Training Forge: reviewing model updates
A public TypeScript state machine for reviewing model-update decisions.
promote_or_rollback = gated
Public reference implementation with a synthetic executor. No model is trained by the example.
Problem
Keep review and promotion decisions explicit as a candidate moves through a workflow.
Training Forge is a pure state machine over explicit records. It tracks reviewed lessons, screening results, externally supplied trial decisions, promotion and rollback. The public example runs against lookup tables with no model, network or training process. A real integration must supply its own executor and evaluation.
How it works
Review lessons
A lesson needs a reviewer and a probe that reproduces a failure on the incumbent.
Screen candidates
An injected executor checks the candidate against admitted lessons. Unknown results and uncovered lessons block promotion.
Supply a trial decision
An external evaluation decides whether the candidate improved. The state machine records that decision and its incumbent.
Promote or roll back
Promotion preserves the previous incumbent. Rollback restores it with a reason, retaining the decisions and lessons.
Tradeoffs
State machine & execution
The reference manages decisions; a real model executor and evaluation remain separate integrations.
Records & proof
The decision log records supplied results. It has no record hash or replay proof that an executor will answer the same way later.
API rules & mutable state
The API preserves admitted lessons, but code can rewrite the plain state object. This is not cryptographic enforcement.
Checks
- Lesson screening
- Promotion while admitted lessons are broken, unknown or uncovered
- Trial decision
- Promotion without a recorded improvement decision for the current incumbent
- Decision log
- Losing the recorded reason for an API refusal
- Rollback
- Discarding the previous incumbent when promoting a candidate
My contribution
I extracted the review workflow into a standalone public reference.
- explicit state
- reviewed lessons
- injected execution
- promotion gates
- decision records
The repository exposes the TypeScript implementation, a synthetic end-to-end example and tests for its state transitions. Its public example does not establish model-training quality or an employment deployment.
The public example produces a decision log covering lesson admission, screening, promotion and rollback against synthetic behaviour tables. It demonstrates workflow rules, not the quality of a trained model. Kordane Forge is a separate product.
- TypeScript
- explicit records
- injected executor
- lesson screening
- promotion checks
- decision log