Security and evidence

Cintian / Security and evidence

An invention disclosure is the most sensitive document a company writes. Where it goes matters.

A foundation model is the right tool for reading evidence and drafting. It is the wrong tool for finding the evidence, and the wrong place to send an unfiled disclosure as a search query. Cintian does the retrieval on its own models. What goes to a foundation model afterwards is the retrieved record, on your terms.

Our models for retrieval. Your choice of model for everything else.

Multiple vertical models trained on patent text find and rank the evidence. No third-party model sees the query to do that. The quoted sentences then go to whichever foundation model you use, inside your environment if that is where it runs.

Your disclosure stays yours

Nothing you search is retained, logged for training, or used to improve a model. Unfiled material is handled as unfiled: confidential, and gone when the search is done.

Reproducible

The index and the retrieval models are versioned. A search run today returns the same evidence in two years, whatever model drafted the report. An FTO opinion or a licensing position has to be defensible after the fact, and the part that has to hold is what was found.

On your infrastructure if you want it

The Vectors tier puts the corpus and the index inside your environment, next to your own models. The query never crosses your boundary. Where Cintian drafts the report, the drafting model and where it runs are agreed with you in the engagement terms.

A foundation model as the search engine

Cintian retrieval, then your model

Your disclosure is the query, sent to a third party. Retention, logging and training terms are theirs, and they change.

Retrieval runs on Cintian's own models. What reaches a foundation model is the retrieved evidence, where and how you decide.

Answers from what the model recalls. The patent record is consulted only if the wrapper bothers to.

Answers from the record. The model reads quoted sentences with their sources, which is what it is good at.

General-purpose vocabulary doing the search. Weakest on exactly the technical terms that carry the invention.

Patent-trained models do the search, routed to the vertical specialist. The foundation model reads the result.

What was found changes when the upstream model does. Yesterday's search cannot be reproduced.

Versioned index and retrieval models. The same search finds the same evidence, whichever model you draft with.

The vendor's interface holds your results. You rent access to the layer.

Vectors you can hold. APIs you call from your own tools. Evidence your own models can use.

Evidence

Measured as retrieval performance. Stated with its limits.

Performance is tested at three levels, each proving a different thing: training coverage across the corpus, coherence of the embedding space on withheld passages, and the end-to-end pipeline against known closest art on independent searches.

The holdout withheld passages, not whole documents.
Recall measures one known target per search, so a family sibling retrieved in its place counts as a miss. Recall is understated.
Retrieval performance is not relabelled as hallucination reduction. That was never measured.

Evidence summary

Retrieval performance on independent searches, with held-out validation and every caveat stated, in a one-page summary sent on request.

Request the evidence summary