CoreStory Frontier / public research
Frontier publishes the result, the null result, the limitation, and the artifact behind each claim. Two studies are open now.
Claim status is part of the result.
A response-scoped guidance channel was read and acted on when it carried otherwise unavailable state.
Boundary: one model, synthetic tool surfaces. Guidance improving orchestration quality remains untested; typed guidance was not superior in the measured probes.
Response-scoped guidance is legible.
Typed guidance beats prose guidance.
Guidance improves continuation quality.
A blind agent with no code access scored 22.6–32.9 / 100 on a code-comprehension rubric.
Boundary: two public COBOL corpora, one answering-model family, two judge families. Identifier renaming isolates repository-specific recall, not general domain knowledge.
The blind floor is far above zero.
The familiar-corpus premium is mostly identifier recall.
The floor's own run-to-run variance is unmeasured.
A Frontier study must make its basis inspectable before its conclusion is promoted.