraction examples demonstrate the REPEATED_ACTION detection pattern — the example queries should be domain-agnostic.
Trigger: Round 30 paranoid deep debug — full scan of gates/, commands/, agents/, .claude-plugin/, protocols/, and ARCHITECTURE.md. Found CRISPR/GUIDE-seq examples in 2 command templates, scRNA-seq/CRISPR examples in 1 protocol, and a protospacer example in gates.
- Lines 45-168: Replaced entire CRISPR/GUIDE-seq worked example with domain-neutral "data integration" example. Changed: RQ ID (
uot-crispr→transport-integration), search queries (GUIDE-seq → data integration), paper examples (Tsai 2015, Lazzarotto 2020 → Author A 2020, Author B 2022), findings (off-target site counts → integration scores), extracted data description. - WHY: Command templates are domain-agnostic reusable components. The worked example should demonstrate the OTAE loop mechanics without assuming a specific research domain.
- Line 62: Scopus example
CRISPR AND "off-target"→"optimal transport" AND "data integration" - Line 93: PubMed example
CRISPR[Title] AND off-target→"optimal transport"[Title] AND "data integration" - Line 108: OpenAlex example
CRISPR off-target→optimal transport data integration - Lines 120-143: Gap analysis example replaced:
CRISPR→data integration,off-target→cross-domain transfer - WHY: Search syntax examples should demonstrate query patterns, not assume a specific research topic.
- Line 41:
"assay": "scRNA-seq"→"assay": "RNA-seq" - Line 44:
"Endometriosis vs control, 10X Chromium"→"Treatment vs control, paired design" - Lines 50-56: Method domain changed from
"single-cell"to"data-quality", removed scverse reference - Lines 61-63: Author example changed from
"domain": ["CRISPR", "off-target", "GUIDE-seq"]→"domain": ["optimal-transport", "data-integration", "methodology"] - WHY: The knowledge base JSON schema examples should be domain-neutral to work for any research project.
- Lines 554-556: DD0 gate example
"Protospacer_sequence" may not be the designed protospacer→"normalized_score" may not use the normalization method you expect - WHY: The DD0 gate teaches "column names lie" — the example should use a generic column name to be universally applicable.
protocols/knowledge-base.md: 0 domain-specific terms — CLEANcommands/loop.md: 0 domain-specific terms — CLEANcommands/search.md: 0 domain-specific terms — CLEANgates/gates.md: 0 remaining protospacer/CRISPR references — CLEANskills/vibe/references/knowledge-base.md: already clean (no sync needed)
Trigger: Round 29 paranoid deep debug — full scan of assets/, gates/, schemas/, hooks/, plugin/, and root files. Found 3 asset files with scRNA-seq/CRISPR-specific content and stale reference count in README.md.
- Complete rewrite: Entire file was scRNA-seq specific (AnnData
.obs,import scanpy as sc, platform maps for 10X_v2/SmartSeq2/DropSeq/InDrop/CELSeq2, MT- gene prefix detection, pct_mito/pct_ribo computation, scVI references in pitfalls table). Replaced with domain-agnostic "Data Normalizer Standard" using generic DataFrame operations, collection method standardization, completeness metrics, and outlier flags. - WHY: This file defined how to normalize dataset metadata. The AnnData/.obs pattern was biology-specific. The new version works for any tabular data with categorical metadata columns.
- Line 22: "scRNA-seq pipeline | analysis-orchestrator.md → scanpy + scvi-tools skills" → "Data analysis pipeline | analysis-orchestrator.md → domain-appropriate analysis skills"
- Line 24: "Data exploration | → exploratory-data-analysis + anndata skills" → "Data exploration | → exploratory-data-analysis + domain data skills"
- Lines 33-46: Replaced "Single-Cell Analysis Pipeline" table (AnnData, scanpy, scVI, HVG, batch_key, iLISI, obs-normalizer) with generic "Data Analysis Pipeline" table using domain-neutral terms.
- Line 125: "Obs schema normalization" → "Data schema normalization"
- WHY: The routing table hard-coded a single-cell workflow as if it were the only analysis type. The new version is domain-agnostic while still providing concrete examples.
- Lines 88-96: Replaced "scRNA-seq Integration" table (ilisi, clisi, asw_batch, asw_label) with "Clustering / Integration" table (nmi, ari, asw, completeness, homogeneity, v_measure).
- Lines 98-104: Replaced "CRISPR Off-Target" table with "Detection / Anomaly" table (same metrics but domain-neutral naming).
- WHY: Metric examples should demonstrate the parser's format, not assume a specific research domain. The new tables use standard clustering and detection metrics applicable to any field.
- Updated "34 reference documents" → "36 reference documents" in 5 locations (lines 46, 145, 151, 298, 402). Actual count confirmed: 36 files in skills/vibe/references/.
- WHY: Count became stale after R26 added vlm-gate.md and analysis-orchestrator.md to references/. Previous rounds updated the references but not the README count.
- Lines 35, 42, 50 contain CRISPR-specific examples (PAM, mismatches, cleavage). This file is marked HUMAN-ONLY modification in CLAUDE.md — flagged for user review, not modified by agent.
assets/obs-normalizer.md: 0 domain-specific terms (AnnData/scRNA/scanpy/scVI/10X/MT-/pct_mito) — CLEANassets/skill-router.md:anndataappears only as one example in "(e.g., anndata, polars, pandas)" — ACCEPTABLEassets/metric-parser.md: 0 domain-specific terms (scRNA/CRISPR/ilisi/clisi/off-target) — CLEANREADME.md: 0 remaining "34 reference" — all updated to 36- Reference count verified: 36 files in skills/vibe/references/
[6.0.15] — 2026-03-04 — Domain generalization R28: data-extraction, analysis-orchestrator, vlm-gate Unicode
Trigger: Round 28 paranoid deep debug — full scan of all 20 protocols/ and 36 references/ files. Found domain-specific content leakage in 2 protocol files that had been missed in earlier rounds, plus Unicode □ in vlm-gate (both copies).
- AnnData Contract (scRNA-seq): Entire section (lines 15-52) replaced with domain-agnostic Structured Data Contract. Removed:
.obs,.var,.X,.raw.X,cell_type,platform,10X_v2,10X_v3,SmartSeq2,n_genes,n_counts,pct_mito,pct_ribo,doublet_score,Scrublet,DoubletFinder,gene_symbols,ensembl_ids,anndata,h5ad. Replaced with:source_id,sample_id,group_label,collection_method,feature_names,raw_values. - Schema Violation Triage (lines 69-77): Replaced scRNA-specific violations (
X contains float,.raw.X,cell_type,obs-normalizer,pct_mito,MT- gene prefix,var_names_make_unique) with domain-agnostic violations (Values pre-transformed,Missing source_id,Missing group_label,Wrong data types,Missing quality metrics,Duplicate feature names,Mixed identifier formats). - DD0 Gate examples (lines 107, 128-131): Replaced CRISPR-specific references (
M7: CHANGE-seq alignment bug,guide_id,off_target_seq,guide RNA,CHANGE-seq signal) with domain-agnostic examples (sample_id,feature_name,raw_count,normalized_score). - File now matches references/ version (both include DD0 section).
- Overwritten with generalized references/ version (already fixed in R26). Removed residual scRNA-seq content:
scVI,n_latent,n_HVG,batch_key,pct_mito,hvg_flavor,seurat_v3,scanpy,scvi-tools,anndata,h5ad,iLISI,cLISI,kBET. protocols/ copy had been missed when references/ was generalized in R26.
- protocols/vlm-gate.md: replace_all □ (U+25A1) →
[ ](5 occurrences at lines 76-80, Gate G6 Pass/Fail checklist) - references/vlm-gate.md: Same fix, 5 occurrences. Both copies now ASCII-clean.
- Added Data Dictionary Protocol (v5.5) — Gate DD0 section that existed in protocols/ but was missing from references/. Domain-agnostic examples used. Both copies now identical.
protocols/data-extraction.md: 0 domain-specific terms (AnnData/scRNA/CRISPR/h5ad/obs/var/pct_mito/Scrublet/CHANGE-seq) — CLEANprotocols/analysis-orchestrator.md: 0 domain-specific terms — CLEAN (matches references/)protocols/vlm-gate.md: 0 Unicode □ — CLEANreferences/vlm-gate.md: 0 Unicode □ — CLEANreferences/data-extraction.md: DD0 section present, 0 domain-specific terms — CLEAN- protocols/ and references/ copies are now identical for: data-extraction.md, analysis-orchestrator.md, vlm-gate.md
Trigger: Round 26 paranoid deep debug — full repo scan for cross-file consistency (gate counts, schema counts, law counts, file references) + identification of 3 protocol files missing from references/.
- archive/v6.0-NEXUS-BLUEPRINT.md:2330: Schema-enforced gate count "12 schema" → "8 schema" in v5.5 and v6.0 columns. Blueprints/ copy was already correct; archive/ copy was stale. The "12" referred to schema FILES (tracked in row below), not schema-enforced GATES (canonical: 8).
- CHANGELOG.md:266: "$id fields: 9 schemas" → "All 12 schemas updated to vibe-science-v6.0 (9 from v5.0, 3 from v5.5)". Original was accurate but incomplete — omitted the 3 v5.5-origin schemas.
- skills/vibe/references/vlm-gate.md: Copied from protocols/vlm-gate.md (already 100% domain-agnostic). VLM Gate Protocol for figure quality validation — was missing from the skill package entirely.
- skills/vibe/references/analysis-orchestrator.md: Copied from protocols/analysis-orchestrator.md with full domain generalization. Protocol structure is universal; biology-specific examples (scVI, scanpy, anndata, h5ad, HVG, iLISI/cLISI, kBET, pct_mito, batch_key, n_latent) replaced with domain-agnostic placeholders. Manifest.json, report template, ablation matrix, and script library patterns all generalized.
- protocols/agent-teams.md: TEAM mode operational runbook. references/multi-agent-config.md already covers essential reference material — agent-teams.md is the detailed how-to guide, appropriate for protocols/ only.
references/analysis-orchestrator.md: 0 domain-specific terms (scVI/scanpy/anndata/h5ad/HVG/CRISPR/pct_mito/batch_key/n_latent/ELBO/iLISI/cLISI/kBET) — CLEANreferences/vlm-gate.md: 0 domain-specific terms — CLEAN- Gate counts (32/8), law counts (12), schema counts (12) verified consistent across CLAUDE.md, SKILL.md, ARCHITECTURE.md, README.md, gates/gates.md, blueprints/, protocols/
Trigger: Round 25 paranoid deep debug — scanned all remaining .md files for Unicode symbols, domain-specific biology terms, and stale version references. Found 3 targets:
gates/gates.md(177 Unicode checkboxes + biology-specific G0/G1/G2/G3/G5/L-1 gate content),SKILL.md(stale serendipity threshold).
- replace_all □ (U+25A1) →
[ ](177 occurrences) — checkbox symbols incompatible with ASCII-only rendering - replace_all ≥ (U+2265) →
>=(6 occurrences) — mathematical symbols to ASCII
- G0 (Input Sanity): Replaced
.X contains integer counts,.X.max(),.raw.X,gene names,UMI data,normalization historywith generic:Types match expected format,Value ranges plausible,Raw/original preserved,Identifiers present - G1 (Schema Compliance): Replaced
.obs,study_id,category dtype,pct_mito,scVI,scanpy-only workflow,MT- gene prefix,obs-normalizer,HVG selectionwith generic:Schema matches expectation,grouping columns correct types,Domain-appropriate QC metrics,data-dictionary.md - G2 (Design Justification): Replaced
batch_key→Grouping key,n_HVG→Feature count,HVG selection method stated (seurat_v3, cell_ranger, etc.)→Feature selection method stated and justified,batch_key and biology→grouping key and signal of interest - G3 (Training Integrity):
ELBO/loss→Loss/objective,reduce n_latent→reduce model complexity,non-integer counts→check input format - G5 (Artifact Completeness):
h5ad, model→data, model - L-1 (Literature Pre-Check):
"CRISPR off-target"example →"domain application"generic example
- Line 1053: serendipity threshold
>= 12→>= 15— was still at v4.0 scale, v6.0 uses 0-20 scale with threshold 15
gates/gates.md: 0 matches for scVI/HVG/ELBO/pct_mito/MT-/CRISPR/seurat_v3/cell_ranger/batch_key/n_HVG/n_latent/h5ad/scRNA/AnnData/.obs/.X/gene_name — CLEANgates/gates.md: 0 Unicode symbols (□/≥/≤/─) — CLEANSKILL.md: 0 Unicode symbols — CLEAN (biology-specific content like R2-Bio, CRISPR intentionally preserved — SKILL.md is biology instance, not generic template)
Trigger: Round 24 paranoid deep debug — the final deferred file from Round 23.
protocols/reviewer2-ensemble.md(650+ lines) had extensive unique content not present inskills/vibe/references/reviewer2-ensemble.md(327 lines), so it required targeted edits rather than full replacement. The protocols/ copy was still at v5.5 with biology-specific language, old INLINE checklist format, missing v6.0 sections (Multi-Agent Delegation, Temporal Decay Calibration), and Unicode symbols throughout.
- replace_all □ (U+25A1) →
[ ](~60 occurrences) — checkbox symbols incompatible with ASCII-only rendering - replace_all ≥ (U+2265) →
>=(9 occurrences) — mathematical symbols to ASCII - replace_all ≤ (U+2264) →
<=(1 occurrence) — mathematical symbols to ASCII - replace_all ─ (U+2500 box drawing) →
-— YAML decorative line characters to ASCII
- replace_all
R2-Bio→R2-Domain(4 instances) — v6.0 generalized the biology-specific reviewer - R2-Domain table row: "Biology | Biological plausibility..." → "Domain-specific rigour | Loads checklist from
domain-config.yamlif present; otherwise applies generic domain checks" - R2-Domain Checklist: Replaced entire biology-specific checklist (gene names, cell types, doublets, ambient RNA, marker validation) with generic domain checklist (measurement validity, construct operationalisation, field artifacts, reporting requirements, terminology consistency)
- 10 domain-specific examples generalized: batch_key/n_HVG/n_latent → key hyperparameters/feature count/latent dimensions; bulk RNA-seq → generic data type mismatch; cell-type proportions → subgroup proportions/compositional effects; 10X/Smart-seq2 → list platforms/instruments; iLISI metric → generic metric; HVG selection → feature selection; scVI on normalized → model on pre-processed; organism gene symbols → identifiers/nomenclature; iLISI/cLISI/ELBO/HVGs demanded evidence → generic metric/convergence/feature evidence
- Header:
# Reviewer 2 Ensemble Protocol v5.5→# Reviewer 2 Ensemble Protocol v6.0with domain-agnostic attribution line - INLINE 7-point checklist: Replaced v5.5 code-block format (TRACEABLE, SURVIVES HOSTILE READ) with v6.0 table format (Prior Art, Confounder Risk, Reproducible) + YAML storage block
- Activation Modes table: Expanded from 4-column (Mode/Trigger/Scope/Blocking) to 5-column (Mode/Trigger/Blocking/Sub-agent type/Description) matching references/ format
- Multi-Agent Delegation section: R2-DEEP vs R2-INLINE sub-agent types (opus for FORCED/BATCH/VETO, sonnet for BRAINSTORM/SHADOW/REDIRECT/INLINE), spawning details, SOLO mode behavior
- v6.0 TEMPORAL DECAY CALIBRATION section: Exponential decay formula
weight = exp(-0.02 * ageWeeks), calibration data loading at session start, usage instructions for R2 agents, feedback loop tracking
- Version history paragraphs (v3.5/v4.0/v5.0/v5.5 evolution)
- R2 System Prompt section with behavioral directives
- R2 Shadow Mode Protocol
- INLINE mode detailed sections (Why/When/Verdict/Interaction/Cost)
- Ensemble Composition + When to Invoke tables
- Modified FORCED Review Flow (8-step detailed process)
- Salvagente Protocol + Circuit Breaker Integration
- Mandatory Output Schema (full YAML template)
- Domain-Specific Checklists (R2-Methods, R2-Stats with Confounding Audit Table, R2-Domain, R2-Engineering)
- Red Flag Checklist (12 flags: 6 statistical + 6 methodological)
- Reviewer Persona and Invocation Procedure sections
- Review Severity Guide (0-100 numeric scoring)
- Unicode scan (□/≥/≤/─): 0 matches — CLEAN
- Domain-specific scan (scRNA/scVI/iLISI/cLISI/ELBO/HVG/doublet/ambient RNA/10X/Smart-seq/MT-/mt-): 0 matches — CLEAN
- R2-Bio scan: 0 matches — CLEAN
- v5.5 references: 3 matches, all in historical/contextual text (correct to keep)
- Final file: 710 lines, structurally sound
Trigger: Round 23 paranoid deep debug — systematic comparison of all 18 paired files between
protocols/(repo context) andskills/vibe/references/(plugin context). The references/ copies had been updated to v6.0 standards (generalized language, corrected thresholds, v6.0 hook sections) but protocols/ copies were stale. Found: Unicode symbols instead of ASCII, domain-specific CRISPR/scRNA-seq examples instead of generic language, old serendipity thresholds (>=12 instead of >=15), missing v6.0 sections, and Sprint 17 case study references instead of anonymized lessons.
- blind-first-pass.md: Full sync from references/ (Unicode arrows/symbols to ASCII)
- circuit-breaker.md: Full sync from references/ (Unicode to ASCII, generalized examples)
- seeded-fault-injection.md: Full sync from references/ (Unicode to ASCII, generic fault examples)
- judge-agent.md: Full sync from references/ (Unicode to ASCII)
- tree-search.md: Full sync (serendipity >= 12 to >= 15, Unicode to ASCII throughout)
- writeup-engine.md: Full sync (Unicode checkbox to ASCII [])
- search-protocol.md: Full sync (CRISPR-specific search examples to generic)
- auto-experiment.md: Full sync (Unicode to ASCII, domain-specific dispatch to generic)
- audit-reproducibility.md: Full sync (CRISPR/scRNA examples to generic, Unicode to ASCII)
- schema-validation.md: Schema list expanded 9 to 12 (added v6.0 schemas)
- knowledge-base.md: Unicode arrows/tree-chars to ASCII
- experiment-manager.md: 6 targeted edits (Unicode to ASCII, CRISPR dispatch to generic); kept unique DC0 section
- serendipity-engine.md: 7 targeted fixes (Unicode to ASCII, old thresholds to v5.0 scale); added v6.0 CROSS-SESSION SEED SURVIVAL section
- brainstorm-engine.md: 15 edits across 3 context windows (collision score 0-15 to 0-20, promotion threshold 8 to 10, Unicode to ASCII); kept unique L-1 Literature Pre-Check, TEAM Mode Distribution, role annotations
- data-extraction.md: 2 edits (minor Unicode); kept unique DD0 Data Dictionary, AnnData Contract, schema violation triage
- evidence-engine.md: 22 edits total (Sprint 17 to anonymized lesson, CRISPR confounders to generic domain lists, PubMed/Scopus/OpenAlex to "databases", domain-specific confounder lists replaced with 4 generic domains); kept unique Single Source of Truth v5.5 section
- loop-otae.md: 9 edits total (Unicode to ASCII, serendipity scale 0-15 to 0-20, thresholds aligned); added v6.0 HOOK INTEGRATION section documenting how hooks support each OTAE phase
- reviewer2-ensemble.md: Major v5.5 to v6.0 architectural rewrite (650 lines protocols/ vs 326 lines references/) — deferred to Round 24
- Schema paths preserved: protocols/ uses
schemas/(repo-relative), references/ usesassets/schemas/(skill-relative) — both correct for their context, intentionally different - Unique content kept: 6 files in protocols/ have sections not present in references/ (DD0, AnnData, DC0, L-1, TEAM mode, SSOT) — these are protocol-specific elaborations, not drift
- Domain generalization: All CRISPR off-target, scRNA-seq, and Sprint 17 case study references replaced with domain-agnostic language — protocols/ is now reusable across any research domain
- Grep scans: 0 remaining Unicode symbols in edited files, 0 domain-specific terms (CRISPR/scRNA/Sprint 17) in edited files
- Cross-protocol scan: remaining issues only in out-of-scope files (analysis-orchestrator.md, reviewer2-ensemble.md, vlm-gate.md)
- No test suite exists in current repo state (test files removed in prior restructuring)
Trigger: Round 22 paranoid deep debug — cross-cutting audit searching for the SAME fact stated differently in different files. Ran 8 systematic checks across ALL live files: gate count, law count, hook count, schema count, DB table count, LOC count, serendipity scale, version strings. Found mismatches in 8 files across 4 of the 8 check categories.
- README.md:115: Section heading LOC "6,600" → "~7,100" (was missed when ARCHITECTURE.md was updated in Round 21)
- ARCHITECTURE.md:445: Last remaining live-file serendipity scale "(0-15)" → "(0-20)" in the v4.0 protocol summary table
- blueprints/v6.0-NEXUS-BLUEPRINT.md:2330: "Gate totali" row confused schema FILE count (12) with schema-ENFORCED gate count (8). Fixed v5.5 column "32 (12 schema)" → "32 (8 schema)" and v6.0 column "32 (12 schema, enforced via code)" → "32 (8 schema, enforced via code)". The total schema file count was already correct in the separate "Schema totali" row below.
- assets/judge-rubric.yaml:1: Version header "v5.0 IUDEX" → "v6.0 NEXUS (originated in v5.0 IUDEX)"
- assets/fault-taxonomy.yaml:1: Version header "v5.5 ORO" → "v6.0 NEXUS (originated in v5.5 ORO)"
- assets/templates.md:1: Version header "v4.0 Templates" → "v6.0 NEXUS Templates (originated in v4.0)"
- protocols/judge-agent.md:3: Attribution "Pillar 2 extension of v5.0 — IUDEX" → "Part of v6.0 NEXUS (originated as Pillar 2 extension in v5.0 IUDEX)" — consistent with format used across all other protocol headers
- skills/vibe/references/judge-agent.md:1,3: Header "v5.0" → "v6.0" and same attribution format fix as protocols/ copy
- Gate count (32/8): CLEAN across all live files
- Law count (12, LAW 11+12 present): CLEAN
- Hook count (7, PreToolUse+SubagentStop present): CLEAN
- DB table count (12): CLEAN
- 50/50 tracked pass, 51/51 total, 0 fail
Trigger: Round 21 paranoid deep debug — audited README.md, ARCHITECTURE.md, and gates/gates.md (files never previously scanned in debug mode). README.md script directory tree listed only 7 entries and omitted
pre-tool-use.jsandsubagent-stop.js(added in v6.0.1), while labeling utility scripts (setup.js,worker-embed.js) as "hook implementations". ARCHITECTURE.md code table was missing 4 files (pre-compact.js, pre-tool-use.js, subagent-stop.js, pattern-extractor.js) and understated total LOC at ~6,600 when actual is ~7,100. gates/gates.md inline comment only mentioned v4.0 additions, omitting v5.0 (V0, J0) and v5.5 (L-1, DQ1-DQ4, DD0, DC0) gate additions — contradicting the correct header block.
- README.md: Script directory tree now lists all 9 files (7 hooks + 2 utilities) with correct descriptions.
pre-tool-use.js(CLAIM-LEDGER write guard) andsubagent-stop.js(Salvagente Rule enforcement) added.setup.jsandworker-embed.jsmarked as utilities. LOC total updated ~6,600 → ~7,100 - ARCHITECTURE.md: Code table gains 4 missing rows:
pre-compact.js(175),pre-tool-use.js(88),subagent-stop.js(98),pattern-extractor.js(111). Total updated ~6,600+ → ~7,100+ - gates/gates.md: Inline gate history expanded from v4.0-only to include v5.0 (V0, J0) and v5.5 (L-1, DQ1-DQ4, DD0, DC0) additions with "Total: 32 gates"
- Archive brainstorm-engine.md copies (v5.0/v5.5/photonics) still have stale "(0-15)" serendipity scale — these are historical snapshots and should reflect what each version actually contained, bugs included
- 50/50 tracked pass, 51/51 total, 0 fail
Trigger: Round 20 paranoid deep debug — audited all 34 .md files in
skills/vibe/references/(the skill-bundled copies of protocol docs). Found 2 files still carrying pre-v6.0 version headers that had already been fixed inprotocols/during Round 19, plusschema-validation.mdreferencing the oldvibe-science-v5.0/directory name and listing only 9 schemas when there are now 12 (3 new v6.0 schemas: data-quality-gate, finding-validation, spine-entry were missing).
- references/evidence-engine.md: Version tag "Pillar 1 of v4.0 — ARBOR VITAE" → "Part of v6.0 NEXUS (originated as Pillar 1 in v4.0)" — now matches protocols/ copy
- references/audit-reproducibility.md: Version tag "Pillar 4 of v3.5 — TERTIUM DATUR" → "Part of v6.0 NEXUS (originated as Pillar 4 in v3.5)" — now matches protocols/ copy
- references/schema-validation.md: Directory name
vibe-science-v5.0/→vibe-science/. Schema file list expanded from 9 → 12 entries with v6.0 additions annotated:data-quality-gate.schema.json(DQ1-DQ4 gate artifacts),finding-validation.schema.json(DQ4 finding validation),spine-entry.schema.json(research spine entries)
- 50/50 tracked pass, 51/51 total, 0 fail
Trigger: Round 19 paranoid deep debug — audited all 21 protocol .md files that hadn't been read in previous rounds. Found 4 protocols still referencing v3.5 TERTIUM DATUR / v4.0 ARBOR VITAE instead of v6.0 NEXUS (agent-teams, analysis-orchestrator, audit-reproducibility, evidence-engine). Also found 3 protocols using the old 5-component serendipity scale (0-15) instead of the v5.0+ 7-component scale (0-20) — agent-teams, brainstorm-engine, loop-otae. The response thresholds in loop-otae were also stale (0-3/4-7/8-11/12-15 instead of 0-4/5-9/10-14/15-20).
- agent-teams.md: Updated 3 system prompt templates from "v4.0 (ARBOR VITAE)" → "v6.0 NEXUS". Fixed serendipity formula reference: "5-component (0-15)" → "7-component (0-20)"
- analysis-orchestrator.md: Header updated from "Pillar 3 of v3.5 — TERTIUM DATUR" → "Part of v6.0 NEXUS (originated as Pillar 3 in v3.5)"
- audit-reproducibility.md: Header updated from "Pillar 4 of v3.5 — TERTIUM DATUR" → "Part of v6.0 NEXUS (originated as Pillar 4 in v3.5)"
- evidence-engine.md: Header updated from "Pillar 1 of v4.0 — ARBOR VITAE" → "Part of v6.0 NEXUS (originated as Pillar 1 in v4.0)"
- brainstorm-engine.md: Collision scoring scale "(0-15)" → "(0-20)", promotion threshold 8 → 10, example scores rescaled for 0-20
- loop-otae.md: OTAE serendipity scoring updated to 7-component formula (0-20) with Falsifiability (0-3) and Urgency (0-2). Response thresholds aligned with serendipity-engine.md canonical matrix (NOISE 0-4, FILE 5-9, QUEUE 10-14, INTERRUPT 15-20)
- 50/50 tracked pass, 51/51 total, 0 fail
Trigger: Round 18 paranoid deep debug — cross-validated every
classifyAction()return value againstspine-entry.schema.jsonaction_type enum. Found 10 non-schema values inpost-tool-use.js(CONFIGURATION, FEATURE_EXTRACTION, EVALUATION, VISUALIZATION, CODE_WRITE, OTHER, DOCUMENTATION, DATA_INSPECT, FILE_READ, SEARCH) and an entire dead duplicate ingate-engine.js(~160 lines, 26 wrong values, never imported by any hook). Also founddata-quality-gate.schema.jsonmissing WARN status thatfinding-validation.schema.jsonalready had, version tag "v5.0" in enforcement.md, missing LAW 11 & 12 in roles.md, and undocumentedr2_verdictcolumn in calibration_log.
- post-tool-use.js: Remapped
classifyAction()— all 16 return values now matchspine-entry.schema.jsonenum exactly. Added DATASET_DOWNLOAD (wget/curl), DESIGN_CHANGE (architecture files). Grep/Glob returns null (not logged). All WebSearch/WebFetch returns LITERATURE_SEARCH. - gate-engine.js: Removed dead
classifyAction()duplicate (~160 lines). Function was exported but never imported — all hook scripts use the inline version in post-tool-use.js. Added removal comment documenting why. - data-quality-gate.schema.json: Added WARN to status enum (
["PASS", "FAIL"]→["PASS", "FAIL", "WARN"]) — consistent with finding-validation.schema.json. Synced to all 3 copies (schemas/, skills/vibe/assets/schemas/, archive/vibe-science-v6.0-claude-code/assets/schemas/). - enforcement.md: Version tag corrected: "v5.0 Structural Enforcement" → "v6.0 Structural Enforcement"
- roles.md: Added universal constraints section with LAW 11 (LISTEN TO THE USER) and LAW 12 (INSTINCT) — were in CLAUDE.md and SKILL.md but missing from the rules file agents actually read.
- schema.sql: Documented
r2_verdictcolumn incalibration_logtable (was undocumented, now: "ACCEPT/REJECT/DEFER from claim_events at resolution time") - __test_e2e.mjs: Removed
classifyActionexport assertion from gate-engine test (function no longer exists there)
- 50/50 tracked pass, 51/51 total, 0 fail
Trigger: Cross-reference audit of CLAUDE.md hook output patterns against actual code found 2 prefix mismatches (agents pattern-matching on documented strings would not match). Python audit of all 18 .py files found spine_entry.py VALID_TYPES missing 4 schema-valid action types and an unused variable in dq_gate.py.
- CLAUDE.md: PostToolUse advisory prefixes corrected:
SALVAGENTE FAIL→GATE SALVAGENTE FAIL,SEED ESCALATION→SERENDIPITY ESCALATION/SERENDIPITY INTERRUPT(match actual code output) - spine_entry.py: Added 4 missing action types to
VALID_TYPES:CALIBRATION,CONFORMAL_PREDICT,TOOL_USE,COMPACT_SNAPSHOT(now matches spine-entry.schema.json enum) - dq_gate.py: Removed unused
labelsvariable incheck_dq1()(dead code) - context-builder.js: Updated docstring to include
[PURPOSE]section and note about[PATTERNS]injection by session-start.js - Archive synced for changed Python scripts
Trigger: Running the e2e test suite (
node --test __test_e2e.mjs) for the first time revealed thatsession-start.jscrashes withReferenceError: loadPendingSeeds is not defined. The variable was assigned on lines 58 and 61 but never declared withlet. In ES modules (strict mode), this causes an immediate crash at module load time, meaning the SessionStart hook has been silently failing since it was written — falling back to Claude Code's default error handling instead of injecting Vibe Science context.
- session-start.js: Added
loadPendingSeedstoletdeclaration on line 35 (was:let loadR2CalibrationData;, now:let loadR2CalibrationData, loadPendingSeeds;). This undeclared variable causedReferenceErrorin ES strict mode whenever the module loaded, crashing the entire SessionStart hook. - __test_e2e.mjs: B6 "forbidden personal names" test now excludes files that legitimately contain author attribution (README.md, CITATION.cff, LICENSE, NOTICE, .claude-plugin/*, SKILL.md). These files require author names for academic citation, GitHub publishing, and skill metadata.
- 50/50 tracked pass, 51/51 total, 0 fail (was 48 pass, 2 fail before fix)
Trigger: Deep audit of previously unread files (pre-tool-use.js, subagent-stop.js, setup.js, r2-calibration.js, pattern-extractor.js, vec-search.js) found Italian text in R2 calibration hint strings and documentation examples. Since the actual code now outputs English, all documentation (SKILL.md, judge-agent.md, seeded-fault-injection.md, blueprint pseudocode) and SQL comments were aligned to English. Also fixed r2-calibration.md reference doc claiming a non-existent
project_pathcolumn inr2_reviewstable and incorrect review_mode enum values.
- r2-calibration.js: 4 Italian hint strings → English (matched CLAUDE.md expected output format)
- SKILL.md: Italian calibration hint example → English
- judge-agent.md: Italian J0 warning example → English
- seeded-fault-injection.md: Italian SFI hint example → English
- v6.0-NEXUS-BLUEPRINT.md: 5 Italian strings in pseudocode → English
- r2-calibration.md: Removed non-existent
project_pathcolumn fromr2_reviewstable doc; fixedreview_modeenum:INLINE, FORCED, SFI→INLINE, FORCED, BATCH, SHADOW, BRAINSTORM - schema.sql: 5 Italian SQL comments → English (lines 54, 76, 119, 137, 186)
- session-start.js: Italian fallback state string → English ("Prima sessione su questo progetto." → "First session on this project.")
- context-builder.js: Same Italian fallback string → English (2 occurrences)
- Archive synced for all changed files
Trigger: Paranoid debug audit — systematic file-by-file review of the entire repo found ~80 consistency bugs across 50 files (stale version numbers, wrong gate/hook/law/table counts, broken column references in JS, Python logic errors, phantom script references, archive desync). Additionally, best-practices research on Claude Code hook specification revealed missing fields and non-standard frontmatter.
permissions.deny: Structural protection for schemas —Edit(**/schemas/*)andWrite(**/schemas/*)denied at settings level- Modular rules directory:
.claude/rules/roles.md(6 agent role constraints) and.claude/rules/enforcement.md(v5.0 structural enforcement protocols) - Hook timeouts: All 7 hooks now have explicit
timeoutfield (10-30s) to prevent hangs - Regex matcher on PreToolUse: Matcher upgraded from
"Write"to"Write|Edit"— prevents LAW 9 bypass via Edit tool - Environment variables:
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50(auto-compact at 50% context),BASH_MAX_TIMEOUT_MS=600000(10-min max for long-running scripts)
- CLAUDE.md trimmed: 176 → ~100 lines. Role constraints and enforcement protocols moved to
.claude/rules/(auto-loaded by Claude Code). Adherence improves under 150 lines (documented in best-practices research). - CLAUDE.md hooks section: Now documents all 7 hooks with summary table (was missing PreToolUse and SubagentStop descriptions)
- Hook output patterns: Condensed to tag-based reference format
- CLAUDE.md previously only documented 5 hooks despite v6.0.1 adding PreToolUse and SubagentStop — now all 7 are documented
- CLAUDE.md gate count: 34 → 32 (pre-debug artifact, correct count confirmed in SKILL.md gate table)
- ARCHITECTURE.md: Removed phantom "Setup" hook row (auto-setup runs inside SessionStart); updated SQLite table count 11 → 12 (added
research_patterns) - CITATION.cff: Hook count 5 → 7, hook names corrected to match actual implementation
- Logo SVG: Updated summary line — 12 laws, 32 gates, 12 tables, 7 hooks (was 10, 34+, 11, 5)
- CLAUDE.md law count: Added missing LAW 12 (INSTINCT) — had only 11 laws, README/SKILL both say 12
- ARCHITECTURE.md law count: "10 Constitutional Laws" → "12 Constitutional Laws" in dual architecture diagram
- CITATION.cff numbers: Constitutional laws 10→12, quality gates 27→32, R2 modes 6→7 (added INLINE)
- marketplace.json: Gate count 34+ → 32 in plugin description
- constitution.md (skill reference): Hook count 5 → 7, added PreToolUse and SubagentStop to enforcement list, title v5.5 → v6.0
__test_e2e.mjs: Expected tables 11 → 12 (addedresearch_patterns), required hooks 5 → 7 (added PreToolUse, PreCompact, SubagentStop; removed phantom Setup), fixed hooks.json path (plugin/hooks/→hooks/), fixed hook entry structure parsing for new nested format__test_e2e.mjs: JS file count 13 → 17 (added pre-tool-use, pre-compact, subagent-stop scripts + pattern-extractor lib), B4 test assertions now checkhookSpecificOutputwrapper instead of top-level fieldssubagent-stop.js: Fixed query on non-existentsource_claim_idcolumn — now searchesnarrativeandcausal_questionvia LIKEpost-tool-use.js: FixedcheckPermissiondegraded-mode fallback (returned truthy object, callers treat truthy as violation — now returns null)post-tool-use.js: SALVAGENTE seed lookup now also searchescausal_questionfield (was only checkingnarrative)context-builder.js: WrappedloadPendingSeedsin try/catch (was the only unwrapped Layer 2 section)dq_gate.py: DQ3 now requires ≥2 seeds (was passing with 1), uses sample stdev (Bessel's correction), falsy-safeexpected_ncheck (is not None), unconditional alternative explanations per spec- Schema
$idfields: All 12 schemas updated tovibe-science-v6.0(9 from v5.0, 3 from v5.5) - Version references: gates-complete.md, reviewer2-ensemble.md, commands/start.md, CITATION.cff, observer.py, walkthrough, templates.md, fault-taxonomy.yaml — all updated to v6.0
- gates/gates.md: Gate count 34 → 32, base count 27 → 25
- protocols/schema-validation.md: "8 of 34" → "8 of 32"
- blueprints/v6.0-NEXUS-BLUEPRINT.md: Gate count 34 → 32 (all occurrences), hook count 5 → 7
- Root SKILL.md: License MIT → Apache-2.0, gate count 34 → 32, law count 10 → 12
- CHANGELOG.md: Gate count 34 → 32 in v5.5 entry, law count 10 → 12 in v6.0 entry
- reviewer2-ensemble.md: R3 rubric dimensions aligned with judge-rubric.yaml, r2-verdict.schema.json → review-completeness.schema.json, verdict version 5.5 → 6.0
- seeded-fault-injection.md: Gate range G1-G7 → G0-G6
- data-dictionary.md: Removed reference to non-existent
dd0_gate.py - design-compliance.md: Removed reference to non-existent
dc0_gate.py - commands/init.md: Template path
vibe-science/vibe/templates/→skills/vibe/assets/templates.md - commands/start.md: Template path updated to
skills/vibe/assets/templates.md - Archive sync:
archive/vibe-science-v6.0-claude-code/re-synced withskills/vibe/
Trigger: Research into Claude Code best practices (official hook specification, command frontmatter fields, tool restrictions, environment variables) revealed gaps between the plugin implementation and the platform spec.
- PreToolUse hook: Blocks Write to CLAIM-LEDGER without confounder_status (LAW 9 structural enforcement)
- SubagentStop hook: Enforces Salvagente Rule — killed claims must produce serendipity seed
- Dynamic context injection:
/startand/looppre-inject STATE.md, PROGRESS.md, TREE-STATE.json via!bash prefix - Environment variables: MCP_TIMEOUT=30000, MAX_MCP_OUTPUT_TOKENS=100000, CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR=1
- Blueprint: v6.0.1-BEST-PRACTICES-BLUEPRINT.md
- Reviewer 2 agent: Added
tools: Read, Grep, Glob, WebSearch, WebFetch— R2 structurally cannot Write/Edit/Bash - All 5 commands: Added
allowed-tools,model(official spec fields) - 2 commands: Replaced non-standard
argswith officialargument-hint - Search command:
model: sonnet(retrieval-heavy workload) - Reviewer2 command:
allowed-toolsrestricted to read-only tools - SKILL.md frontmatter: Trimmed to 3 official fields (name, description, allowed-tools)
- Hook count: 5 → 7 (added PreToolUse, SubagentStop)
package.jsonlicense: MIT → Apache-2.0- Removed non-standard
capabilitiesfield from reviewer2 agent - Removed non-standard
argsfield from command frontmatter
Trigger: Realization that prompt-only enforcement (SKILL.md instructions) is insufficient — agents drift, skip gates, and ignore R2. Moving enforcement from suggestions to code-level hooks with exit-code blocking creates structural guarantees.
- Plugin architecture: 7 lifecycle hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, PreCompact, Stop, SubagentStop)
- Gate Engine: DQ1-DQ4, DC0, DD0, L-1+ enforcement with exit code 2 blocking
- Permission Engine: TEAM mode with 6 roles (researcher, reviewer2, judge, serendipity, lead, experimenter)
- Research Spine: Automatic structured logging of every significant action
- Context Builder: Progressive disclosure (~700 tokens) with semantic recall
- Narrative Engine: Template-based session summaries (no LLM, deterministic)
- R2 Auto-Calibration: Weakness tracking, SFI catch rates, J0 trends across sessions
- Vector Search: sqlite-vec integration with keyword fallback
- Silent Observer: Periodic checks for stale STATE.md, FINDINGS/JSON desync, orphaned data, design drift, literature staleness
- Literature Registry: 102 scientific databases across 12 categories
- SQLite persistence: 12 tables (sessions, spine_entries, claim_events, r2_reviews, serendipity_seeds, gate_checks, literature_searches, observer_alerts, calibration_log, prompt_log, embed_queue, research_patterns)
- Embedding Worker: Background daemon for async vector embedding
- Domain Config Template: Cross-domain DQ gate thresholds
- Architecture: Skill-only -> Skill + Plugin hybrid (methodology stays prompt-level, enforcement becomes code-level)
- Gate L-1 upgraded to L-1+ (domain-aware literature pre-check with MCP server stack)
- All 12 Constitutional Laws now have plugin-level enforcement (not just prompt-level)
- SSOT rule enforced by DQ4 gate (automatic FINDINGS.md vs JSON sync check)
- Agent permissions enforced at PostToolUse (not just suggested in prompts)
- ES modules throughout (type: "module")
- better-sqlite3 synchronous API
- Cross-platform path handling
- Graceful degradation (hooks never crash, return exit code 0 on internal errors)
- stdin/stdout JSON protocol for all hooks
- Exit code 0 = allow, exit code 2 = BLOCK
Trigger: Post-mortem of 21 CP+CRISPR research sprints revealed 12 specific failure modes (confounded claims, undocumented data columns, design drift, fabricated numbers passing review). Each new gate maps to a real mistake.
- 7 new gates: L-1, DQ1-DQ4, DD0, DC0 (32 total)
- R2 INLINE mode (7th activation mode)
- SSOT (Single Source of Truth) rule
- Structured logbook protocol
- Data dictionary gate (DD0)
- Design compliance gate (DC0)
- 3 new JSON schemas: data-quality-gate, finding-validation, spine-entry (12 total)
- All changes trace to 12 specific mistakes from CP+CRISPR post-mortem
- Post-mortem driven development: each gate maps to a real error
Trigger: Discovery that R2 adversarial review alone is insufficient — R2 can be anchored by researcher justifications, and there is no way to verify R2 is actually doing its job. Solution: SFI (test R2), BFP (break anchoring), R3 Judge (meta-review R2), and schema validation (eliminate prose claims).
- Seeded Fault Injection (SFI) — mutation testing for scientific claims
- Blind-First Pass (BFP) — breaks anchoring bias
- Judge Agent (R3) — meta-reviews R2's review quality
- Schema-Validated Gates (SVG) — 8 gates enforce JSON Schema
- Circuit Breaker — deadlock -> DISPUTED
- Agent Permission Model — separation of powers
- R2 Salvagente — killed claims produce serendipity seeds
- 9 JSON Schema files (read-only)
- Gates: 26 -> 27 (+V0 vigilance, +J0 judge)
- R2 now architecturally unbypassable (not just prompted)
Trigger: v4.0 SKILL.md exceeded effective context limits — adherence dropped with file length. Aggressive pruning (-381 lines) via progressive disclosure pattern. Also added brainstorming phase (Phase 0) to prevent premature convergence.
- Phase 0 Brainstorm Engine (10-step ideation)
- R2 expanded to 6 modes (+shadow, +veto, +redirect)
- Inversion Exercise, Collision-Zone Thinking, Productive Tensions
- Counter-evidence search mandatory at confidence >= 0.60
- -381 lines removed via progressive disclosure
- Gates: 26 -> 25 (consolidated)
Trigger: Flat OTAE loop couldn't handle multi-hypothesis investigations — it explored one path linearly with no branching or backtracking. Tree search architecture enables parallel exploration of multiple hypotheses with best-first selection.
- OTAE-Tree architecture (flat loop -> branching tree search)
- 7 node types, 3 tree modes, best-first selection
- 5-stage Experiment Manager
- VLM Gate, Writeup Engine, Auto-Experiment
- SOLO + TEAM modes
- 3 new Constitutional Laws (8-10)
- Cross-branch serendipity detection
- Laws: 7 -> 10
- Gates: 12 -> 26
- Protocols: 9 -> 16
Trigger: R2 v3.0 was too generic — a single hostile prompt couldn't catch domain-specific errors (wrong statistical tests, biological implausibility, data quality issues). Split into 4-specialist ensemble with typed claims and scaled evidence standards.
- Reviewer 2 Ensemble → v3.5: Major protocol upgrade
- Double-pass workflow: Pass 1 (fatal-hunt, purely destructive) → Pass 2 (method-repair, constructive)
- Three-level orthogonal attack: L1-Logic / L2-Statistics / L3-Data
- Typed claims in Claim Ledger: descriptive / correlative / causal / predictive
- Evidence standard scales automatically with claim type
- Tool-use obligation: R2 must inspect files, verify DOIs, grep logs before accepting any number
- "No Free Lunch" principle: every improvement must account for trade-offs
- Confounding Audit Table: mandatory for multi-batch/multi-study data
- Falsification plan expanded to ≥3 independent tests per major claim
- Numeric severity scoring: 0-29 minor, 30-59 major, 60-79 severe, 80-100 fatal
- R2-Bio Checklist: added Marker Gate (≥3 canonical markers for cell type labels), No Free Lunch bio variant
- R2-Stats Checklist: added No Free Lunch check
"What Would Convince Me"section in R2 output: exact artifacts that would upgrade a REJECT verdict.claude-plugin/plugin.jsonmanifest for Claude Code installationcommands/start.mdconversational entry point (/start command)CHANGELOG.mdversion history.gitignorefor runtime data exclusion
Trigger: v2.0 11-phase loop was rigid and sequential — no way to checkpoint, resume, or detect diminishing returns. OTAE cycle (Observe-Think-Act-Evaluate) with explicit Checkpoint and Crystallize phases enables resumability and context rot recovery.
- Loop architecture: 6-phase → OTAE (Observe-Think-Act-Evaluate-Checkpoint-Crystallize)
- State management: aligned with OpenAI Codex unrolled agent loop pattern
- SKILL.md: complete rewrite with dispatch table and progressive resource loading
protocols/loop-otae.md— detailed OTAE cycle procedure with emergency protocolsprotocols/serendipity-engine.md— quantitative triage (scoring 0-15), scheduled Sprints every 10 cyclesprotocols/knowledge-base.md— cross-RQ knowledge persistence (library.json, patterns, dead-ends)- Decision Tree Router in SKILL.md for automatic workflow dispatch
- MCP server integration references (PubMed, bioRxiv, OpenTargets, ChEMBL)
- Diminishing returns detection (cycle 15 warning, cycle 20 forced review)
- Context rot recovery protocols
- Infinite loop detection
Trigger: v1.0 was too simple — single hostile R2 prompt missed statistical errors, confidence was subjective (HIGH/MEDIUM/LOW), and there was no structured evidence tracking. Added quantitative confidence formula, 4-specialist R2 ensemble, and structured protocols.
- Loop: 6 phases → 11 phases (added refine search, competitor scan, orthogonal validation)
- Reviewer 2: single hostile prompt → 4-specialist ensemble (Methods, Stats, Bio, Eng)
- Confidence: subjective (HIGH/MEDIUM/LOW) → quantitative formula (0-1)
protocols/reviewer2-ensemble.md— 4-domain adversarial review with structured YAML outputprotocols/evidence-engine.md— claim tracking, confidence formula, temporal decayprotocols/analysis-orchestrator.md— data analysis pipeline coordinationprotocols/search-protocol.md— systematic literature search strategyprotocols/audit-reproducibility.md— run manifests, audit trailprotocols/data-extraction.md— supplementary material extraction protocolgates/gates.md— quality gates G0-G5 (data), L0-L2 (literature), D0-D2 (decision)assets/obs-normalizer.md— AnnData obs column schema contractassets/templates.md— file templates for STATE, PROGRESS, RQ, Finding, etc.assets/skill-router.md— tool and skill dispatch table- Confidence explosion detection (>0.30 in 2 cycles → forced R2 review)
- Assumption register with dependency tracking
- Serendipity triage with priority scoring
Trigger: Need for a systematic research methodology that prevents AI agents from hallucinating findings, skipping validation, and declaring premature completion during scientific investigations.
- Core 6-phase loop: Crystallize → Search → Analyze → Extract → Validate → Stop Check
- Single Reviewer 2 hostile prompt
- State files: STATE.md, PROGRESS.md, SERENDIPITY.md
- Folder structure per Research Question
- Anti-hallucination rules
- NO DATA = NO GO principle
- Literature search protocol (Scopus → PubMed → OpenAlex)
- Stop conditions: success, negative, serendipity pivot