What you pay for, and who warrants what
You run agents that hold real keys. Three parties can stand behind that work, and each one warrants only its own act.
We charge for one thing: the licence, per agent-year. The warranty on our own act rides inside it and is never sold on its own. $20 a seat, read from Stripe on 2026-09-21.
+Three warrantors, each on its own act3 parties · 1 built · 1 proposed · 1 yours
A promise about someone else's outcome is insurance. A promise about your own product is a warranty. So each party below promises only what it makes.
+ThetaDrivenBUILT
Our own act: the tape is written, the trigger fires on the threshold you declared, the placement re-runs from the commit to the same bytes, and the dispatch arrives inside the stated window.
A service credit when our own act misses, inside the licence.
+The model providerPROPOSED
Its own model's adherence to the spec you declared, read off a record the model did not write.
A credit in its own tokens, issued automatically on our reading. No provider offers this today.
+Your carrierYOURS
The outcome: what your agent did in the world.
Indemnity under your own policy, priced off the tape we keep.
+Why a model provider could credit adherenceproposed
Model providers already credit their customers when a service level slips, for uptime and latency, and enterprise buyers negotiate for the credit to issue automatically on measurement data.
Adherence to a declared spec has no such meter, because the only account of what the model did is the model's own. An account produced by a process cannot contain what the process displaced.
The tape is a record the model did not write. That makes adherence a service level a provider can credit on, in its own tokens, at no risk to anyone else.
+The levelsL0 free · L1 $20 a seat · L2 and L3 unpriced
When the trigger fires, it debits nothing you did not buy. Prepaid hours and credits are yours to spend at any time, fire or no fire. A fire only routes what you already bought to the person with grip where it fired; it never pays out a benefit you did not buy. That keeps them a service, not insurance. A model provider's adherence credit, where one exists, lands in the same balance.
The discount is the spread the market earns: providers are paid from hours you redeem, and because the spec is the tape, an hour is priced against the cell it was spent at.
The licence is the only price on record. The L2 and L3 discounts are UNMEASURED until they are set, and none is shown before then.
+L0 Measurefree
The tape, the rings and the readings. Free and open source (MIT), at any size.
+L1 Attest$20 a seat
The licence seat, per agent-year: the notary-witnessed tape, the loss-history packet a carrier reads, and a service credit when our own act misses.
+L2 HoursUNMEASURED
Expert hours, prepaid at a discount, redeemable whenever you choose. A fire routes them to the provider with grip where it fired.
+L3 ComputeUNMEASURED
Compute credits, prepaid at a discount, redeemable whenever you choose. A fire routes them to the re-run and the fix.
+Free consulting on your deployment, once your tape can be readfloor 8 placed jobs a room · ours: 5 of 9 rooms read (2026-10-17)
Install the free measurement and let your agents work. When one of your rooms carries 8 placed jobs, its shape can be read, and the founder reads it with you, free: where the work landed against what you declared, and which arm to take. After that, the map of maps runs on your rooms the way it runs on ours.
A room reads at 8 placed jobs. Below that, a shape is noise wearing a number, so the page says how many jobs are left and draws nothing.
An offer, never a gate: the install and the measurement are free and open source (MIT) at any count.
+vault26 placed · reads
Enough placed jobs to read a shape.
+architect93 placed · reads
Enough placed jobs to read a shape.
+performer0 placed · 8 to go
Below the floor, so no shape is drawn: 8 more placed jobs and this room reads.
Enough placed jobs to read a shape.
+network0 placed · 8 to go
Below the floor, so no shape is drawn: 8 more placed jobs and this room reads.
+voice49 placed · reads
Enough placed jobs to read a shape.
+builder18 placed · reads
Enough placed jobs to read a shape.
+laboratory2 placed · 6 to go
Below the floor, so no shape is drawn: 6 more placed jobs and this room reads.
+operator0 placed · 8 to go
Below the floor, so no shape is drawn: 8 more placed jobs and this room reads.
+Your ten answers, and the part each one sets10 questions · fill them in on the warranty page
Your readiness check. Ten questions; paste any one into your own AI and let it draft from your repo, or think it through yourself. Every saved answer is yours, and each sits beside the reading that tests it.
+1. The plumberThe spec
Sets the envelope adherence is read against: what the provider would credit on.
+2. The electricianThe routing
Sets who a fire calls, and how far the work may step before your halt acts.
+3. The demolisher and your haltYour halt
Sets what your halt does on your machine; we warrant the signal, never the halt.
+4. The blast radiusThe size
Sets the most the work can move unattended: how large a balance is worth holding.
+5. The causeYours alone
Sets no warrantor reads why the work succeeded; only you hold it.
+6. What only a person decidesYours alone
Sets the decisions no program makes, and who makes them.
+7. The witnessThe record
Sets the out-of-failure-domain record every warrantor reads from.
+8. The first hourThe window
Sets the dispatch window our warranty is measured against.
+9. The wordingYour carrier
Sets what your policy excludes today: the gap the tape lets a carrier write into.
+10. The level, the scope, and the stopYour level and your stop
Sets L0 to L3, the threshold that fires, and the date you would stop paying.
Sufficient for: Who warrants what, what you pay for, and which of your answers sets each part.
Not sufficient for: Whether your agent will behave (undecidable, Rice 1953), a price for L2 or L3, or any provider having agreed to the adherence credit. Five counsel questions on the warranty stay open.
We measure where your agents’ work landed and attest it. We do not control or halt an agent; the halt is yours to wire.
Sufficient for: where it landed, re-runnable from the commit: same bytes, same hash.
NOT sufficient for: whether the work was good, which is undecidable (Rice, 1953); nothing here claims it. Rice (1953): no program can decide, in general, whether another program is correct.