Detect syntax failures, test breakages, or system telemetry alerts to isolate the defect.
Every progress, across every workspace, as one board — mined backlog, the forge in flight, the HITL gate, done, failed, and the ratchet clicks. Each card carries its workspace. HITL cards are resolved in place; nothing is applied without a human.
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.
The forge mines tasks from the codebase, an agent writes a change, the judge reviews it, and nothing is applied until a human approves it at the HITL gate. Every applied change is an AVS snapshot.
Each faculty invocation emits a training signal. Interactions roll up into the signal store and ship to the training service. The models to train and the endpoint are set under Configure → Models.
The ratchet only moves forward. An AVS snapshot freezes the new standard; the next pass begins on top of it. History is content-addressed and hybrid-signed — verifiable at ratchet.axiom.farm.
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.
Captured signals and AVS snapshots/diffs → a dataset with public/private isolation. No secrets in public weights.
Train a candidate on the configured model and endpoint (Configure → Models). No training runner is registered on this host yet.
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.
GET /api/kaizen/status?workspace=<id>
training.total_signals · training.pending
/api/kaizen/snapshot · /api/kaizen/board
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.
Operator configuration and controls. Reads are open to everyone; changing
anything — codebases, models, policy, snapshots, the forge — requires an
authenticated operator. The session is shared across .axiom.farm,
so signing in once at the auth portal authorizes every AXIOM surface.
The self-dev loop: mine → forge → judge → HITL apply → ratchet. Start to mine and work tasks; stop cooperatively (the in-flight cycle finishes).
Take an AVS snapshot of the selected codebase. The ratchet only moves forward.
Every codebase the loop works on, and the AVS repos it ratchets. Saved immediately and opened live — no restart needed.
The models the forge uses — dev writes, judge reviews — at the inference endpoint. Training consumes the same endpoint once a training runner is configured.
Content-addressed snapshots per codebase. The full chain explorer lives at ratchet.axiom.farm.
Mining cadence, poll interval, learning tick. Saved to gateway.toml
and applied live — a running forge is restarted so the new cadence takes
effect immediately.