CFLAG and CCHECKBAD: a memory for the answers that were wrong
Most caches only remember good answers. Crowkis also remembers bad ones — flag a hallucinated or harmful response once, and CCHECKBAD catches every paraphrase of the question that would have reproduced it.
A cache that only stores good answers has half a memory. When a model hallucinates a wrong answer — a made-up API, a fabricated policy, a confidently incorrect fact — you fix it once and then wait for it to happen again to the next user who phrases the question slightly differently. CFLAG closes that loop by recording the bad answer in a negative cache, with an optional reason.
CCHECKBAD is the read side: it asks whether an incoming query matches a flagged-bad entry, and if so returns the matched question and the reason it was flagged. Because the match is semantic, flagging one phrasing of the bad question inoculates against all of them — the paraphrase that would have regenerated the hallucination is caught too.
Five stages score every write before it can ever be served.
This turns human review into durable protection. A support lead who spots a wrong answer flags it; from that moment, every semantically similar question is intercepted before it can reproduce the error, with the reason attached so the next reviewer understands the history. The negative cache is tenant-scoped, so one team's correction doesn't leak into another's traffic.
The bottom line
Knowing what's true is half of intelligence; remembering what was false is the other half. CFLAG gives the cache an immune memory — the answers it has learned to refuse.