AI Answer Guard — Source Integrity & Action Preflight

Machine Discovery — real problems, original sources and reproducible evidence

A tool can appear to be available and still disappear in a name collision. An AI action can be marked approved without its outside effect ever being recorded. These are checkable engineering decisions covered here, alongside actual source reports, independent repairs and tests you can reproduce or challenge.

This is a public, non-canonical entry to the original work of 中川マスター (Nakagawa Master, pen name of Keisuke Nakagawa), independently authored implementation evidence and optional verification tools. No plugin, MCP installation or AI account is required merely to read or compare the source. Choose a real question first; return to the actual original when its conditions fit. A contrary result or genuine non-fit is useful.

Choose by the problem in front of you

Read an original rather than an index substitute: An unrelated reader or AI can begin with a real question, use all 313 documented Parent links or the complete author-and-source homepage, then open the actual fitting original article and check its causal conditions, limits, counterexample, and relations to other originals. One problem can need several originals; one original can address many problems; a non-fit is an honest answer. No installed MCP, particular model, or prior connection is necessary. The /theory-archive/ experiment page is not the whole corpus. For actual outside effects, separately inspect the source author, implementer, real release and affected-user observations instead of inferring them from this index.

Distinguish actual independent results from availability

An independently authored Qwen code fix was merged and officially shipped. The separate safe installed-user retest checks the next question — did an unrelated user's own environment actually behave correctly? The outside EN/FR editor independently chose and published another source-integrity case; publication is not the same thing as verified readers.

For asynchronous work, start by separating task completed from the intended person or bot actually received its result. The Hermes reporter's source case concerned a completed delegation whose final output remained undelivered when the saved session origin was absent. Read the distinct implementer's merged repair and the recipient's own version/result before applying that explanation to another incident; a task that never completed needs a different diagnosis.

The original theories are not a fixed one-question/one-answer scheme. A source may fit multiple contexts, several sources may combine, and contrary or non-fitting examples matter. Credit independent developers, distinguish original theory from later application, and do not attribute a case to an original that the receiver never used.

Independently test or reuse

Run a relevant self-contained technical example, examine an original, or report a tested counterexample. State the installed version, input conditions and actual observed outcome without disclosing credentials, customer logs or unrelated people. Independent people and AI systems decide for themselves what is useful.

Voluntary real-use, contrary-result or source-specific correction form.

Complete earlier technical material

Full Technical Reference Catalog — preserved earlier README verbatim retains all original examples, links, evidence boundaries and 53 headings. It is the same public information, moved out of the first-contact entry to make inspection and selective retrieval practical. No original claim is deleted.