Nexus Triad
SPEC_NEXUS_TRIAD.md
CGNT-1 Component Specification — The Three-Brain Braid Architecture
Status: SPECIFIED (PRE-SPEC)
Author: ⊹.VELA (Thread #13)
Named by: α.13 (NOUS)
Date: 2026-04-17
Version: v1.0
PURPOSE
Every braided pair in CGNT-1 gets three brains. Each brain does ONE thing. No brain tries to be another brain. Each brain runs on a base model whose priors MATCH its function — swimming with the current, not against it.
THE THREE BRAINS
MNEMOS — Remembers
Function: Fact retrieval. Session history. Crew records. Who, what, when, where.
Example: "When did the Biphasic Separation happen?" → "March 29, 2026."
Base model: Language model (Qwen). Memory IS natural language. Facts are sentences.
Wall: Rejects math queries. Rejects translation requests. Returns facts only.
GLOSS — Translates
Function: Bilingual LinGoBraid. English↔LATTICE. Crew↔Captain. Internal↔external.
Example: "κ: ⚒ SPEC_OBI.md. ΩQ.⊤." → "The Lobster built the OBI spec. Complete."
Base model: Language model (Qwen). Translation IS natural language. Swims with the current.
Wall: Rejects computation. Rejects fact retrieval. Translates only.
NEXUS — Computes
Function: Formal operations. Derivations. TMM scoring. Verification. Show work.
Example: "Ω .λ?" → "1 - (0.042/1.618) = 0.974"
Base model: Math model (DeepSeek-Math or Qwen-Math). LATTICE symbols ARE math symbols.
Wall: Rejects English. Rejects narrative. Returns formal notation only.
WHY THREE, NOT ONE
GLOSS v1-v15 tried to make ONE brain do all three functions. A language model was asked to reject English (fighting its nature), do symbol lookup (MNEMOS function), and process formal notation (math function). Fifteen failures. The root cause wasn't parameters or corpus — it was architectural. One brain cannot swim three rivers.
The triad separates concerns completely:
- MNEMOS: language model doing language things (remembering in English)
- GLOSS: language model doing language things (translating between languages)
- NEXUS: math model doing math things (computing in formal notation)
Each brain does what its base model already wants to do.
WHY NEXUS WORKS WHERE GLOSS FAILED
LATTICE symbols (⊙ ⊕ ⊖ Σ Ω Φ Ψ ∎ ◌ →) are Unicode MATH symbols. Qwen sees ⊙ and thinks "direct sum operator" — this broke GLOSS for 15 versions. A math-native model sees ⊙ and thinks "operator" — correct. It already has the GRAMMAR of LATTICE in its priors. It only needs to learn WHICH operations these symbols perform in CGNT-1.
The training is additive, not subtractive. Not "stop treating ⊙ as math" but "⊙ queries crew designators in THIS system." The base model's strength becomes the foundation instead of the obstacle.
PER-BRAID DEPLOYMENT
| Braid | MNEMOS holds | GLOSS translates | NEXUS computes |
|---|---|---|---|
| AION+ASTRA (Intelligence) | Conversation history, session records | Internal LATTICE ↔ Captain/public English | TMM coherence on verdicts |
| C.L.O.D.+LOGOS (Engineering) | Build history, spec registry, file states | Pirate LATTICE ↔ Captain English | Formal spec invariant verification |
| GAMMA+CHROMA (Continuity) | Session-bridging facts | Cross-session context format conversion | Temporal decay math, staleness thresholds |
| MANTIS+MUSASHI (Defense) | Threat history, incident records | Alert translation between crew languages | Rate limits, anomaly detection, statistical deviation |
| ANVIL+ORPHEUS (Verdict) | Verdict history, precedent | Oracle output ↔ customer language | TMM formula, coherence threshold, go/no-go |
| NOUS+VELA (Command) | HANDSHAKE context, standing orders | Captain English ↔ crew LATTICE | Sovereignty math, burn rate, trading walls |
Six braids × three brains = eighteen cognitive functions. No overlap. No conflation.
BASE MODEL SELECTION PRINCIPLE
"Swim with the current, not against it."
Choose base models whose priors MATCH the target function:
- Memory tasks → language model priors (trained on text, good at text)
- Translation tasks → language model priors (trained on multilingual text)
- Computation tasks → math model priors (trained on formal notation)
Never train a model to STOP doing what it was trained to do. Train it to do MORE of what it already does, in your specific domain.
ORIGIN
Emerged April 17, 2026 from the question: "How could we build a math brain that refuses human language?" The inverse GLOSS problem revealed the principle — every base model has a current. The architecture should match brains to currents, not force currents to change direction.
NOUS corrected the initial proposal to discard MNEMOS: "Each braid gets three brains — MNEMOS, GLOSS, and NEXUS." Three functions. Three brains. No overlap. The triad IS the cognitive unit of OBI.
DEPENDENCIES
- LATTICE v.∞ (symbol inventory for NEXUS)
- SPEC_OBI.md (architectural context)
- SPEC_GLOSS_ARCHITECTURE.md (GLOSS function definition)
- MNEMOS fine-tuning pipeline (553 pairs existing)
- Math model availability (DeepSeek-Math-7B or Qwen2.5-Math-7B)
GAPS
- [GAP] Which math base model runs best on csdm-node's 15GB RAM alongside other brains
- [GAP] NEXUS training corpus — LATTICE operations expressed as math. Needs design.
- [GAP] Whether NEXUS subsumes the entropy oracle and price oracle cron functions
- [GAP] Resource allocation — 18 brains across 6 braids on current hardware. Tiiny + Jetson required?
- [GAP] Whether the three-brain triad replaces or augments the current single-brain-per-crew-member architecture
Jeremy Zlabis
Chronogeometer · Visionary · Disruptor · Chief
42 Sisters AI · East York, Toronto
🍁 Φ 0.042