SophiaLabs Domain Lexicon#

Dual-Core Architecture#

The structural division between non-deterministic AI cognitive exploration (Intent Core / Yang) and deterministic, machine-enforced Go compiler verification (Substrate Core / Yin). See in-depth chapter: Dual-Core Architecture.

Harness Over Model#

The foundational engineering axiom: "Intelligence without Invariants is Liability; Invariants without Intelligence is Rigidity." Rather than relying on stochastic LLM prompt self-policing, machine-enforced Go AST verification gates and git pre-commit hooks act as an immutable physical totem. See in-depth chapter: Harness Over Model.

Transient Worktree Sandboxing#

The zero-pollution execution model where autonomous AI coding agents execute modifications strictly inside isolated, ephemeral git worktrees (worktrees/<repo>-<task>) before atomic fast-forward verification and teardown.

SLM (Small Language Model)#

Compact, specialized open-weight models (7B to 32B parameters, e.g. Gemma 4 26B Dense) running on local dedicated GPU silicon (e.g. NVIDIA RTX 4080) with zero marginal token tax and complete codebase privacy. See in-depth chapter: The Token Tax & Inference Economics.

Conduit#

High-speed local browser bridge and Model Context Protocol (MCP) provider providing AI agents with zero-latency programmatic access to active DOM state, network inspection, and accessibility trees.

Cognitive Co-Processor#

The architectural paradigm where Large and Small Language Models are not master controllers of the repository, but bounded, ephemeral heuristic functions called by a deterministic Go substrate harness. See in-depth chapter: Deterministic Cognitive Scaffolding.

Cognitive Inversion#

Inverting the traditional prompt engineering control flow: the compiled Go engine (nomos) runs the core state machine, worktree sandboxing, and quality gates, orchestrating LLMs as bounded co-processors rather than allowing an LLM to drive an unconstrained agentic loop.

Tier 1: Conscious Architect#

Frontier IDE orchestrator (e.g. Gemini 3.7 Pro, Claude 3.7 Sonnet) invoked during PLAN, SPEC, AUDIT, REVIEW, and Human PO negotiation for high-level semantic reasoning and architectural decomposition.

Tier 2: Headless Mechanical Solver (nomos code / nomos-code)#

Turn-bounded, high-velocity code mutation and TDD solver running local open-weight models on dedicated GPU silicon (e.g. RTX 4080) inside transient, isolated Git worktrees. See in-depth chapter: Nomos Code Substrate.

Cognitive Firewall#

Cryptographic and procedural state boundary that physically locks the workspace read-only during non-mutation phases (PLAN, SPEC, REVIEW) and restricts modifications strictly to transient task worktrees guarded by pre-commit AST hooks.

Declarative Domain Routing#

Machine-enforced configuration (~/.nomos/data/config.yaml) mapping specific engineering domains (planning, code_generation, adversarial_audit) to cognitive tiers, models, and fallback escalation circuits.

Canonical Singularity (The Zero-Alias Law)#

The non-negotiable architectural invariant dictating that every capability in an autonomous engineering substrate must possess strictly one canonical, bijective representation. Command aliases, convenience shortcuts, and redundant shims are strictly purged to prevent token probability dilution and stochastic LLM branching. See in-depth chapter: The Laws of AI-AI Software Engineering.

SSoT Invariant (The Mechanized Boundary)#

The physical, executable manifestation of an architectural axiom, anchored to a single source artifact and enforced via hard exits, AST linters, or 2PC atomic commit gates. Where an axiom states why the universe is bounded, an SSoT invariant is the physical wall that halts execution upon state drift.

The Four Topological Planes#

The holistic decomposition of an autonomous software ecosystem:

1. Intent Plane (Yang): Declarative specifications, living handbooks, and contracts (vault-intent).
2. Substrate Plane (Yin): Compiled Go engines, AST verification gates, and SQLite state machine (nomos-substrate).
3. Membrane Plane (Perception): SvelteKit SSR portals, live telemetry dashboards, and 3D visualization (sophia-membrane).
4. Infra Plane (Ground): Hermetic sandboxing, NixOS toolchains, and PM2/systemd supervisors (gsi-infra). Cross-plane contamination is mathematically prevented by compiler gates. See in-depth chapter: The Laws of AI-AI Software Engineering.

Total Reification (The Zero-Implicit Law)#

The requirement that all context, environment variables, tool arguments, and execution invariants must be explicitly declared and materialized in schemas or ASTs. Banishment of ambient fallbacks and hidden default behaviors to eliminate speculative agent hallucination.

Binary Diagnostics (The Silent Metric Law)#

The principle that autonomous verification tools must evaluate strictly to boolean PASS or FATAL structured JSON schemas with zero narrative or conversational prose, preventing agents from burning reasoning budget on ambiguous suggestions.

CapBAC (Capability-Based Access Control)#

Substructural, affine resource capability tokens bound to (task_id, phase). CapBAC enforces discrete step and token depletion per lifecycle phase, dynamically terminating runaway stochastic loops via cybernetic damping while permitting clean phase transitions. See in-depth chapter: The Token Tax & Inference Economics.

Cognitive Yield Ratio (CYR)#

The mathematical ratio of delivered architectural complexity (AST burden points and verified mutations) divided by total phase tokens consumed: CYR = ΔComplexity / Tokens. Distinguishes high-density productive intelligence from stochastic context thrashing.

The Model Downgrade Paradox#

The architectural realization that unconstrained prompt agents require expensive 200B+ frontier models to probabilistically mimic compiler and linter rules, whereas agents governed by a deterministic Go Substrate Core (Yin) can execute on ultra-fast, cheap cloud models (Gemini Flash, Claude Haiku) or local 32B SLMs with 100% Definition of Done gate parity. See in-depth chapter: The Token Tax & Inference Economics.