Legal

Terms, corrections and right of reply

This registry is built to publish findings about named companies once it is safe to stand behind them. It does not yet: what is below states who would publish them, what a score does and does not mean, and how to check or correct anything that already appears — under a pseudonym, not a name.

Findings about a named company are not published here. What is published is the method, aggregate figures over the whole set, and rows shown under pseudonyms — none of which asserts anything about an identifiable party. Findings attached to a named company are shown to the operator only. The entity named below publishes this site; findings stay withheld regardless, because professional and media liability cover is not yet bound — a decision about risk, not about whether any finding is true.

Nothing below was drafted by a lawyer. It states accurately what the software does, which is what a lawyer works from and not a substitute for one.

For information only. Not advice, and not to be relied on.

Everything here is published for general informational and educational purposes. It is a description of what documents say, not a recommendation about what to do.
Nothing here is legal, compliance, regulatory, financial, investment, procurement or safety advice, and reading it creates no advisory or professional relationship of any kind.
Do not act, or decline to act, on the basis of anything here without your own verification. Every claim names the document it came from and the date it was read, so that verification is possible rather than merely recommended.
Information ages. A score describes documentation as it stood on a stated date and may not describe what a vendor publishes today.
It is one input among several. It is not a substitute for your own diligence, your own testing, or advice from someone qualified and accountable to you.

What this section does, and what it does not. A disclaimer of reliance answers a claim that someone was entitled to act on this and was harmed by doing so. It does not answer a claim that something here about a named company is false — no wording makes an untrue statement acceptable, and nothing in this section is offered as though it did. That risk is answered elsewhere and differently: every claim names the document it came from, its date and a fingerprint of it; the tiers separate what a vendor asserted from what someone independent verified; and anyone can have a finding corrected. The two protections cover different things and neither substitutes for the other.

Who publishes this

Publisher: Rtavan AI LLC (NC SOSID 3388209)
Governing law and venue: North Carolina, United States
Contact: [email protected]
Corrections: [email protected]

What a score is

A GRI score measures the documentation this registry located about a system at a stated URL on a stated date, against a published method. It does not measure the system.
A low score means the located documentation is thin. It is not a finding that a vendor governs its models poorly, that a system is unsafe, or that anyone has broken a law or a rule.
Documentation a vendor holds privately, publishes elsewhere, or supplies to customers under contract is not counted, because this registry cannot see it.

The scoring rule is arithmetic, published in full at /method, and applied identically to every entry. Where a number is an evaluative judgement — that a given piece of evidence sits at one tier rather than another — the underlying fact is published next to it, with the source and the date, so a reader can reach their own conclusion from the same material.

What this is not

Not an audit, an assurance opinion or an attestation.
Not a certification, accreditation or approval.
Not a credit rating or any other regulated rating.
Not investment, financial, safety or legal advice.
Not a statement that any system is unlawful, unsafe or non-compliant.

No relationship of any kind is implied between this registry and any company named in it. Product and company names are the marks of their owners and are used to identify the systems described, which is what a registry is for. Nothing here should be read as an endorsement, a disparagement, or a recommendation to buy, avoid or dispose of anything.

How a finding can be checked

Every published claim is meant to be re-derivable by someone who does not trust us.

Source and date. Every scored claim names the document it came from and the date that document was fetched.
A fingerprint of what was read. Each scanned entry records the SHA-256 of the exact document scored. If the document later changes, the registry can still show what it read, and so can you.
A versioned method. Scores are comparable only within a method version and a weighting profile. Both are shown wherever a score is.
Scope, applied as a rule. Only systems published by organisations are scored. Individual accounts and test namespaces are excluded — determined mechanically, not case by case, so no entry can be said to have been singled out.
Nothing invented. The registry holds 71 systems and 1673 claims, of which 0 are fabricated. Seven invented demonstration systems were removed in migration 018; the counter in the site header is computed, so it reports a fabricated entry if one is ever inserted again.

Corrections and right of reply

If something here about you or your product is wrong, we want to fix it, and quickly.

Use the request a correction form on the entry itself, or write to [email protected].
You do not have to be the vendor. Anyone who can point at a factual error may raise one.
We aim to respond within 10 working days. Where a claim is shown to be wrong it is corrected at once, and the correction is published in the corrections log with its date.
Where we review and do not change a finding, that is published too, with the reason. A log that records only the corrections we were happy to make would be worth nothing.
Pointing us at documentation we could not locate is the fastest route to a different score, because a score is a measurement of located documentation. Publishing that documentation changes the number; asking us to change the number does not.

No correction has been published yet. The log is at /corrections.

Personal data

The registry describes systems published by organisations, and holds the organisation account names under which they were published. Accounts belonging to individual people are outside its scope and are excluded automatically. If you believe personal data about you appears here, write to [email protected] and it will be removed unless there is a clear public-interest reason to keep it, which we will explain.

Requests to the router are logged as a salted hash of the prompt, never the prompt. Inbound enquiries store the address you give and a salted hash of your IP, kept to throttle abuse.

Requests to this site

Since 13 September 2026, each request to a page or API on this site writes one line to a server log. It exists to show which pages are read, and to tell people apart from crawlers and from automated attacks, which make up most requests to any public site.

Each line records:

the time, the request method, and the path requested — with the credential in an invitation or examiner link replaced, so a link that grants access is never written down;
what kind of request it was: a page, an API call, a background fetch made by the page itself, or a request for software this site does not run, of the kind automated attacks send;
of the query string, only profile and state, which change what a page shows and say nothing about you;
a broad category for the software making the request — a browser, a search or AI crawler, a link preview, or another automated client — and, where a crawler names itself, that name; and, for a request claiming to be Googlebot or Bingbot, whether its address matched the ranges those companies publish (the address itself is not kept);
the site a link was followed from, as a hostname only; for a link followed from another page of this site, the path of that page;
the country Cloudflare reports for the connection;
a visitor identifier: a salted hash of your IP address and browser identification string, which changes every day.

It does not record:

your IP address, or your browser’s full identification string;
cookies, form contents, or any other query parameter;
anything you paste into /scan or /before-you-paste, which run in your browser and send nothing to this server.

The visitor identifier lets a day’s visits be counted without following anyone from one day to the next. It is pseudonymous, not anonymous: combined with other records it could in principle be linked to a connection, which is why it is kept for a limited time and read by no one but the operator of this site.

The log is rotated when it reaches about 2 MB, and only the twenty most recent files are kept; older ones are deleted. At current traffic that is roughly two months, but the limit is the number of files, not a date. No analytics cookies or third-party tracking scripts are used.

Every request also passes through Cloudflare, which carries traffic to this site and processes connection data, including IP addresses, under its own terms. To ask about any of this, write to [email protected].

Terms of sale

Covers the Practitioner and Firm subscriptions sold at /pricing. Enterprise runs on terms written for the engagement, agreed separately.

Contracting entity. A subscription is a contract with Rtavan AI LLC (NC SOSID 3388209), not with any individual.
What a subscription grants. A console tenant for the paying organisation — its own policies, keys and users, the policy router with the hash-chained evidence log and exports, examination packs, and the document corpus with its change history — for as long as the subscription is active and paid. It does not grant a licence to any content published elsewhere on this site, which stays free to read regardless of subscription status.
Provisioning. Paying through Stripe Checkout starts a subscription; it does not itself create a tenant. Access is granted by hand in the console against the Stripe payment record, usually the same day the payment is recorded. An entitlement is never granted without a note saying how it was settled.
Term and renewal. Monthly subscriptions renew monthly; annual subscriptions renew annually, at the price then in effect for new terms. A price change never changes what you are already paying for the term you are in — it applies only to the term that starts at the next renewal, and Stripe notifies you before a renewal charge.
Cancellation. Cancel any time from the Stripe customer portal or by writing to [email protected]. Cancelling stops the next renewal; access continues until the end of the period already paid for, then the tenant is disabled. Cancelling does not delete the evidence already generated inside it — a separate request, handled the same way as any other data request under Personal data above.
Refunds. No refund is given for a period already underway once a tenant has been provisioned, except where the law of North Carolina, United States requires one. If a tenant was never provisioned against a payment — the error is ours — that payment is refunded in full on request.
What a tier changes. A tier changes settings and limits, never the code that produces a score or an evidence record: every tier runs on the same router and the same audit chain.

Marked for a lawyer, not drafted as settled: the refund position above states our intent, not a clause tested against consumer-protection law in every place a customer might be; and nothing here yet addresses what happens to a tenant’s evidence log on non-payment (suspension before deletion, and how long evidence is held before it is purged) or limits our liability for a subscription specifically, as distinct from the general no-warranty section below. Have counsel draft these three before relying on this section commercially.

No warranty, and the limits of liability

This registry is provided as is. Its contents may be incomplete, out of date, or wrong, and a score computed on one date may not describe what a vendor publishes on another. No warranty of accuracy, completeness or fitness for any purpose is given.

Do not use it as the sole basis for a procurement, deployment, compliance or investment decision. It is one input, and it measures documentation. To the fullest extent permitted by the law of North Carolina, United States, the publisher accepts no liability for loss arising from reliance on it. Nothing here excludes liability that cannot lawfully be excluded.

This page is not legal advice and was not drafted by a lawyer. It states what the software actually does and what the publisher intends, in plain language, so that a lawyer has something accurate to work from. Have one review it — and the method — before this registry is published anywhere real.