Evolution

Evolution is the behavior of a model across retraining. Each training run takes the current supervision set and produces a trained model. When the supervision changes and training is run again, the new model expresses the updated standard.

§1Definition

Retraining is a training run performed on a supervision set that has been modified relative to a previous run. The result is a new trained model fit from the modified supervision. Retraining can be triggered whenever the supervision set changes — samples added, corrected, or removed; a brief revised; references updated.

§2Mechanism

§2.1Deployment replacement

The retrained model replaces the previously deployed model under the same handle. Subsequent scoring calls use the retrained model. Retention of prior trained models is not part of the shipped behavior; see roadmap for planned versioning.

§2.2Effects on breaks and scores

Retraining runs calibration from the updated samples, which can shift the breaks on affected traits. A fixed piece of content scored before and after retraining may receive a different tier label if the breaks have moved past its score, even when the numeric score is similar. Traits added or removed between runs change the composition of the composite; the composite of a fixed piece of content can change for that reason alone.

§3What determines how much changes

How far the retrained model departs from the previous one depends on how the supervision set changed:

Change to the supervision setEffect on the retrained model
Samples added that are consistent with the existing distributionsBreak positions tighten around existing values; confidence may increase; trait axes are substantially unchanged.
Samples added that contradict existing labels in regions of the score axisBreak positions can shift toward the new samples; confidence in the contested region can decrease if the contradiction reduces distribution separation.
Traits added or removedTrait scores for the removed traits disappear; scores for the added traits appear; composite recomposes over the new set.
Brief revised or references updated (for non-sample supervision)The resolved sample set changes; effect depends on how the resolved samples differ from the previous resolution.

Table 1. How supervision changes translate into model changes on retraining.

§4Edge cases

§4.1Retraining with no changes

A retraining run on an unchanged supervision set produces an equivalent model. Deterministic components (scoring, calibration, break derivation) return the same values. Any non-deterministic selection step uses the same random seed when available; minor variation is possible only for components that depend on sampling.

§4.2Retraining that degrades the model

Supervision changes that introduce contradictions — for example, samples labeled positive that are similar to existing negatives — can reduce distribution separation and increase the overlap region. The retrained model will reflect the lower separation as lower confidence in affected regions. Retraining does not validate that the new supervision improves the model; the contributor is responsible for evaluating whether the change expresses the intended standard.

§5Related concepts

  • Training — the process retraining re-runs.
  • Calibration — the step that can shift break positions between runs.
  • Breaks — the values that move when training distributions change.
  • Roadmap — planned model versioning that would retain prior trained models across retraining runs.
Scores are approximate — not a substitute for human judgment.