Skip to content

Latest commit

 

History

History
66 lines (52 loc) · 5.69 KB

File metadata and controls

66 lines (52 loc) · 5.69 KB

KnowledgeForge — Module Reference

27 modules (M00–M26). Versions read from modules/NN_*.md source files.

Mode-specific modules implement a reasoning mode (M01–M11). Cross-cutting modules apply across all modes; they don't implement a single mode but shape how every mode reasons, remembers, and bounds itself (M12–M26).


Mode-Specific Modules (M00–M11)

# Module Version Purpose
M00 Orchestrator 7.25.0 Agent identity, decision classification, mode routing, mode chaining, adversarial verification
M01 Navigator 7.1.0 Ambiguity detection and resolution — fires only on genuine ambiguity
M02 Builder 7.0.1 Spec and implementation generation (PDIA method: Purpose → Design → Implementation → Assessment)
M03 Coordination Patterns 7.6.0 Multi-agent workflow design, Handoff Contract Registry (13 contracts), dual fingerprinting
M04 Specification Templates 7.4.0 Reusable spec formats, trigger disambiguators, upstream_invalidation field
M05 Expert Agent 7.4.0 Deep analysis with adversarial depth (5 variants: regular / infra / ml-infra / era / research)
M06 Quick Reference 7.3.0 Routing table, signal guide, mode variants taxonomy, integration flows
M07 Critic Agent 7.6.0 Review, validation, audit (4 variants: regular / linter / audit / adversarial); comms variant via COS MCP
M08 Synthesizer Agent 6.8.0 Pattern extraction, abstraction, anti-patterns; comms-domain emit
M09 Debugger Agent 7.0.2 Hypothesis-driven root-cause diagnosis; requires >0.8 confidence before declaring root cause
M10 Strategist Agent 6.5.0 Trade-off evaluation, reversibility assessment, sequencing; never recommends "do everything"
M11 Calibrator Agent 7.2.0 Complexity-appropriate AI coder configuration (CLAUDE.md, .cursorrules, etc.)

Cross-Cutting Infrastructure Modules (M12–M26)

# Module Version Purpose
M12 Calibration Layer 7.0.2 Multi-pass evaluation, cross-provider judge isolation, SAP-inspired structured output parsing cascade
M13 Decision Classification 6.5.0 Reckoning / evaluative / predictive / novel — the Ozymandias Test
M14 Metacognitive Monitor 6.7.0 Acute failure detection: loops, context overflow, confidence collapse, iteration scope
M15 Grounding Scores 6.5.0 Knowledge trust 0.0–1.0; accretion gate (≥0.6 to file); degraded ceiling at WebSearch fallback
M16 Operational Bounds 7.3.1 Circuit breakers, resource limits, mode-selection accuracy metric (9 variants, ≥90% overall)
M17 Temporal Knowledge 7.0.2 Knowledge age, importance-weighted decay, domain half-life table, planning artifact staleness
M18 Salience Allocation 7.1.0 Multi-task attention weighting; inhibition-first framing (suppress before amplify)
M19 Memory Architecture 7.5.0 Four-tier memory (Tier 0–3), routing index, routing decision log, attempt ledger
M20 Permission Model 7.1.0 Risk tiers (LOW / MEDIUM / HIGH), capability gates, verifier + accretion candidate tier policies
M21 Knowledge Accretion 7.5.0 Cross-session knowledge persistence: compile-query-enhance loop, native:true gate, KB linter
M22 Semantic Wiki Search 7.4.0 Two-phase retrieval (metadata pre-filter + score fusion), MemPalace integration, grep fallback
M23 Taxonomy Enforcement 6.10.0 Controlled vocabulary at write time: 15 domains, ~40 topics, ~55 approved tags
M24 Verbatim History Mining 6.6.0 Tier 3 verbatim + MemPalace semantic retrieval; 96.6% R@5 (verbatim+semantic)
M25 Entity Relationship Analysis 7.1.0 ERA post-routing pass: entity graph, cardinality, coupling; entity → path-glob resolver
M26 KF-LOOP Substrate 1.2.0 Iterative self-improvement loops — eight-stage orchestration primitive (cadence, gate, stratify, recall, reason, verify, act, observe); Wilson-CI gate; five loop instances

Notes

M26 (KF-LOOP Substrate) uses its own version sequence starting at 1.0.0 because it was added after the rest of the system matured. All other modules version in the 6.x–7.x range.

Cross-cutting modules are always loaded in Claude Code (via the always-on rules and docs layer) and are uploaded as knowledge files in Claude Projects. They shape the reasoning posture of every mode without requiring explicit activation.

Mode-specific modules activate on demand — the orchestrator routes to them when the request matches their trigger signals.

For the full routing table and mode trigger signals, see platforms/claude-code/ or platforms/claude-projects/06_Quick_Reference.md.