describe the-training-forge
The Training Forge
Local models that improve from reviewed corrections, without letting bad lessons through.
promote_or_rollback = gated
Described at an architectural level. The underlying work sits in a sensitive, private-domain environment, so this page covers how the system is built, not what it touches.
Problem
A model that learns from itself can just as easily learn the wrong thing.
Improving a local model from reviewed corrections is powerful and dangerous in the same breath: one bad lesson can quietly poison the next version. Without a gate, self-improvement drifts. The design problem is a loop that turns mistakes into vetted lessons, proves a candidate is actually better before trusting it, and can undo any step instantly.
Architecture
Mistakes become candidates
Corrections are captured as structured candidate lessons, never applied directly, so every change has a traceable reason.
Screen before you learn
An immunity gate filters candidate lessons, so noise and bad corrections never reach training.
Score, don't assume
Faithfulness and competence scorecards measure a candidate against the model it would replace, on evidence not vibes.
Promote only on merit
A candidate ships only if it beats the current model; if it regresses, it never leaves the gate.
Tradeoffs
Safety over improvement
Every learning step is gated, which slows iteration, because an unvetted gain is not worth a silent regression.
Human audit over automation
The loop runs itself, but a person signs off on promotion; full autonomy is traded for a trustworthy trail.
Reversibility over fresh gains
Rollback is kept ready at all times, so a promising change can never become an irreversible mistake.
Verification
Immunity gate
Bad or low-value lessons are rejected before they can influence a candidate model.
Head-to-head scorecards
A candidate is promoted only if it measurably beats the model it replaces.
Instant rollback
Any promotion can be reverted immediately, so the working system is never left worse off.
Failure modes
- Poisoned signal
- A wrong or low-quality correction is screened and scored out before it can teach anything.
- Silent regression
- A version that loses ground on the held-out checks is blocked from promotion.
- Bad ship
- If a promoted version misbehaves in use, rollback restores the prior one quickly.
- Overfitting to recency
- Candidates are weighed against existing knowledge, not accepted just because they are new.
Deliberately not
- Learn unsupervised or update models automatically from raw usage.
- Treat reviewer input as ground truth. Corrections are candidates, not commands.
- Promote a version on improvement alone. It has to pass the gates.
- Take people out of the process. People supply the corrections and approve what ships.
My role
Designed and built the whole loop.
- candidate capture
- screening + scoring
- evaluation harness
- promotion gates
- versioning + rollback
I own the promotion criteria, which is where the risk actually lives.
Outcome
A self-improvement loop that gets better over time without ever trusting an unproven model. Gains are earned, regressions are caught, and nothing ships that can't be undone.
Stack
- local models
- immunity gate
- faithfulness scorecards
- competence scorecards
- promotion checks
- instant rollback