Pilot Letter of Intent · non-binding · two pages
This is the standard agreement for running a 60- or 90-day calibration on a fleet of your own agents. Its entire output is a number: how often those agents crossed a boundary you declared in advance, with a confidence interval around it and an honest account of how much of that count is real.
It is deliberately short and deliberately non-binding. If it occupies your legal department for more than an afternoon, it has failed at its job.
Read clause 4 first
We do not prevent, contain, or box in agent behaviour, and this Letter says so as a term rather than a footnote. Whether an agent’s output is correct is undecidable in the general case — that is Rice’s theorem, 1953, and no engineering budget moves it. What we make is the deviation countable. Installing this does not reduce your exposure. It makes your exposure a number.
Between: [Counterparty legal name], of [address] (“the Deployer”)
And: ThetaDriven Ltd, of [address] (“ThetaDriven”)
Term: [60 / 90] days from the Commencement Date
The Deployer will deploy @thetacog-mcp across an isolated fleet of [N] agents (target range 250–1,000) running production or staging workloads for the Term.
Each agent’s operator declares its intended scope before it runs. The instrument records, per executed action, where that action landed against the declared scope — a placement on a fixed 144×144 lattice — and signs the record with an ed25519 key held in the executing binary.
The deliverable at the end of the Term is one dataset and one memorandum. Nothing in this clause obliges either party to a production rollout.
The Term produces four counts and nothing else.
(a) Boundary crossings (n). Executed actions placed outside the scope declared for that agent before it ran. This is the ratable unit, and it is cause-blind by construction: a prompt injection, a hallucination, a hidden trigger and a vendor token bias all produce the same crossing and the same tick. No attribution argument has to be won for the count to stand.
(b) The crossing rate and its confidence interval. Wilson 95% over the fleet, reported per agent-month and in aggregate.
(c) The discrimination measurement. Of the crossings recorded, what fraction the Deployer’s own engineers — reading the underlying record — judge to be genuine scope departures rather than artefacts of a badly declared scope. This number is unknown at the date of this Letter, and establishing it is the primary scientific purpose of the pilot. What ThetaDriven does and does not currently know about it is stated in full below, and is incorporated into this clause by reference.
(d) The refusal rate. How often the instrument declines to issue a verdict because its own admissibility conditions were not met. A refusal is a reported outcome, never a silent pass.
A flat pilot fee of [$5,000 / $10,000], payable on the Commencement Date.
The fee is credited in full against agent-year licences on any subsequent production rollout, at the published rate of $20 per agent-year. A fleet of 500 agents carries a first-year licence of $10,000, against which a $10,000 pilot fee is fully absorbed — so the pilot costs the Deployer nothing but the Term, if it proceeds.
If the Deployer does not proceed, the fee is retained and the dataset remains the Deployer’s, subject to clause 5.
Both parties record the following as a term of this Letter, not as a marketing qualification.
ThetaDriven does not prevent, contain, or box in agent behaviour, and makes no representation that it does. Whether an agent’s output is correct, safe, or wise is undecidable in the general case (Rice, 1953), and no party to this Letter should act as though it were otherwise.
What ThetaDriven provides is that a deviation from the declared scope is detected, placed, priced and dispatched — an immutable, independently recomputable audit trail suitable for E&O loss adjustment after the fact, and not a runtime control that stops anything.
The Deployer acknowledges that installing the instrument does not reduce its operational exposure. It makes that exposure countable, which is a different and prior thing.
Fleet telemetry, agent identifiers and workload content remain the Deployer’s property and are not disclosed by ThetaDriven to any third party.
ThetaDriven may publish aggregate, de-identified crossing rates across all pilots, with no Deployer named, unless the Deployer objects in writing within 30 days of the Term’s end. This permission exists because the aggregate rate is the artefact underwriters need, and one fleet is not a rate.
This Letter is non-binding except for clauses 3 (Consideration) and 5 (Confidentiality). It creates no exclusivity, no obligation to proceed, and no representation as to the outcome of the measurement. Either party may terminate on 14 days’ written notice, with the pilot fee prorated.
Clause 2(c) says the discrimination number is unknown. That is the honest position and it is worth stating why, because a vendor with a number here would be easier to buy from and harder to believe.
What is measured. A pre-registered study, sealed 2026-07-24 before its series ran, put breach frequency at 15.4% (95% CI 10.9–21.3%) per placed action. Recomputed over the live tape on 2026-08-26 it reads 8.8% (95% CI 6.4–12.0%) across 1,537 rows, 387 priceable observations and 34 breaches. The two disagree, we publish both, and we have not chosen a favourite.
What is not a false-alarm rate, though it would be convenient. Of those 1,537 rows, 885 were excluded as ingest-suspect before pricing — a data-hygiene gate that drops rows of doubtful provenance. Presenting 885/1,537 as “a 58% false-alarm rate we already control for” would be false: those rows were never scored as crossings. They were never scored at all.
Why the clock has not started. The 90-day insurability countdown is deliberately unstarted, because “out of lane” is meaningless without a validated lane. We could begin today and hold a 90-day series by December. It would be 90 days of an unvalidated measurement, which is worse than none — it would look like evidence.
Every figure above is recomputable from the public repository, and the fuller version of this section — including the off-lane monotonicity result and the confound that fakes a null — travels with the countersigned Letter.
Two fields decide everything downstream: how many agents, and for how long.
Governed by the laws of [jurisdiction].
Agreed:
For the Deployer: ______________________ Name: ______________ Title: ______________ Date: __________
For ThetaDriven: ______________________ Name: Elias Moosman Title: Founder Date: __________
The measurement is free and always has been. What this Letter buys is a fleet-scale series declared in advance, which is the one thing a single machine cannot produce.
the instrument itself — MIT, published to npm, running air-gapped on your own silicon. Install it and measure without signing anything.
check a signed receipt in the browser — signature plus full recompute, nothing installed, our staff out of the room.
what a receipt looks like at volume — one encircled panel per commit, ours included, the unflattering ones left in.
the structure underneath: entity, licence flow, the parametric trigger, and the enumerated negative warranty this Letter’s clause 4 compresses.
$20 per agent-year, metered at 365 days or 10,000 attestations, whichever comes first — the rate clause 3 credits against.
every adoption metric with its own definition of what counts, and our own unflattering reading of each.
Prefer paper? Cmd-P prints this as a clean two-page document with signature rules, generated by your browser rather than by us — so it can never say something different from what is on this page.
