01
Proposal
Models, learned fields, heuristics, and working beliefs may rank candidates or direct attention.
Current research programme
Verifier-governed cognition under uncertainty.
PRIME studies how a bounded system can choose what to inspect, change its own working representation, preserve useful state, and abstain—without allowing the proposing mechanism to certify itself.
Invariant
Learned systems may propose. Declared verifiers decide what becomes accepted state.
02
Evidence
Fresh outcomes, sequential tests, hard-risk filters, and minimum-sufficient information constrain the candidate.
03
Authority
Declared verifiers decide whether state is repaired, retained, retired, restored, or left unresolved.
04
Receipt
Accepted changes preserve provenance, hashes, replay paths, abstentions, and the exact boundary of the result.
Evidence snapshot
PRIME v1.0.0 ships the argument and the pressure test.
The first public release corresponds to internal scientific generation v18. It includes its frozen evidence, an initially rejected RC1 candidate, tamper checks, deterministic replay, and an explicit scope statement.
Open the release97/97
authorised integrated passes
0
wrong authorised worlds
100/100
continuing lifecycles
500/500
quotient theorem checks
Claim boundary
What this programme does not claim.
The scope is part of the result. Wider language would make the research sound larger while making it less credible.
Finite symbolic and controlled-Markov settings—not unrestricted real-world cognition.
Selective information acquisition—not universal POMDP-optimal planning.
Bounded representation repair—not unrestricted ontology or causal discovery.
A mounted real-prose proposal field—not unrestricted language understanding.
Strong reproducibility evidence—not evidence of AGI or general safety.