These records examine specific LARI architecture components, from structured knowledge and bounded execution to persistence and model governance. Each approved experiment includes its method, baseline, measurements, provenance, and what the result did not establish. Only approved records appear here. Explore the architecture →
An isolated candidate began with zero implementation modules, induced a four-program typed bundle from local immutable tests and a project contract, created three JavaScript source files, and passed all 12 acceptance checks. The construction run passed twice. A serialized candidate then rebuilt a second blank workspace in a fresh process with induction disabled; removing the generated command-line entry point restored failure. The candidate remains separate from the active model and unpromoted.
An October 23 attempt acquired seven public sources and recovered positive examples and rejection conditions from the immutable Go tests, but the shared typed inducer produced no planner. Later diagnostics still produced no candidate: one direct typed search reached its 100,000-step limit, and a separate relational route rejected the task shape before generating candidates. These are distinct runs and budgets. The compiler’s generic nested-loop sketch is not a Go planner.
2026-10-31T01:11:19.307Z · Program induction and search · execution · Architecture ↗
The shared typed proposer induced one repeated keyed-state update program without receiving the target AST. After an ambiguous first hypothesis, the review retained that failure and added a separately frozen refinement split; three acceptance holdouts stayed out of search. The selected candidate passed all six induction/refinement examples and all three holdouts. A fresh process reloaded the exact record with research and induction disabled and passed the holdouts; removing that record returned performance to zero of three. The candidate was not promoted.
2026-10-31T00:00:00Z · Relational and stateful learning · state · Architecture ↗
The 2026-10-31 source report describes five named coding lanes in active production: traceback-based issue localization, isolated test creation when a native harness is blocked, coordinated multi-file repair, behavior-preserving refactoring, and bounded long-horizon multi-defect repair. It reports 5/5 reuse-only lane checks, three causal ablations, reload, rollback, and same-hash selection across five public surfaces. The active model identity matched the report at review time; the underlying per-lane qualification suite was not freshly rerun for this release.
The 2026-10-31 source report describes 38 atomic RECAP records across seven bounded language families as present in active production. It reports inherited ancestor qualification: 6/6 training families, 12/12 unseen semantic variants, 32/32 new-record ablations, 12/12 reload checks, 17/17 hidden transfer cases, five-surface parity, and no family regressions. A bounded recursive-composition record is also reported active with inherited qualification. The active-model identity matched the report at review time; the family qualification suite was not rerun for this release.
2026-10-31T00:00:00Z · Language generation and discourse composition · state · Architecture ↗
LARI’s documented governance separates evidence, isolated candidates, qualification, active-model state, and explicit promotion. A current qualification receipt verifies one isolated candidate on two workspace languages: fail-before/pass-after, persisted-candidate reload, and exact-operator removal restoring failure. The active model, registry, and runtime remained read-only, the candidate was not promoted, and external-model calls were zero. The report separately describes rollback controls for bounded production lanes; those rollback checks are inherited report evidence, not a fresh rollback test in this receipt.
2026-10-31T10:00:58.566Z · Persistence, verification, and model governance · governance · Architecture ↗
Discuss this page
Read and join this page’s discussion here. Sign in to GitHub when you want to reply or react.
Loading discussion from GitHub…
The embedded discussion could not be loaded. Browse discussions on GitHub ↗