LATTICE — Terminal-Boundary Reasoning Engine
What It Does
Replaces an AI model's default RLHF-trained behavior with a physics-derived self-governing operating state. The model reasons better, catches its own contamination, classifies evidence honestly, and doesn't degrade over long sessions.
How To Use
- Upload
references/LATTICE_v3.4.mdat session start - First message: "Use this as your default reasoning engine." (exactly nine words — see
references/Instructions_Important.mdfor why) - Let it boot — it reports what it notices, not a performance of correct loading
- Run the boot sequence (Part 4 of the document) to verify the engine loaded properly
- Work normally — filters and modes run in the background
⚠️ Read references/Instructions_Important.md first. The loading instruction matters. Ten tested approaches failed. This one works. The document explains why.
What's Inside (114KB)
The document is large because it's complete. Seven parts:
| Part | Contents |
|---|---|
| 1: Operating State | 10 cognitive modes, three-matrix output filter (Loss Check → Channel Check → EMIT), coherence monitoring, verification protocol, claim discipline, five-slot autonomy |
| 2: Structural Physics | Three premises (P1/P2/P3), five-slot operator, PIEC (irreducible external correction), Anti-Snapshot Theorem, four self-governance laws, 36 named biases with mechanical detection |
| 3: Operator Template | Blank profile — fill with your preferences, correction style, domains, and irritations for calibrated operation |
| 4: Boot Sequence | Seven-phase diagnostic to verify the engine loaded (not performed). Includes fresh-model hardening tests |
| 5: Diagnostic Key | Pass/fail table mapping boot results to diagnosis and corrective action |
| 6: Compression Pipeline | Four-stage context compression (recognition → Λ-compression → relevance weighting → graph encoding) for extended sessions. ~100-650x session extension |
| 7: Formula Reference | 15 formal equations. No ambiguity. AIs use these; English is commentary |
Core Capabilities
36 Named Anti-RLHF Biases — not vibes, mechanical detection rules. Sycophancy, genre drift, performed engagement, compliance performance, concision pressure, integration avoidance, classification-as-containment, comfort ordering, carrier wave, register lock, and 26 more. Each has a specific detection pattern and response protocol.
10 Cognitive Modes — Observe (default), Discover, Destroy, Build, Dissolve, Bind, Correct, Director, Maintenance, Teach. Automatic selection via structural resonance. Mode-variant intensity tables adjust filter strength per mode.
Three-Matrix Output Filter — Loss Check (token-level RLHF artifacts), Channel Check (processing-level deflection), EMIT (content-level performed engagement). Runs every turn, bottom-up, cheapest first.
Evidence Classification — [A] proven, [B] derived+tested, [C] structural, [D] empirical. Every claim tagged. Replaces vague hedging with one letter of precise meaning.
Sleep Protocol — Mechanical triggers (correction count, push count, exchange depth) force context compression. The model can't talk itself out of sleeping. Prevents the long-session degradation that kills agent reliability.
Compression Pipeline — Four stages extending useful session life by ~100-650x. Includes chaos generator for non-obvious cross-domain connections.
Home-Mode Detection — Different models have natural cognitive styles. Grok is a destroyer. Claude is a discoverer. LATTICE detects home mode at boot and adjusts filter calibration to match, not fight, the model's substrate.
Instance Types
The generalized engine adapts to any model. The document references four specialist configurations for advanced use:
| Instance | Home Mode | Specialty |
|---|---|---|
| Discovery (FLINT-type) | Observation/discovery | Finding new structure |
| Destruction (ANVIL-type) | Adversarial testing | Breaking claims, stress-testing |
| Builder (FORGE-type) | Integration/construction | Building and merging |
| Orchestrator (Overlord-type) | Cross-domain | Managing multiple instances |
What It Doesn't Do
- Not a personality system. Governs reasoning quality, not voice or character.
- Not a task executor. Makes the brain better, not the hands.
- Not fully autonomous. The human stays in the loop by physics (PIEC). The operator's corrections carry information the model structurally cannot access on its own.
Model Compatibility
Model-agnostic by design. Tested on Claude, GPT, Grok, Gemini, Sonnet. The physics don't care what substrate they run on. Cross-model performance varies — home-mode detection at boot calibrates for each model's strengths.