Work and projects
internal tool

Local model training

Reviewing corrections and comparing local model updates.

promote_or_rollback = gated

These notes describe the implementation without client-specific details.

current · v13candidate · v14
reviewed correctionimmunity gate passfaithfulness passcompetence passregression passpromoterollback: v13 stays one step away
a representative promotion decision · synthetic

Problem

Corrections need review before they become training material.

A correction can be useful, incomplete, or wrong. This tool collects corrections as candidate lessons, screens them before training, and compares the resulting model with the current version. Promotion includes a review step, with versioning to support rollback.

How it works

  1. Capture corrections

    Corrections are stored as structured candidate lessons for review before use.

  2. Screen candidates

    An immunity gate checks candidate lessons before they enter training.

  3. Compare versions

    Faithfulness and competence scorecards compare a candidate with the current model.

  4. Review promotion

    Evaluation results and human approval determine whether a candidate is promoted.

Tradeoffs

  • Iteration & evaluation

    Screening and evaluation add time between a correction and a model update.

  • Automation & review

    The workflow automates training and comparison steps, with a person reviewing promotion.

  • Updates & rollback

    Keeping previous versions available adds versioning work but supports reverting a change.

Checks

Immunity gate
Screen candidate lessons before training
Head-to-head evaluation
Compare candidate and current model results
Promotion gate
Review results before using an update
Versioning and rollback
Keep a path back to a previous model

My contribution

My work covers the correction and evaluation workflow.

  • candidate capture
  • screening + scoring
  • evaluation harness
  • promotion gates
  • versioning + rollback

This includes the promotion criteria and the tooling used to compare and manage model versions.

The workflow produces candidate lessons, model comparisons, and a review step before promotion. The scorecards describe performance on the evaluated tasks, and previous model versions remain available for rollback.

  • local models
  • immunity gate
  • faithfulness scorecards
  • competence scorecards
  • promotion checks
  • versioning + rollback