featuresJune 9, 2026· 3 min read

CPIN: golden answers that are served verbatim, with an audit trail

For the questions where 'close enough' is unacceptable — pricing, legal, brand lines — CPIN serves a human-approved answer verbatim, records who approved it, and never lets the model improvise.

Some answers cannot be paraphrased. The exact refund policy, the regulated disclosure, the brand's one approved description of itself — for these, a semantic cache's 'similar enough' is exactly the wrong instinct, and a freshly generated answer is a liability. CPIN carves out a verbatim lane: a human-approved answer that's served word-for-word when a query matches.

In plain words: For answers that must be exact — pricing, legal, brand lines — Crowkis serves a human-approved version word-for-word and logs who approved it, so the AI never improvises on the sentences that matter.

The mechanics are deliberately strict. CPIN stores the approved answer with a BY field naming who pinned it; CPINGET checks whether an incoming query matches a pinned answer and returns it with that attribution; CPINLIST gives an auditable inventory; CUNPIN removes one. The match still uses semantics so paraphrases of the question hit the pin — but the answer never varies.

the write-trust pipeline

Five stages score every write before it can ever be served.

Pinning sits ahead of generation in priority, which is the point: a pinned question never reaches the model, so there's no chance of an improvised legal answer or an off-brand description slipping out. The audit trail means that when compliance asks 'who approved this language and when,' the answer is a query, not an investigation.

The bottom line

It's the feature that lets a probabilistic system make deterministic promises about the answers that matter most. The model handles the long tail; CPIN handles the sentences a lawyer wrote.