Skip to content
☰

Validation

Every model is L1 today. Here is what that means.

Grades are earned per metric and use, with evidence and sign-off. They are Alive's internal terms, not an external certification.

Draft pending scientific sign-off

Grade definitions and promotion criteria will be reviewed by a named scientific reviewer before launch.

Evidence grades

L1

Executed and tested

The behavior ran; a trace and assertions exist.

L2

Internally verified

Invariants, extreme cases, distribution checks and replay all pass, with a documented report.

L3

Calibrated

Compared with a documented external dataset within stated tolerances, for a stated use.

L4

Externally validated

Independently checked against real outcomes for a stated population and use, with prediction errors disclosed.

Tests every model passes today

  • Byte-identical replay of every reference world on a locked runtime
  • No phantom events: every completion has a matching creation (trace assertions)
  • Every intervention request settles exactly once
  • Sanity extremes: zero staff serves nobody; a population of one runs
  • Variants share the same synthetic people and per-person random draws
  • Paired confidence intervals cross-checked against an independent statistics library
  • Probability parameters confined to [0, 1]; units checked against each model

These run on every change to Alive. They show the software does what its models say; they do not show the models match reality. That is what calibration (L3) and external validation (L4) are for.

What reaching L3 will require

A documented external comparator, stated tolerances, a holdout comparison, a reviewer's sign-off, and a version-locked benchmark report. Parameter defaults already cite published evidence where it exists (see each model), but citing evidence for an input is not the same as calibrating an output.