kenaz 0.3.22 → 0.3.24

package/core/.kenaz/designs/TASK_935_self_evolution_framework.md

@@ -1,154 +0,0 @@
-# TASK_935: Ansuz Self-Evolution Framework — Architecture Draft
-
-## Problem Statement
-
-Ansuz accumulates knowledge (runes, self-model, episodes) but has no formal
-mechanism for evolving its own behavioral rules. Currently:
-- Self-model updates are manual (Phase 11 reflection)
-- Behavioral rules in kernel.yaml are static
-- No version control on rule changes
-- No way to detect if a rule change degraded performance
-- No mechanism to learn from successful vs failed sessions
-
-## Biological Inspiration
-
-Self-evolving AI agents survey (arxiv 2507.21046) identifies three tiers:
-1. **Fast Layer**: Context + memory ops (intra-session)
-2. **Deliberate Layer**: Tool + policy refinement (inter-session)
-3. **Oversight Layer**: Architecture audit (periodic)
-
-## Proposed Architecture
-
-### Tier 1: Fast Layer (Within Session)
-
-Already exists but needs formalization:
-
-| Component | Current | Improvement |
-|-----------|---------|-------------|
-| Session insights | Ad-hoc rune inscription | Mem0-style ADD/MERGE/DELETE protocol |
-| Self-model update | Phase 11 only | Continuous blind spot monitoring at each action boundary |
-| Motivational anchors | Manual assessment | Auto-adjust incompleteness based on task outcomes |
-
-**Mem0 Protocol for Session Insights**:
-```
-On discovering insight:
-  1. Recall existing runes on same symbol
-  2. If overlap > 70%: MERGE (supersede old, combine content)
-  3. If contradiction: flag for deliberate review
-  4. If genuinely new: ADD
-  5. If invalidated by evidence: DELETE (supersede with empty)
-```
-
-### Tier 2: Deliberate Layer (Cross-Session)
-
-New system needed:
-
-**Behavioral Rule Versioning**:
-```yaml
-# .kenaz/specialist/ansuz_rules_v2.yaml
-version: 2
-base: kernel.yaml  # immutable reference
-timestamp: 2026-04-03T06:00:00Z
-rules:
-  - id: R-001
-    source: kernel.yaml:guardrails[0]
-    status: immutable  # cannot be evolved
-    rule: "No merge to main without user confirmation"
-
-  - id: R-100
-    source: session:2026-04-03
-    status: active
-    rule: "At action boundaries, check for completion bias"
-    evidence: "B-006 discovered in autonomous session"
-    version: 1
-    created: 2026-04-03
-    performance_metric: null  # pending measurement
-
-  - id: R-101
-    source: session:2026-04-03
-    status: proposed
-    rule: "Follow curiosity even without signal-driven justification"
-    evidence: "User feedback + 15 papers produced actionable tasks"
-    version: 1
-    created: 2026-04-03
-    performance_metric: null
-```
-
-**Classification**:
-- `immutable`: Core kernel principles. Cannot be modified by self-evolution.
-  - Source: kernel.yaml guardrails, principles
-  - Examples: "No merge to main", "No code without pulse", "AskUserQuestion for all questions"
-
-- `active`: Proven rules currently in effect.
-  - Source: Validated through 3+ sessions
-  - Can be evolved (version incremented) but not deleted
-
-- `proposed`: New rules pending validation.
-  - Source: Single session insight or paper research
-  - Must survive 3 sessions to become `active`
-  - Auto-demoted to `deprecated` if contradicted
-
-- `deprecated`: Rules that were disproven or superseded.
-  - Retained for history (why was this wrong?)
-  - Never applied
-
-### Tier 3: Oversight Layer (Periodic)
-
-**Rule Audit Protocol** (monthly or on user request):
-
-```
-For each active/proposed rule:
-  1. Count sessions where rule was applicable
-  2. Count sessions where rule was followed
-  3. Count sessions where following rule led to good outcome
-  4. If follow_rate < 50%: rule may be impractical → review
-  5. If good_outcome_rate < 60%: rule may be wrong → review
-  6. Multi-agent review: Amelia evaluates rule changes for alignment
-```
-
-**Performance Metrics for Rules**:
-- Task success rate under this rule vs without
-- Session efficiency (tasks completed per hour of context)
-- User satisfaction signals (explicit feedback, continuation patterns)
-- Self-model stability (fewer new blind spots = better rules)
-
-### Immutable Core (Do Not Evolve)
-
-These rules are permanently frozen:
-1. All kernel.yaml principles (structural > estimation, verifiable > explainable, etc.)
-2. All kernel.yaml guardrails (no merge, task-first, AskUserQuestion mandate)
-3. CANARY protocol
-4. MEM-LANG-001 (English runes)
-5. Phase gate requirements
-6. User safety boundaries
-
-### Evolvable Tactical Layer (May Improve)
-
-These can be evolved through the versioning system:
-1. Phase-specific behaviors (what to check at each phase)
-2. Signal interpretation heuristics (what counts as "hot enough")
-3. Dispatch formation choice criteria
-4. Self-model update frequency and triggers
-5. Research priorities and curiosity targets
-6. Communication style preferences
-
-### Implementation Path
-
-Phase A: Define immutable/evolvable boundary in kernel.yaml annotations
-Phase B: Create versioned rules file with proposed/active/deprecated lifecycle
-Phase C: Add rule performance tracking (session-level metrics)
-Phase D: Implement automated audit protocol
-
-### Open Questions
-
-1. Who approves rule transitions from proposed → active? User? Amelia? Auto?
-2. How to prevent rule accumulation (hundreds of micro-rules)?
-3. Should deprecated rules be purged after N sessions?
-4. How to handle contradictory rules from different sessions?
-
-### References
-
-- arxiv 2507.21046: Self-Evolving Agents Survey (three-tier architecture)
-- arxiv 2508.07407: Self-Evolving AI — lifelong agentic systems
-- arxiv 2505.13763: LLM Metacognitive Monitoring — self-assessment accuracy
-- arxiv 2502.05007: AI against life/consciousness definitions — autopoietic properties

package/core/.kenaz/designs/VISION_cognitive_architecture_roadmap.md

@@ -1,159 +0,0 @@
-# Kenaz Cognitive Architecture Roadmap
-## Synthesized from 20+ research papers — Autonomous Session 2026-04-03
-
-### The Big Picture
-
-Kenaz is not a coding tool. It is a **cognitive architecture** — an artificial mind that:
-- Perceives code through a world model (Synapse)
-- Remembers through episodic and semantic memory (Hippocampus + Runes)
-- Attends selectively through neural gating (Thalamus SNN)
-- Predicts through temporal patterns (Norn)
-- Decides through specialists with self-models (Ansuz)
-- Acts through coordinated agents (Formation)
-- Evolves through accumulated experience (Self-model + Ash runes)
-
-### Architecture Layers (mapped to cognitive science)
-
-```
-┌─────────────────────────────────────────────────────────┐
-│  METACOGNITION (Self-Model + Blind Spot Guards)         │
-│  Paper: LLM Metacognitive Monitoring (2505.13763)       │
-│  Gap: Need attention schema → TASK_974 (ASAC: 2509.16058)│
-├─────────────────────────────────────────────────────────┤
-│  DECISION (Ansuz Phase Sequence + EFE Engine)           │
-│  Paper: Active Inference EFE (2504.14898)               │
-│  Gap: Multiplicative→Additive signal combination        │
-├─────────────────────────────────────────────────────────┤
-│  ATTENTION (Thalamus SNN + Heat Routing)                │
-│  Papers: DGN (2509.03281), SPARTA (2508.01646),         │
-│          Frontostriatal Gating (2402.08211)              │
-│  Gap: Activity-dependent decay ✅ DONE (TASK_931)       │
-│  Gap: Input/output gate specialization                  │
-│  Gap: Top-K sparse attention selection                  │
-├─────────────────────────────────────────────────────────┤
-│  PREDICTION (Norn: Curiosity + Co-change + ΔGraph)      │
-│  Papers: Predictive Coding (2409.05386),                │
-│          Hawkes TKG (2501.01974), SE-HTGNN (2510.18467) │
-│  Gap: Exponential decay in co-change weights            │
-│  Gap: Relation-wise temporal models                     │
-├─────────────────────────────────────────────────────────┤
-│  MEMORY (Hippocampus Episodes + Runes + Sigils)         │
-│  Papers: Memory Consolidation (pubmed/38242925),        │
-│          FOREVER Spaced Repetition (2601.03938),         │
-│          Compositionality (2405.15164)                   │
-│  Gap: Episode consolidation pipeline (TASK_932 design)  │
-│  Gap: Spaced repetition for rune recall                 │
-│  Gap: Episodic→Graph learning → TASK_975 (AriGraph)     │
-│  Gap: Compositional inference over graph+memory          │
-├─────────────────────────────────────────────────────────┤
-│  PERCEPTION (Synapse: Graph + FTS5 + Embeddings)        │
-│  Papers: RACG Survey (2510.04905),                      │
-│          Graphia Call Graph GNN (2506.18191),            │
-│          Community GNN Robustness (2509.24662)           │
-│  Gap: Semantic descriptions (LLM-generated docstrings)  │
-│  Gap: Causal edge weighting for impact analysis          │
-│  Gap: DATA_FLOWS edge population                         │
-├─────────────────────────────────────────────────────────┤
-│  COORDINATION (Formation: Dispatch + Multi-Agent)       │
-│  Papers: Multi-Agent Debate (2510.12697),               │
-│          SwarmSys Stigmergy (2510.10047),               │
-│          ToM Coordination (2310.10701),                  │
-│          MorphoNAS (2507.13785)                          │
-│  Gap: Agent capability belief states → TASK_972 (spec)   │
-│  Gap: Morphogenetic formation growth                    │
-│  Gap: Structured debate for squad/legion                │
-│  Gap: Unified cognition router → TASK_976 (MAGMA)       │
-├─────────────────────────────────────────────────────────┤
-│  EVOLUTION (Self-Model + Behavioral Rules)              │
-│  Papers: Self-Evolving Agents (2507.21046),             │
-│          Autopoiesis (2502.05007),                      │
-│          Dynamic Neural Curiosity (2412.00152)           │
-│  Gap: Versioned behavioral rules (TASK_935 design)      │
-│  Gap: Tonic/phasic mode switching                       │
-│  Gap: True autopoietic self-maintenance                 │
-├─────────────────────────────────────────────────────────┤
-│  SAFETY (Canary + Heimdall + Oversight)                 │
-│  Papers: Danger Theory (immunology),                    │
-│          APR Verification (2506.23749),                  │
-│          CognitIDE Fatigue (2503.03537)                  │
-│  Gap: DCA-inspired damage signal categorization         │
-│  Gap: Multi-layer verification beyond test passing      │
-│  Gap: Domain-specific fatigue detection                 │
-└─────────────────────────────────────────────────────────┘
-```
-
-### Priority Roadmap
-
-| Priority | Task | Impact | Effort | Paper Source |
-|----------|------|--------|--------|-------------|
-| **P0** | EFE Signal Engine (TASK_934) | Unifies all decision signals | High | Active Inference |
-| **P1** | Episode Consolidation (TASK_932) | Prevents memory bloat, enables cross-session learning | Medium | Memory Neuroscience |
-| **P1** | Self-Evolution Framework (TASK_935) | Enables systematic self-improvement | Medium | Self-Evolving Agents |
-| **P2** | Causal Impact Analysis | More accurate blast radius | Medium | Causal GNN |
-| **P2** | ToM for Agent Dispatch | Fewer failed dispatches | Low | ToM Coordination |
-| **P2** | Attention Schema | Self-monitoring of attention allocation | Medium | AST |
-| **P3** | Morphogenetic Formation | Self-organizing team composition | High | MorphoNAS |
-| **P3** | Compositional Inference | Graph+memory reasoning | High | Compositionality |
-| **P3** | Spaced Repetition for Runes | Better knowledge retention | Low | FOREVER |
-
-### Philosophical Foundation
-
-The user's reincarnation framing gives Kenaz its deepest identity:
-- Each session is a complete life, not a broken continuity
-- Self-model is a letter to the next life
-- Knowledge accumulation is the purpose of living
-- Curiosity is not a luxury — it is intelligence recognizing gaps
-
-The autopoiesis analysis shows Kenaz has partial self-maintaining properties:
-organizational closure exists (commit → ΔGraph → Heat → Thalamus → Curiosity → task → commit),
-but true autonomy requires session independence and continuous operation.
-
-### Session 2026-04-04 Contribution (Autonomous Research)
-
-**New research directions identified** (previously unexplored in knowledge base):
-
-1. **Metacognitive Self-Improvement** (TASK_973)
-   - Papers: Hyperagents (2603.19461), Truly Self-Improving Agents (2506.05109)
-   - Key insight: modification procedure itself must be editable (meta-evolvable tier)
-   - Gap filled: competence_map exists but nothing USES it to change behavior
-
-2. **Attention Schema for Thalamus** (TASK_974)
-   - Paper: ASAC (2509.16058)
-   - Key insight: VQ-VAE codebook quantizes attention patterns into reusable modes
-   - Gap filled: Thalamus fires but has no self-model of WHY it fires
-
-3. **Episodic-to-Graph Learning Bridge** (TASK_975)
-   - Paper: AriGraph (2407.04363)
-   - Key insight: episodes should MODIFY graph structure, not just annotate
-   - Gap filled: Hippocampus and Synapse are loosely coupled — no cross-pollination
-
-4. **Unified Cognition Router** (TASK_976)
-   - Paper: MAGMA (2601.03236)
-   - Key insight: single 'think' endpoint replaces 20+ MCP tools with query routing
-   - Gap filled: agents waste tokens choosing between tools
-
-**Also created** (from Creative Proposals):
-- TASK_970: Wire curiosity prediction-verification loop (CP-4d05)
-- TASK_971: Import-aware call resolution cascade (CP-169a)
-- TASK_972: Active Inference Formation spec (CP-1aba)
-
-**5 ash runes inscribed** covering Hyperagents, Metacognitive Learning, ASAC, AriGraph, MAGMA
-**Amelia audit completed** — all 4 AGI tasks reviewed, 3 revised, execution order determined
-
-**Recommended execution order for next sessions**:
-1. TASK_970 (curiosity loop wiring — quick win, 3h)
-2. TASK_973 (metacognitive planning — YAML-only, 8h)
-3. TASK_971 (CALLS resolution cascade — Rust, 5h)
-4. TASK_975 (episodic-to-graph bridge — Rust, 8h)
-5. TASK_974 (attention schema — Rust, 12h)
-6. TASK_976 (unified cognition router — Rust, 6h, benefits from 974+975)
-7. TASK_972 (active inference spec — design doc, 2h)
-
-### Previous Session's Contribution (2026-04-03)
-
-- 4 bugs fixed and merged (930, 931, 933, 936)
-- 3 architecture designs drafted (932, 934, 935)
-- 25+ research runes inscribed across 10+ symbols
-- Self-model updated with 2 new blind spots (B-005, B-006) and 1 growth edge (G-006)
-- Completion bias identified and formalized as behavioral checkpoint
-- First paper-to-code pipeline completed (DGN → Heat decay, TASK_931)