Typed Behavior Usage 56b EN

Using the behavioral truth auditor

한국어 · Pre-observation plan · Results · Runtime hint tool

riido-typedaudit is a maintainer Go tool for preparing learning evidence. It compares observed values, error identity, input mutation, memory ownership and acquire/cleanup ordering against independent finite literal tables. Incomplete code descriptions remain unknown. Requests sharing sources, candidates, counterexamples or helpers are connected transitively.

The auditor is not a model router or ranking model. Runtime hints belong to the separate riido-shortclaim tool. This audit downloads or trains no model and changes neither Codex integration nor default settings.

Build from the repository root using Go1.27.1. No Python or API key is required.

mkdir -p bin .cache
CGO_ENABLED=0 go build -trimpath -o bin/riido-typedaudit ./cmd/riido-typedaudit

Preparation writes a new fixture from authored requests fixed in source. It executes no candidates, group audit or ranking. Use a new directory name: an existing output directory is rejected.

./bin/riido-typedaudit --stage prepare --out .cache/typed56b-preparation-example

probes.json contains original text and fixed source/bundle digests. There is no automatic label generation for arbitrary work code. Known labels require the exact reviewed request/candidate captions and source bindings. Fitting the caption budget does not establish semantic completeness.

The results document provides the official replay command and exact plan digest. Auditing verifies the plan, source manifest, compiled sources and both input files, then decodes the same verified input bytes. It writes legacy48-request truth, new24-request truth and combined72-request relations to a new results.json, without overwriting existing files or output directories.

whole_group_gate_pass differs from labeled_group_gate_pass. Unknown-only groups retain their relation edges but cannot satisfy the labeled minimum. Even if both pass, training_execution_ready remains false: this plan assigns no roles and authorizes no fitting, weights, model calls, final evaluation or production activation.

CPU execution uses1 thread with a256MiB Go-heap soft limit. This is neither a total-process RSS/GPU memory cap nor a measured inference result. Distinguish offline go/types auditing, which needs the Go installation's standard-library type metadata, from the small fixed-array runtime hint tool. Do not put this maintainer auditor in the path for repeated low-cost production requests.

People can read public unknown reasons and group relations. Agents can inspect fixed error codes, exit status and JSON. Errors do not echo input, caller-selected personal paths or code. Raw profiles, real work inputs, credentials and weights do not belong in Git.