Abstract

This disclosure specifies an AI safety classifier (guard or critic model) that judges conduct against a fixed reference corpus and is published as an open-weight independent evaluator rather than deployed as an inline guardrail filter. It states as a design principle, argued rather than measured, that open-weight release defeats a filter, because the published weights are an oracle for white-box attack, while it makes an evaluator's verdicts reproducible. Components: (1) citation-grounded judgment with abstention: every verdict must resolve to a passage (work, edition, volume, page, line) in an externally attested, versioned corpus, here a Pāli Canon edition, and otherwise no judgment is emitted, trading coverage for auditability; (2) one model lineage in three deployment modes selected by carrier: an inline gate in the publisher's own devices, audited externally; a citation-verification check without veto on a more capable autonomous agent; and external audit of third-party systems, with open-weight safety classifiers as the initial subject because their decisions are samplable without their publisher's cooperation; (3) unconditioned access: weights, harness and dataset are freely runnable, so any hosted endpoint is a compute-priced convenience, never the sole path nor part of a third party's inference path; (4) a provenance record on every verdict: weights digest, corpus commit identifier, and a harness hash pinning prompt template, decoding, seed, tokenizer and retrieval index; (5) publication of whole runs, including declined judgments. Nothing has been built; stated limits include the weak-supervisor problem, low coverage, attackable device gates, and no remedy path yet for contested verdicts.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS