Score the whole system.
Model, harness, worker, tools, permissions and verifier are one evaluated system.
THE RESEARCH MACHINEMake every claim
Make every claim
earn its confidence.
Every recommendation can be overturned by stronger evidence. Prices, boundaries, failures and unknowns remain visible instead of being averaged into one magic score.
THE EVIDENCE LADDER
A demo starts at the bottom.
E0LeadSomeone says it exists.
E1Source-backedCode or first-party machinery exists.
E2ReproducedWe ran the exact thing.
E3ProductionReal throughput or traces exist.
E4IndependentA neutral operator confirms it.
E5LongitudinalIt survives months, incidents and change.
FACTORYBENCH
One obstacle course. Same rules.
1Intent
2Plan
3Build
4Verify
5Deploy
6Incident
7Recover
8Learn
bad requirementsecret leakAPI changelate datafalse RCAfailed restore
Keep authorities separate.
The proposer cannot silently approve, verify and ship its own work.
Preserve every failure.
Unknowns, reverts and contradictions teach more than polished demos.
Count the humans.
Minutes, expertise, interrupts and approvals are part of the cost.
THE COMPLETE SOURCE LAYER
Open raw research ↗Open every caveat.
Read the ontology, dossiers, evidence ledger, benchmarks and research passes behind this teaching layer.