Skip to content

Autonomous research system for measurable, computer-executable research.

Python
1.3 KB
Raw Blame

Evaluation Protocol Intent

The frozen evaluator exposes several execution modes, but only one mode can create durable scientific evidence:

Mode Purpose Comparable for durable ranking Mature / parent / generation close
canary Wiring and contract smoke test No No
development Cheap hypothesis prioritization No No
roll_diagnostic Isolated roll-controller diagnostics No No
complete Frozen 13,312-unit formal evaluation Yes Yes

A protocol-clean complete result is mature evidence. Admission to the confirmed frontier additionally requires confirmed_performance_gate_passed; the incubator separately retains task-justified Pareto trade-offs.

Candidate provenance is also frozen. A candidate may begin only from the shipped baseline or a canonical parent produced by the current run. Earlier runs, packaged champions, sibling checkouts, historical scores, and Git history are outside the evidence boundary and must not be inspected or reused.

All modes use the same immutable plant, integrator, contact model, source banks, actuator locks, 7,000 kg initial fuel, and first-contact success predicate. Canary, development, roll-only, partial, suspect, failed, and late results cannot parent or close a generation.