Terms of service

Draft. This document has not been adopted and is not in force. It is published here so it can be read and reviewed before it is.

Also: Neutrality · Privacy

Status: DRAFT. Not adopted, not in force. These terms bind nobody until Sparks & Sawdust LLC adopts them. They were drafted by an agent, not by a lawyer, and §11 lists the clauses that need one. docs/legal/README.md says which parts are still open.

Version: draft-1, drafted 2026-09-17. MODEL-70.

1. Who you are contracting with

The ModelSpec API and the modelspec.dev and benchgraph.dev sites are operated by Sparks & Sawdust LLC ("we", "us"). "You" is whoever calls the service, including an autonomous agent calling it on somebody's behalf. An agent that accepts these terms binds the person or organisation it acts for.

2. What the service is

ModelSpec catalogues AI models and ranks them against a published method. Two things are live today, and both are free and open to anyone:

• The static export — versioned JSON under https://modelspec.dev/api/, including the catalogue, the rankings and the ranking policy. No account, no key, no charge.
• The rank endpoint — POST https://api.modelspec.dev/v1/rank and GET https://api.modelspec.dev/v1/health. It computes a ranking per request from that same published export. As of the date above it requires no key and costs nothing.

Nothing else is live. In particular there is no paid plan, no API key, no metered tier, no billing and no payment rail in operation on the date above. §6 states the rules that will apply to a charge when one exists; until then, no charge exists. We will not bill you under terms you did not see first.

What the service is not. We do not run inference. We do not proxy, relay or resell access to any model, and your prompts and inference traffic do not pass through us (§4, §7). We are not your agent, your broker or your advisor, and a ranking is information, not a recommendation to enter any transaction.

3. The honest-broker rule

This is the rule the rest of these terms exist to keep, and it is the reason we will never sell the thing a ranking is worth most to the people it ranks:

Charging the consumer of a recommendation is compatible with being an honest broker. Charging the subjects of one is not.

We charge the consumer of a recommendation. We do not charge its subjects.

4. The neutrality commitment

No referral fees, no paid placement, no provider-paid visibility, permanently.

No model provider, hosting platform, gateway, inference vendor or other subject of a ranking can pay us for a better position, for inclusion, for visibility, or for anything else that touches what an answer says. There is no rate card for this, because there is no price at which it is for sale.

Neutrality does not stop at money. No provider, gateway, hosting route or inference vendor is privileged at any stage of a recommendation: not in the ranking, not in the tie-breaks, not in a default hosting suggestion, and not by any commercial relationship of ours.

We also do not earn on your inference. We recommend and hand off: your calls to a provider, gateway or local runtime go there directly and never through us. A margin that grows with your token volume would be a reason to steer you, so we have arranged not to have one.

This commitment is machine-readable. It is published as the neutrality block of the ranking policy, in the same object as the ranking floors, at https://modelspec.dev/api/rank/profiles.json, and in the policy block of every rank response. It is fetchable with no key and no account, so you can check it rather than trust it. The ranking method itself (api/ranking/engine.py) is public for the same reason.

If we ever break this commitment, the JSON says so before the prose does.

5. Using the service

You may call the API, script against it, and build products on it. You may not:

• present our output as something other than what it is, or attribute to us a claim we did not make;
• misrepresent the neutrality commitment, or imply a paid or endorsed relationship with us that does not exist;
• attempt to obtain data you are not entitled to, interfere with the service, or evade a limit or a refusal;
• use the service in a way that breaks the law where you are or where we are.

We may refuse or withdraw service for any of the above.

6. Billing

None of this is in operation on the date above. It is stated now, before any money moves, so that it cannot be written to suit the first dispute.

6.1 Only a successful result is charged. A charge is incurred when, and only when, the service has delivered a result to you. A request that does not produce a delivered result is not charged, whatever caused it.

6.2 What counts as a delivered result. An HTTP 200 response carrying a ranking. A ranking that reports ranking_status: "partial" is a delivered result: it means some models lack the evidence to be ordered, which is the answer, honestly labelled, and not a degraded one. We would rather tell you what we do not know than charge you for a guess.

6.3 What is never charged. None of the following is a delivered result, and none of them is charged:

• a refusal for a malformed or invalid request (HTTP 400, 413);
• a request to an endpoint or method that does not exist (HTTP 404, 405);
• a well-formed request that no model survives (HTTP 422). You are told which constraint emptied the pool; you are not billed for the finding;
• any failure on our side, including an unreachable or unreadable catalogue (HTTP 5xx), a timeout, or a response you never received;
• a rate-limit refusal. Being told to slow down costs nothing (HTTP 429), and being refused for a missing, invalid or revoked credential costs nothing (HTTP 401, 403). A refusal is not a service.

6.4 Settlement before delivery. Where the service is prepaid, settlement is verified before the answer is delivered, and the balance is drawn down per delivered result. An unsuccessful call leaves your balance where it was.

6.5 Partial failure. If a charge is taken and no result reaches you, that charge is refunded in full. You do not need to show that the failure was ours. If you were charged and are not sure you were served, tell us and we will look; where the record is ambiguous, it resolves in your favour.

6.6 Refunds. Charges for delivered results are not refundable merely because you disliked the answer — the ranking is the product, and a ranking that only charged when it flattered would not be worth having. We refund: anything charged without a delivered result (§6.5); anything charged in error, including duplicate or mis-metered charges; and unused prepaid balance, on request, minus nothing. If we withdraw the service or these terms change to your material detriment, unused balance is refunded.

6.7 Prices. No price is in force. Any price will be published before it applies, and a price change will not apply retroactively to balance you already bought.

6.8 Taxes. Prices will state whether tax is included. You are responsible for taxes on your own side of the transaction.

7. Your data

The privacy statement is at https://modelspec.dev/legal/privacy/ and describes only what the service does today. Two points belong here because they are terms, not just practice:

• We do not receive your prompts. A rank request carries a profile — use case, environment, constraints — computed on your side. Prompt text is not a field of this API. Widening it toward prompt text would be a breach of this commitment, not a feature release.
• We do not proxy your inference tokens, so we do not see, store or meter the content of your model calls. There is nothing for us to hold.

8. Data, licences and what you may do with the output

The catalogue data is licensed CC BY-SA 4.0 (LICENSE-DATA). The code, including the ranking method, is licensed MIT (LICENSE). Those licences govern the data and the code; these terms govern the hosted service, and nothing here takes away a right either licence grants you.

Rankings are computed from that data, so redistributing a substantial part of it carries the CC BY-SA attribution and share-alike conditions with it. Acting on a ranking, quoting one, or building a product whose output is informed by one does not put your product under that licence.

Card data is compiled from third-party sources, each cited with the date it was read. We do not own those sources and grant you no rights in them.

9. Accuracy, and the absence of any warranty

Rankings are computed from published benchmark results and vendor-stated facts, which are incomplete, sometimes wrong, and always out of date by some margin. We publish the floors, the provenance of each row and the date each source was read so you can judge for yourself. A model's position is evidence about evidence. It is not advice, and it is not a promise about how the model will behave for you.

The disclaimer below is the one from the MIT License, which is the licence this project's code already carries (LICENSE), reproduced verbatim rather than rewritten:

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

There is no uptime commitment and no support commitment. Neither is offered, so neither is promised.

10. Changes, suspension and termination

We may change the service. A change to the ranking method is published in the method itself, which is public. A change to these terms is published here with a new version, and a change that materially reduces what you get takes effect for you only after it is published, never before.

You may stop using the service at any time. We may suspend or end access for a breach of §5. §4 does not expire, is not suspended, and does not change with a new version of these terms.

11. Open — needs a lawyer before adoption

Left blank on purpose rather than guessed at. Sparks & Sawdust LLC should have counsel settle each of these before any money moves:

• governing law, jurisdiction and venue;
• limitation of liability, and any cap;
• indemnity;
• dispute resolution, and whether arbitration or a class-action waiver is wanted;
• consumer-protection, distance-selling and cooling-off rules in the places paying callers will actually be, and whether the refund position in §6 clears the strictest of them;
• sales tax and VAT registration and collection;
• the data-protection basis and any controller/processor terms (§7 and the privacy statement describe what the system does; whether that is all that is required is a legal question);
• whether §4 should be a contractual undertaking that survives a change of control, which is the only version of "permanently" a buyer cannot quietly drop.

12. Contact

Sparks & Sawdust LLC. Contact details to be filled in before adoption.