Detect syntax failures, test breakages, or system telemetry alerts to isolate the defect.
KAIZEN is not a pure loop — it is a spiral. Each pass ends higher than it began. The cycle lives between AEGs (AXIOM Execution Graphs), not inside them. Steps 1–5 are one AEG. Step 6 is the ratchet click (AVS snapshot). Then the next AEG begins on the newly-standardized state.
Internal model axiom-kaizen runs on local GPU hardware (Ollama/vLLM), combined with foreign model swarm (Gemini, Claude, GPT) for complex logic decomposition and code modification.
Git commits → SFT JSONL dataset with strict public/private isolation. No API keys, no secrets in public weights. Internal dataset trains the next axiom-kaizen iteration.
KAIZEN autonomously designs and tests new ML approaches, dataset processing, and hosting strategies. Verifies performance metrics and self-upgrades code, schemas, and models.
Faculties are equal-rank cognitive units. Each produces a typed output and a training signal. The Orchestrator is distinguished only by its output type (AEG), not by authority.
Pending approvals from the KAIZEN loop. Review diffs, approve or reject candidates. Approval requires signing with your ax-id keypair.
Every Faculty invocation produces a training row: prompt, raw model response, parse result, downstream verification outcome.
Git diffs → SFT JSONL. ReAct trajectory synthesis. PII scrubbing. Public/private isolation barrier.
QLoRA training on local GPU (Unsloth) or cloud GPU (RunPod). Produces candidate model artifact.
Candidate model goes through review → HITL → ratchet gate. Operator must sign deployment with ax-id.
New model promoted to production. AVS snapshot locks the new standard. Ratchet clicks forward.
POST https://train.axiom.farm/api/submit
GET https://train.axiom.farm/api/status
GET https://train.axiom.farm/health
Multi-model swarm audit with monotonic progress guard. Faculties are scored by a swarm of models, blended with human votes to produce a planetary consensus score. The score can only go up — regression is blocked by the ratchet.
If the new planetary score is lower than the current high-water mark, the regression is blocked. The progress score remains at the previous maximum.
50% AI swarm agreement + 50% human validation rate = planetary consensus score. If no human votes exist, score defaults to AI-only.
When a concept reaches 90%+ consensus, it is elevated to Planetary Axiom status — the highest designation in the system.