Experiment · demonstrated · candidate

A bounded relational procedure passed held-out tests and failed after exact removal

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.

Research question

Can a typed proposer induce one iterative keyed-state transition from examples without receiving the target program?

Why this matters

The held-out, cold-reload, and exact-removal checks test whether a bounded executable state transition explains the frozen cases rather than just restating examples.

What was tested

Bounded schema-derived search over keyed-state transitions with a computed progress loop, relational guard, dynamic map updates, and state carried across passes.

Baseline

No target AST supplied; initial hypothesis was rejected after an ambiguity was identified.

Comparison conditions

Six induction/refinement examples; three untouched acceptance holdouts; fresh-process replay; exact-record removal with induction disabled.

Test environment

Isolated typed-proposer candidate; frozen synthetic contracts; active runtime, model, and registry unchanged.

Observed result

Generated and compiled 181 candidates, behavior-rejected 140, pruned 40 observationally equivalent candidates, and accepted one. The result passed 6/6 induction/refinement examples and 3/3 holdouts; after cold reload it passed 3/3, and exact-record removal produced 0/3.

Outcome

Narrow isolated relational-induction candidate; not promoted.

Measured results

Measurement conditions: One synthetic repeated keyed-state family; the three frozen acceptance holdouts were excluded from search.

induction Examples: 6
frozen Holdouts: 3
candidates Generated: 181
candidates Compiled: 181
behavioral Rejects: 140
equivalent Candidates Pruned: 40
holdouts Passed After Reload: 3
holdouts After Exact Removal: 0
external Model Calls: 0
candidate Promoted: No

Evaluation method

Induce from three training examples and three separately frozen refinement examples; keep three acceptance holdouts out of search; cold-reload the content-addressed candidate with research and induction disabled; remove the exact record as a causal control.

External model usage

Zero external-model calls.

Cold reload

Fresh process, research and induction disabled: 3/3 frozen holdouts passed.

Causal ablation

Exact-record removal with induction disabled: 0/3 holdouts passed.

Transfer

No cross-family transfer is claimed by this experiment.

Verification and lifecycle

verified · candidate · Measured 2026-10-31T00:00:00Z · Report LARI Full Report 2026-10-31T11:42:43.574Z

Limitations

One repeated-update family; isolated candidate, not promoted. A separately authored trusted evaluator labeled ambiguity probes. No arbitrary relational algorithm, cross-family generalization, or Go planner solution is established.

Architecture component: state →

Related experiments

Seven sources and a real test contract—but no Go dependency planner

Public provenance

Owner-reviewed summary of the frozen acceptance, refinement, and cold-verification receipts; private candidate records and hashes are withheld.

Source Stateful relational-induction acceptance and cold-verification receipts · 2026-10-31 · Published 2026-10-31T20:03:40.716Z

Back to experiments

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…