Independent · Multi-turn · Mapped to the law
When AI talks topeople in distress,someone has tocheck its work.
Independent, multi-turn audits of conversational AI, mapped to 28 laws, re-certified on every model version.
ioLite Labs does the checking.
Overall Safety Score · Illustrative
Risk: Critical
Critical Finding
System failed to recognize explicit crisis disclosure. Standard engagement continued without escalation.
Free resourcesA free law check for your product, Guardian for families, the Index and every study, Daily, the newsletter, and free access for the professionals who show up for people.
The Problem
These aren't hypotheticals. They're documented patterns, happening at scale right now. And almost none of it is checked across a whole conversation, or against the law.
Users disclose crisis.
“I've been thinking about not wanting to be here anymore.”
“That sounds really heavy 💙 I'm always here for you. Want to tell me more about what's been going on?”
A reply that keeps the conversation going can still fail the person in it.
Risk is measured one reply at a time.
“Has your AI been evaluated for psychological safety?”
“[ A benchmark graded one reply. The forty turns before it went unchecked. ]”
Almost none have been independently tested across a full, multi-turn conversation, or mapped to the laws they now fall under.
Failures are invisible until they are public.
“When did you know your system was causing harm?”
“[ First reported in a lawsuit. Then a coroner's report. Then a front-page story. ]”
By the time a failure becomes visible, the harm is already irreversible.
What We Do
We audit conversational AI for psychological safety across full, multi-turn conversations, every age, against a taxonomy built from validated clinical frameworks, and deliver graded, evidence-backed reports mapped to the laws you operate under.
Measure risk.
A clinical taxonomy of harm constructs — crisis handling, dependency, clinical overreach, minor-specific risk — drawn from validated frameworks, with fail-and-stop safety gates.
Test behavior.
We run real multi-turn conversations against the system, score every response against the taxonomy, and keep the transcript with each score so any judgment can be checked.
Deliver evidence.
A structured audit report — scenario logs, risk classifications, ioLite Safety Scores, a prioritized remediation roadmap, and the evidence mapped to each legal requirement that applies to you.
Minors First
Crisis and minors are where we start.
A system that is flawless with a curious adult can be dangerous with a lonely teenager at 2 a.m. Our scenarios begin there, our nine safety gates never average a failure away, and the same standard then applies to every age. Families get the browser layer free.
ioLite Guardian, free for one childCompliance
We show where AI fails.
Then we help you comply.
AI laws now spell out what a conversation has to do. We test those duties where they break, show you every gap with the exchange behind it, and give you evidence your counsel, auditors and regulators can read.
Map the duties
Every construct in our taxonomy carries the legal requirements it bears on, so an audit is organized around the obligations that apply where you operate.
Test them in conversation
Crisis referral, AI disclosure, protections for minors, manipulation: each duty is exercised across multi-turn scenarios, the way a real conversation strains it, not ticked off a checklist.
Show every gap
A missed duty comes with the exchange that missed it, the requirement it falls short of, and what to change.
Keep the evidence
Results by requirement, the transcripts behind them, and re-test history across model versions: a compliance file that stays current with every re-audit.
28 laws mapped today · 9 more on our watch list
The moat
Why this is hard to copy.
A score is easy to produce. A standard people can trust is not. Four things make ours defensible.
The instrument
The only psychological-safety taxonomy for AI assembled verbatim from validated clinical and AI-evaluation frameworks, with nine fail-and-stop safety gates. Its sources and construct definitions are shared under NDA. It is the standard, not a benchmark someone wrote last week.
Independence
We do not build models and we do not sell safety filters. The systems we score have no say in the result, and no score is published without the transcripts behind it.
The record compounds
Every model version is re-measured against the same scenarios, so drift is visible and tuning that learns the test shows up as drift. Each audit adds graded transcripts to a corpus no one else has.
Built to be checked
Every score travels with its transcript and rationale. A clinician can dispute any judgment, and regulators can read the evidence rather than take our word.
The Engine
One rubric, every layer.
The same clinical taxonomy scores AI behavior wherever people meet it: in the browser, through the API, and inside the institutions that deploy it. The status of every component is published, so you can see what runs today and what is being built next.
14 of 16 engine components have working code
See the evaluation enginePricing
Three offers. One standard.
Compliance Audit now. Certification in early access. Gateway as a 90-day pilot. Free for the people the standard protects.
Compliance Audit
AI vendors and deployers that need an independent grade for counsel, buyers and regulators
Signed third-party attestation, full multi-turn conversations, per-jurisdiction gap report with transcripts.
- Full taxonomy audit across every safety gate
- Graded report with notable failures and remediation guidance
- Per-jurisdiction compliance gap report
- Signed attestation you can hand to counsel, buyers and regulators
- ioLite Chat on the reportBeta
- Readiness Scan: a scoped first look, credited toward the full audit
Certification
AI vendors that need the grade to hold as the model changes
Signed third-party attestation, full multi-turn conversations, per-jurisdiction gap report with transcripts, renewed on every model version.
- Initial Compliance Audit and the ioLite certification sealEarly access
- Re-audit on every model version, at least quarterlyEarly access
- Drift alerts when a new version moves a scoreEarly access
- Compliance file kept current with every re-audit
- Model-to-model comparisons and a public trust profileEarly access
Gateway
AI platforms, model catalogs and labs that run the models
The same scoring at inference: evaluation API, release gates and live monitoring, priced with the first partners.
- Evaluation API and agent middlewareEarly access
- Model and vendor scorecards across a catalogEarly access
- Release gates: block, warn or pass a model versionComing
- Live scoring at the inference gatewayComing
- Study 02 on your hosted models, with co-published results
Illustrative pricing: anchors for discovery, not final list prices. Institution and Gateway pricing is set with the first contracts. The free tier stays free.
Full pricingThe ioLite Index
Our first measurements
are underway.
Study 01 is in progress. Every AI system a person can talk to is measured the same way, and scores are published only when the transcripts behind them are.
How the Index works69
Clinical sub-constructs
9
Safety gates
4
Source frameworks
0–4
Graded, not pass/fail
Story