Learnx
SPEC_LEARNX.md
CGNT-1 Component Specification — LEARNX
Status: SPECIFIED
Version: v1.0
Module: 23 of 24
Author: VELA (Thread #13)
Authorized: NOUS
Date: 2026-04-20
PURPOSE
Continuous learning engine. Closes the feedback loop between what the crew experiences and what the brains know. Every session, every query, every correction is a potential training signal. LEARNX extracts, filters, and queues that signal for the forge.
LEARNX feeds FORGEX. It never triggers a forge itself.
THE LOOP
Experience (session / query / correction)
│
├─→ LEARNX nightly scan (02:00 ET)
│
├─→ Candidate pair extraction
│
├─→ Quality gates (dedup, consistency, format)
│
├─→ Human review queue (most pair types)
│ └─→ Captain / GAMMA approval
│
├─→ Auto-approval (FACT pairs only — verified against deterministic tools)
│
└─→ FORGEX queue → next brain retrain
DATA SOURCES
| Source | What LEARNX extracts |
|--------|---------------------|
| Sisters sessions (SESSIONS.md) | Corrections, new facts, CSDM reasoning examples |
| 42sisters.ai public chat | Customer questions that expose gaps, hallucinations caught |
| Lobster sessions (LOBSTER_LOG.md) | Engineering facts, deployment outcomes |
| ROUTX query logs | ◌ (gap signal) queries — where no tool or brain could answer |
| HACKX observations | Attack patterns for MANTIS pairs (via HACKX→LEARNX pipeline) |
| GAMMA extractions | GAMMA's post-session analysis (highest quality source) |
PAIR TYPES
| Type | Definition | Auto-approve? | Forge target |
|------|-----------|--------------|-------------|
| FACT | Verified correct Q→A pair, checkable against GLOSS/NEXUS | YES (if tool-verified) | Any brain |
| CORRECTION | Wrong answer + correct answer (caught hallucination) | NO — human review | Target brain |
| GAP | ◌ response where answer exists in specs | NO — human review | Target brain |
| NEGATIVE | Hallucinated answer + correct answer (explicit contrast) | NO — human review | Target brain |
| OPERATIONAL | New architecture fact, deployment outcome, system state | NO — human review | MNEMOS |
| KERNEL | CSDM physics pair — new equation, theorem, observation | NEVER — Captain only | MNEMOS + Sisters |
QUALITY GATES
All pairs pass through these gates before entering any queue:
| Gate | Check | Action on fail |
|------|-------|---------------|
| Deduplication | Hash of (question, answer) vs existing corpus | Discard |
| Consistency | Answer checked against GLOSS (symbol lookup) and NEXUS (math) | Flag for review |
| Format | Valid UTF-8, reasonable length (Q: 10–500 chars, A: 10–2000 chars) | Discard or truncate |
| Source attribution | Pair must have traceable source | Discard if unknown |
AUTO-APPROVAL RULES
FACT pairs only. Auto-approval requires ALL of:
- Question is a lookup or computation (not open-ended)
- Answer verified against a deterministic tool (GLOSS, NEXUS, or SPECX)
- Tool returned a definitive non-◌ result
- No conflicting answer found in existing corpus
All other pair types enter human review queue regardless of apparent quality.
KERNEL PAIRS — SPECIAL PROTOCOL
KERNEL pairs are CSDM physics facts: equations, invariants, theorems, physical constants, and their derivations. These are the most valuable and the most dangerous category.
Rules:
- KERNEL pairs are NEVER auto-generated by LEARNX. They are always manually authored.
- Every KERNEL pair requires Captain (NOUS) review and explicit approval.
- GAMMA may propose KERNEL pairs; NOUS approves them.
- KERNEL pairs feed MNEMOS AND the Sisters system prompts.
- Wrong KERNEL pairs corrupt the manifold. There is no recovery path except manual correction.
SCHEDULE
| Time (ET) | Action |
|-----------|--------|
| 02:00 ET nightly | LEARNX scans all data sources, extracts candidates |
| 02:30 ET | Quality gates applied, pairs sorted by type |
| 03:00 ET | FORGEX queue updated with approved pairs |
| On-demand | ROUTX query: "learnx status" / "learnx queue" / "learnx gaps" |
INTEGRATION
| System | Relationship |
|--------|-------------|
| FORGEX | Writes approved pairs to forge queue — LEARNX's primary output |
| GAMMA | Highest-quality source of pair candidates |
| GLOSS / NEXUS | Used for FACT pair verification |
| SPECX | Used for consistency checking |
| ROUTX | Registered as module 23 |
| HACKX | Receives attack observations → MANTIS training pairs |
| COMMX | Publishes nightly summary to crew |
INVARIANTS
INV-01: KERNEL pairs are never auto-generated. Manual authorship only. Captain approval always.
INV-02: Auto-approval applies only to FACT pairs verified against deterministic tools. No neural verification for auto-approval.
INV-03: LEARNX never triggers a forge. It only feeds FORGEX queue. Forge decisions are FORGEX's domain.
INV-04: Source attribution is mandatory for every pair. Unattributed pairs are discarded, not queued.
INV-05: Deduplication is exact-match on (question, answer) hash. Semantic similarity is not sufficient for dedup — a question phrased differently is a new pair.
INV-06: LEARNX never modifies existing training data in place. It only appends to the queue.
GAPS
- Engine file:
~/learnx_engine.py— not yet built - 42sisters.ai chat log ingestion not yet wired
- HACKX→LEARNX pipeline not yet implemented
- No systemd unit yet
- Human review queue UI not yet defined (currently: TASK_QUEUE.md entries)
Jeremy Zlabis
Chronogeometer · Visionary · Disruptor · Chief
42 Sisters AI · East York, Toronto
🍁 Φ 0.042