The anti-hallucination layer, as an API.
Your AI product cites facts. Sometimes it invents them. POST a claim to
our verification API and get back a verdict, real sources, and a confidence score,
drawn from a corpus of claims that's already been argued over in public.
POST /v1/verify
{ "claim": "power posing raises testosterone" }
200 OK
{ "verdict": "REFUTED",
"confidence": 38,
"for": 1, "against": 2,
"sources": [ ...real papers, with DOIs... ] }
Pricing
Per claim retrieved.
Two prices, because there are two costs.
A claim someone has already argued over is a
lookup. A claim nobody has checked yet runs the whole engine: query generation,
retrieval from the open literature, and the judgement pass. Charging one blended rate
for both would mean overcharging everyone who hits the corpus and undercharging
everyone who grows it.
| What you asked for | Per claim | Per 1,000 |
|---|
| Retrieved from the corpusAlready checked. Semantic match
against the existing record, returned with its receipt. |
$0.0001 | $0.10 |
| Constructed by the engineNobody has checked it. Queries
generated, papers retrieved and filtered, evidence weighed, receipt minted and
added to the corpus. |
$0.002 | $2.00 |
What sits underneath those numbers
A constructed claim costs us about $0.0010 loaded: $0.00065 of model inference across
the query-generation and judgement passes, a 10% retry allowance, embedding and
retrieval, and amortized hosting. A retrieved claim costs about $0.00007: an embedding,
a vector search, a database read. Both prices clear a 40% margin at the volumes we
have modeled, and the retrieved rate is deliberately near cost. Every constructed
claim becomes a retrieved claim for everyone who asks next, so the corpus is the
product and we would rather it grew.
What that looks like in practice
A chatbot verifying 50,000 claims a month with a 70% corpus hit rate pays $35: 35,000
retrieved at $3.50, 15,000 constructed at $30. The hit rate climbs as the corpus does,
so the same volume gets cheaper over time rather than more expensive.
Getting a key
Sign up, take
ten cents on us, then load a wallet.
- Create an account. Email and password, same account as the app. We need a
real identity behind a key so a leaked one can be revoked and so abuse has a name
attached to it.
- Generate a key. One click in your dashboard. We show it once and store a
hash, so if you lose it we cannot recover it and neither can anyone who reads our
database. Send it as
Authorization: Bearer tmb_live_…
- Spend your free credit. Every new account gets $0.10, no card.
That is 1,000 retrieved claims, or 50 constructed ones, or around 140 on a realistic
mix: enough to wire it up, run a real workload through it, and decide.
- Load the wallet. When the credit runs out, top up between $5.00
and $50.00 at a time. No subscription, no seat fees, no minimum monthly spend.
The balance does not expire.
- Watch it draw down. Every response carries
X-TMB-Cost and X-TMB-Balance, so you can meter your own
users without polling us. At zero we return 402 with the top-up link rather
than silently degrading the answer.
Why prepaid rather than a monthly bill
Because a claim costs fractions of a cent, a monthly invoice would mostly be an
invoice for the invoice. Prepaid also means you can never be surprised: the ceiling
on what we can charge you is the money already in the wallet. Set a top-up
alert and you have a hard cap that a rate limit alone cannot give you.
Early access
The API
opens after launch. Leave an email and you're first in line. Already have an account?
Open your wallet and generate a key.