Research · Questions and evidence
Research into a developing architecture.
LARI research asks whether an AI model can acquire, construct, compose, and retain executable representations as part of a locally controlled architecture. This is a scientific effort to make a different model—not a claim that current systems already deliver general intelligence on inexpensive hardware. The questions below describe active investigations and evidence boundaries.
01 · Executable learning
Can experience change behavior in a checkable way?
Problem. A system can appear to improve because a developer added a shortcut or a test reused known code. Research must separate those explanations from behavior acquired from examples or failures.
Approach. LARI stores typed rules and programs, then tests candidates against baselines, unseen cases, ablations, reloads, and rollback where applicable.
Demonstrated. The report describes bounded candidates with causal and transfer checks, alongside failures and unpromoted work.
Unresolved. Independent generalization, durable benefit across ordinary use, and broad capability remain open questions.
Browse reviewed experiment records →02 · Representation growth
What structures should a growing model acquire?
Problem. Memorizing outputs does not explain how knowledge can be composed or reused across tasks.
Approach. LARI investigates typed operators, semantic relations, recursive generators, and executable structures that can be selected by a shared kernel.
Demonstrated. Some bounded structures have passed family-specific qualification in historical and current evidence.
Unresolved. Growth must remain useful outside the training contrasts, avoid regressions, and preserve meaning under composition.
Read about structured model state →03 · Program induction
Can a system construct a procedure from examples?
Problem. Procedures can be authored or retrieved; neither alone proves induction.
Approach. Bounded candidate synthesis uses explicit task contracts, constrained search, execution, and immutable checks.
Demonstrated. Certain coding lanes and program families have bounded passing evidence, while object-program and broader repository tasks also show failures.
Unresolved. Search cost, unseen structures, ambiguous requirements, and arbitrary repository work need stronger qualification.
See the observatory status →04 · Language generation
How far can structured generators support language?
Problem. Bounded conversational families are not unrestricted language understanding or generation.
Approach. LARI represents selected discourse and semantic operations as composable records and tests hidden variants and ablations.
Demonstrated. The report documents qualified bounded RECAP families and recursive composition in production, with family-specific gates.
Unresolved. Broad generation, long-horizon dialogue, multilingual coverage, and robust open-domain semantics remain limited.
View capability qualification →05 · Persistent model state
What survives process boundaries and rollback?
Problem. User memory, a candidate learned record, and an active model are different state layers.
Approach. Tests distinguish scoped persistence from model promotion and measure reload, isolation, and rollback separately.
Demonstrated. The internal record includes fixture-backed persistence checks and separate registry operations.
Unresolved. A narrow fixture result cannot establish general durability, multi-user safety, or all entry points.
Read the state model →06 · Verified learning
What does “verified” mean for a learned behavior?
Problem. Passing examples may reflect leakage, a pre-existing intrinsic, or a test-specific patch.
Approach. Use sealed transfer cases, causal ablation, cold reload, parity across surfaces, regression checks, and explicit model binding.
Demonstrated. The report records both qualified bounded results and candidates that remain unpromoted or fail a gate.
Unresolved. Qualification is only as broad as the tested families and environment; it is not a global certification.
Read the claim-scope guide →07 · Generalization
Can a rule travel to genuinely new tasks?
Problem. Paraphrases and near-duplicates can overstate transfer.
Approach. Separate discovery from sealed validation and test distinct domains, schemas, and execution paths.
Demonstrated. Several bounded transfer studies report held-out successes; other studies fail when the task family or interface changes.
Unresolved. The evidence does not establish general-purpose reasoning or robust transfer to arbitrary settings.
Follow reviewed milestones →08 · Efficiency
What are the costs of search and qualification?
Problem. A method that works only with excessive search or expensive repeated verification may not be practical.
Approach. Record bounded search steps, runtime conditions, test counts, and comparable baselines rather than collapsing unlike tasks into one score.
Demonstrated. Current report evidence includes measured bounded search and performance observations, including unproductive searches.
Unresolved. Cross-task efficiency comparisons and end-to-end user value need consistent conditions and fresh evaluation.
Read the evaluation method →
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 ↗