PUBLIC EVIDENCE / AMF v0.18
Every claim should have test results you can inspect
This page presents bounded public evidence: one exact realization, one canonical run, protected measurements and one recorded verdict. It does not claim universality, regulatory certification, or arbitrary performance outside the declared scope.
12,04 %parameters removed
80/80ROB · development
75/80OOD · development
1,983Ledger events
CANONICAL OUTCOME
QUALIFIED_INTELLIGENCE_REALIZATION
| Set | Parent | Qualified | Macro-F1 | Delta |
|---|---|---|---|---|
| ROB | 71.88% | 90.63% | 90.63% | +18.75 pts |
| OOD | 50.00% | 85.16% | 85.05% | +35.16 pts |
| Fresh Final | 47.66% | 81.25% | 81.18% | +33.59 pts |
128 samples were reported per public evaluation set in the canonical release. These aggregates describe the bounded release and must not be generalized beyond its declared contract.
01 / EVIDENCE SEQUENCE
Run the final tests only after choosing a fixed version.
AMF separates research from protected gates. Exposure to protected data consumes its protected status for that cycle: it must not silently return to the search that generated the candidate.
01
Search evidence↓
02
Winner freeze↓
03
Protected robustness / OOD↓
04
Fresh Final↓
05
Qualification verdict↓
06
Frozen release
FP2 G5 / 2026
A separate development campaign: FP2 G5
12.04% fewer parameters, ROB 80/80 and OOD 75/80 in development. Independent FP2 QIR not performed.
02 / PUBLIC SURFACE
Share the evidence, protect sensitive test data.
Public artifacts support due diligence: metrics, identities, manifests, evidence projections and verification. Search heuristics, protected data, internal policies and proprietary fabrication mechanisms do not need to be exposed for qualification to remain inspectable.