nx_extllm_publish_guard.nx

nx_extllm_publish_guard.nx -- RUNG XL8: what we may SAY about a provider we are allowed to USE. This stopped being hypothetical the moment the estate started making real calls. The OpenRouter Stealth AUP forbids publicly disseminating confidential technical information regarding the performance of the models; Cerebras prohibits benchmarking or competitive analysis; SambaNova prohibits publishing capability or performance results without written consent. We now HOLD such numbers -- a 200, a 429, a wall time -- and /compare publishes to the open web on a beat. Nothing stood between those two facts until this organ. THE RULE, three decidable conjuncts on a single line: (1) a token identifying a provider whose terms row reads benchmark_publish_restricted=YES, (2) any metric token from knowledge/extllm_metric_vocab.conf, and (3) a run of two or more digits. The identifying tokens come from the ledger itself, never from a list in this file: the provider name, plus the tier when the tier is written as a model id (it contains a slash, e.g. stealth/ox-alpha). A tier like "free" or "trial" is NOT used as an identifier -- it would match half the corpus and a detector with false positives everywhere is one everybody learns to ignore. THE IMPRECISION IS DECLARED, in the vocab conf and here: a performance claim in pure prose carries no metric token and PASSES, so this is a FLOOR and never a proof of compliance; and a capability row citing a published rate cap trips all three conjuncts and is REFUSED. That direction is deliberate -- an AUP breach is not undoable once a page is served, so the guard fails toward making a human look. exit 0 ALLOW / 1 REFUSE / 3 UNKNOWN / 2 usage. license_tier: ORIGINAL. No hw writes (Rule 26).