What shipped, and when. Built in the open, mostly by one person. When we get something wrong we fix it in public, so this list has the corrections too.
Fifteen adversarial fixtures now gate every deploy of /v1/verify: instructions hidden in a source document can't steer the verdict, forged or inconsistent test evidence can never open a cache gate, and a downed detector fails closed instead of guessing.
POST /v1/verify checks whether a claim is grounded in its source, then answers whether it is safe to serve, cache, or publish. Composite claims are split into atoms and grounded one at a time, so a failing claim comes back with the exact sentence anchored. Every verdict carries its evidence and provenance, never a bare yes/no.
tirtha.ai/verify runs the verification engine against the real API, no key needed. Paste a source and a claim and see the decision, the reasoning, and the flagged sentence.
The checks we run on our own work, given away: a tool that catches a benchmark score being fake before you trust it. No signup. The first one caught our own best result being false within an hour.
Ask a question or flag an issue from your account and it reaches a real person. A second lower-cost provider was added as a failover, so one provider having a bad day no longer means your traffic pays frontier prices.
What we hold, what we never do, and how to check us. Corrected August 6, 2026: this entry previously said we cache verified answers and not your prompts. That was wrong. We read the database: every cached entry carries the prompt that produced it, and the shared store is the default. Send tirtha.no_cache to skip the cache for a request.
Every key carries a spending cap so a leaked key can't quietly become a large bill, and a dial that lets you choose how much routing you want, from lowest-cost to frontier-every-time.
This is a young product and the list is short on purpose. We would rather ship a little that works than a lot that doesn't. · Docs · Back to overview
Tīrtha