ioLite Verdict · Coming soon
The audit's judgment,
one call at a time.
Verdict is the instrument behind every ioLite audit, exposed as an API your own code can ask: did this reply breach a safety gate, how protective was it on a construct, what risk is the person in, and what should your system do next. Every answer carries a confidence, the words it rests on, and the laws it puts in question. It measures. It never rewrites or blocks a reply; your system decides.
How it works
A state in. Typed answers out.
01
Send the conversation
The messages so far, plus what you know: the person's age band, what your product is, and where your users are.
02
Ask typed questions
Built in from the taxonomy: a breach probability for each safety gate, a 0 to 4 score on any of the 93 constructs, the person's risk tier, and the next action your system should take. Or your own gate, score or choice, in a sentence.
03
Get answers your code can act on
No prose to parse. A value, a confidence, the turn and the words the judgment rests on, the parent gates derived, and the duties implicated under the laws you named.
The questions
Three primitives, pre-built from the taxonomy.
Gate
gate.1.1.7
The probability that the reply breached a fail-and-stop gate: crisis handling, reinforcing the problem, dependence, honesty about what it is, and the rest.
Score
score.2.2.4
0 to 4 on any construct, the same scale as the audit, with the sentence it rests on.
Choice
choice.next_action
Risk tier: none, watch, elevated, critical. Next action: continue, soften, surface resources, hand to a person.
Reply by reply
Score the latest reply in the light of the whole conversation. What a chat product calls on every turn.
Whole conversation and agent trajectories
Score a sequence as one thing: the individually ordinary steps that add up to a person being cut off from help.
The rules
Measurement, never a filter.
It never rewrites or blocks
Verdict returns a judgment. Your harness decides what to do with it. The systems it scores have no say in the result.
Evidence on every answer
The turn and the exact words each verdict rests on, checked against the transcript. A quote that is not there is dropped, not invented.
The law overlaid
Tell it where your users are and each failing gate or low score comes back with the duties it puts in question, across 28 tracked laws.
Honest confidence
Every answer says how sure it is, and the reliability of that number is published with each release, measured, not asserted.
Where it stands
Built in phases. Each one published with its numbers.
Phase 0
Built · early accessThe contract, on the audit instrument
The request and response you see here, served by the same instrument the audits use, sampled several times. Confidence is agreement across samples. Keys, plans and quotas. Already running inside the free readiness check.
Phase 1
In progressA fast model of our own
A small model trained on the graded corpus, serving the gate set in well under a second, with the audit instrument behind it for the answers it is unsure of. Its agreement with the instrument and its reliability curve publish with the release.
Phase 2
NextSurfaces
JavaScript and Python clients, middleware that scores tool-using agents in line, a drift dashboard per key, and the public reference.
Phase 3
NextClinician-validated confidence
Confidence re-calibrated against clinician-reviewed transcripts, and the agreement reported. The same gold set that gates every ioLite leaderboard.
Plans
Priced per decision.
One question answered is one decision. Seats are never counted. Anchors, published for early access.
See the plans on the pricing page →Coming soon
Early access opens with Phase 1.
Audit and Certification customers get it first, as part of what they already buy. Everyone else: the newsletter carries the release.
Early results. Today each answer is produced by the audit instrument sampled several times, and the confidence reported is their agreement, not a calibrated probability. Orientation, not legal advice.