Evidence levels and finding types

Review status: draft pending scientific sign-off.

Finding types

Type	Comes from	What it means
verified_failure	an assertion that failed in a run's event trace	It happened in that synthetic run; the cited event ids show where. Whether it happens in reality depends on how faithfully the world represents it.
modeled_outcome	a metric or comparison in a simulation	What the model produces given its stated assumptions. Not a measurement of real people.
ai_hypothesis	an AI-drafted review	A reasoned starting point to test. Never "verified".
expert_assessment	a named human reviewer	A person's judgement, recorded with who and when.

There is no overall readiness score. A single number would hide exactly the distinctions above.

Evidence grades (Alive's internal terms)

Grades apply to a metric for a use case, not to a whole world:

• L1 — executed and tested. The behavior ran, a trace and assertions exist.
• L2 — internally verified. Invariants, extreme cases, distribution checks and replay pass.
• L3 — calibrated. Compared with a documented external dataset within stated tolerances.
• L4 — externally validated. Independently checked for a stated population and use, with prediction errors disclosed.

These are Alive's own definitions, not an external certification. Every current model is L1. A WorldSpec cannot claim a higher grade than its models have earned; the compiler rejects it.

Parameter provenance

Every number in a world records where it came from: measured, public_evidence (with citations), expert_assumption, or synthetic_assumption. An AI can never be a provenance source. Results state how many assumption parameters they depend on.