Using the finite-input property audit

riido-propertyaudit is an offline Go maintainer tool aligning request meaning with truth evidence for tiny claim-model development. It loads no Laya or model weights. Humans and agents can follow request→candidate source→listed input→expected/observed values. It does not authorize general function correctness or safety.

The easiest check from the repository with Go1.27.1 is:

go test ./internal/finiteproperty ./cmd/riido-propertyaudit

This reads public evidence and replays source behavior. Linux/macOS CI checks observation equality, adding no official experiment/model calls. The preserved Darwin executable was not thereby executed on Linux.

Historical collection procedure

This describes the completed56e collection. A new experiment needs a new plan/version preserving historical records. Keep executables out of Git; record build recipes/hashes. Require real source/input Git commits and preserve the same executable between preparation/audit.

CGO_ENABLED=0 go build -trimpath -buildvcs=false -p=1 -o PRIVATE_BINARY ./cmd/riido-propertyaudit
PRIVATE_BINARY --stage prepare --source-commit SOURCE_COMMIT --out NEW_PREPARATION
# Freeze dataset.json and plan.json bytes in a separate input commit, then:
PRIVATE_BINARY --stage audit --input-commit INPUT_COMMIT --in NEW_PREPARATION --out NEW_AUDIT

The official run also used the environment in the frozen build recipe. Preparation emits dataset/plan with zero candidate execution. Audit requires equality with actual Git blobs at experiments/short-claim/probes-56e.json and execution-plan-56e.json. Existing output directories, tampered data, changed executables and bad input commits refuse. The CLI does not prevent repetition into every new directory, so maintain an honest official-attempt/failure ledger.

Interpreting results

Raw evidence includes full expected/got observations in source_results. rows preserves original candidate order and evidence indices. matches_finite_literals means agreement on listed inputs only, not other inputs, general function truth or model probability. finite_no_answer does not instruct removal or reduced checking: preserve candidates and fallback.

This is truth material for subsequent utility/training planning. Training eligibility false, fits/models/rankings0 and new independent parents0. Read English results, Korean and next public-source contracts for scope and remaining work.