Capability
Grounded verification: catch the citation that does not exist
The most damaging thing an AI system does is not give a rude answer. It gives a confident one that cites a source which was never real. Grounded verification is Turrigan's accuracy-assurance layer for exactly that failure. It checks the references an output makes against a source of truth you supply, and it does so inline, offline, and without ever pretending to know more than it does.
The failure it targets
A fabricated URL, a wrong DOI, a case number that looks right and does not exist. In law, medicine, finance, and government, a made-up reference that reads as authoritative is worse than no answer, because it is trusted. This is the class of error that turns one AI output into a headline, and it is invisible to a system that only checks tone and safety.
What it verifies today
Turrigan's first grounding capability is reference-existence. It checks that the URLs, DOIs, and optionally the IDs an output cites actually exist in a list of known-good sources you provide.
- Not-found escalates, it never blocks. A reference outside your list routes the output to a human. It is never blocked and never labelled fabricated, because your list is partial by nature, so "not on the list" means "we could not confirm it", not "it is fake".
- Precision first. It will not flag a reference it did not confidently parse, so it does not manufacture false alarms.
- Silent until you configure it. With no source list supplied, the check is inert and says nothing.
Honest by construction
Turrigan never asserts open-world truth. It affirms only the references you stand behind. You point it at a source of truth, a simple list of the URLs, DOIs, and IDs your assistant is expected to cite, and Turrigan verifies against that.
That source list is hashed into the tamper-evident evidence, so an auditor can later confirm which list a given decision was checked against, while the list contents themselves stay out of the stored record. The check is deterministic, so the same output and the same list always produce the same verdict.
It never blocks, never slows, never phones home
Reference-existence runs inside the core, deterministically and offline. There is no network call and no model in the check, so your AI output is never sent to an external service to run it, and there is no per-token cost. The whole add-on is optional and off by default. With it off, the core product is byte-identical to today. When it is on, it adds an advisory check that can escalate a decision to a human, and it can never override the core into allowing something the core would block.
On the roadmap
We will not describe these as available before they ship. They are where grounded verification is going, and they extend the same principle.
- Faithfulness checking (next). Verifying each claim in an output against the context your own system retrieved. This runs in a separate Verifier sidecar built on local natural-language-inference models: CPU-only, no external API, no internet, and self-hostable in your environment for zero-egress and sovereign deployments. The sidecar is isolated by design and can never block or slow the core. It is reached through a circuit-broken adapter with a hard timeout, so a sidecar that is down, slow, or absent simply means the extra check is skipped. The default is fail-open, and you can choose fail-closed per tenant so an unavailable check escalates instead.
- Attribution (after that). Verifying that a cited source actually supports the claim it is attached to, including against your own knowledge base.
Frequently asked questions
Does grounded verification send my AI output to an external service?
No. Reference-existence checking runs in the core, deterministically and offline, with no network call and no model. Your AI output is not sent anywhere to run it.
Will it block my AI when a citation looks wrong?
No. A reference that is not in your source of truth is escalated to a human for review. It is never blocked and never labelled fabricated, because an allowlist is partial and "not on the list" means "could not confirm", not "fake".
What does grounded verification check today?
It checks that the URLs, DOIs, and optionally IDs an output cites exist in a source-of-truth list you supply. Faithfulness checking, meaning each claim against the context your system retrieved, is on the roadmap and not yet shipped.
Does turning it on change the rest of Turrigan?
No. The add-on is optional and off by default. With it off, the core product is byte-identical to today. It adds an advisory check that can escalate, and it can never block or slow the core.
An assurance layer that inflates its own coverage is worse than none. So grounded verification tells you precisely what it confirmed, what it could not, and what is still on the roadmap. That honesty is not a caveat on the product. It is the product.