Bounded minimum repair
Optional insertion, deletion, and replacement search interprets tables without running actions, then replays one selected repair through the normal runtime.
Measured error UX evidence →EVIDENCE BEFORE PROMOTION
Research features keep visible boundaries.
Experimental surfaces require explicit activation, publish resource budgets and failure modes, and may change or remain outside the stable contract.
Algorithm research
IELR remains preview after its successful spike. Generalized and incremental parsing are not shipped because their measured spike gates were not met.
Read the research decisions →Hard budgets
Ambiguity, counterexample, repair, and parser-session work is bounded. Exhaustion is a first-class result or structured exception rather than an unreported fallback.
Inspect budget APIs →Promotion policy
Promotion requires a public contract, representative use, adversarial tests, performance evidence, migration notes, and complete documentation.
Read the complete policy →