The evidence room
Everything on this page is public and none of it is under NDA, which is not generosity: a compliance standard that is a secret is not a standard. This is the proof of the exposure and the proof of the cure, in the order a decision actually needs them.
If you arrived here holding an advisory invoice instead, this is the same page it points to — and the same rule applies: nothing is owed, and nothing here asks you for anything.
01 · the actuarial reality
An AI loss cannot be priced because it cannot be counted: nobody — not the vendor, not the enterprise, not the underwriter — can say which step in the chain actually failed. That is a falsifiable claim, stated so you can swing at it. Show the count and it withdraws.
Generative-AI exclusion endorsements are attaching to standard policies and carving these losses out of coverage outright. An exclusion is an underwriting confession: we cannot price what we cannot count.
Do not take our word for it — take your carrier’s
Ask your broker for your own generative-AI exclusion form.
It is already on the record, it has your company’s name on it, and it will settle this faster than anything we could send you. If the form does not exist on your policy, you have learned something real and it cost you one email.
What is being priced is the delta between what an agent was specified to do and what it ungroundedly did, permanently on the record. A parametric trigger fires on a failed state-proof — the sealed reading, never an adjuster’s judgment. The payout is remediation in kind: consultant hours and GPU time dispatched to whoever holds the highest-confidence pixel over the failure region. If you want cash instead, that is a broker’s business: a large house can wrap insurance on top of the same reading and carry the liability. Either way the reading is identical. We license the measurement, the broker carries the liability, you pick the payout.
02 · the physical proof · no code required
You do not have to install anything, and you should not have to make a security decision on a work laptop to check a vendor’s claim. Open a signed receipt in the tab you already have: a signature check plus a full recompute, running in front of you.
What comes back is the verdict, the coordinate the action landed on, and a σ that recomputes shape-identical on any stranger’s machine, in perpetuity. Decidable is a higher standard than interpretable: it turns an output into a discrete, recomputable state rather than a qualitative read.
Want to see a wall of them rather than one? Every commit we make mints a receipt — ours included, with the unflattering ones left in.
03 · the board pack · what you forward
ten frames · PDF · attach it
The argument as it is actually delivered: one word, then two pictures, then the control, then the arithmetic, then the exposure and its date — one added inference per frame, never two. This is the file for someone senior with exactly four minutes, and it is a file rather than a link because a board pack takes attachments.
Download the brief (PDF) →thetadriven.com/decks/grip-deck.pdf
who is on the hook · entity by entity
The next question a general counsel asks is not whether the measurement works, it is who holds what and where liability sits when an agent breaks. The walkthrough builds that picture one entity at a time — the open instrument, the carrier, the insured deployer, the licence line between them — so no claim about who is on the hook gets ahead of its own proof. Open the liability walkthrough →
what your risk function will ask for
The lit-mass floor, the portfolio distributional audit and the false-negative protocol: what counts as inside a spec, how a breach rate is measured against a distribution rather than asserted, and the documented procedure for when the trigger fails to fire on a real breach. They sit on the public board with their status and our own unflattering reading of it. The three hardening documents →
independent validation · not our employee, not our investor
Dr. Benito R. Fernandez
PhD MIT • 32yr UT Austin Professor • 6 patents
"FIM and hardware-based trusted execution environments are a match made in secure computing heaven—a software blueprint for transparency paired with hardware's ironclad enforcement."
neural networks · control systems · the whisper company · snowbird declaration signatory
04 · the developer handoff · for your engineering team
The party running the system is never the party that certifies it — which means the logging is your engineers’ job and the certification is yours. The format they need is free, MIT-licensed and theirs to keep: github.com/wiber/thetacog-mcp.
One line for them. It runs locally, nothing leaves the building, and it returns the same receipt you just watched recompute:
npx -y thetacog-mcp@latest attest-demo
The builder’s version of this page, if you want to send them a link instead of a command: thetadriven.com/cta.
05 · the resolution
You are not a fool and this is not aimed at you. You are a professional in good standing operating in a market that keeps selling confidence nobody can back with a count. The target is that condition, never the person reading this.
Check it. Run the browser verification above. If we are wrong you will see it faster than anyone we could hire.
Push back. Show the log a carrier accepted to underwrite an AI loss, or the count that prices one, and the claim withdraws.
Adopt it. Buy the agent-year licences that match the agents you are actually willing to stand behind.
metered at 365 days or 10,000 attestations · thetadriven.com/pricing
We never promise the agent stays in its lane — that question is undecidable and anyone selling you the promise cannot deliver it. What is guaranteed is that the deviation is detected, placed, priced and dispatched.
06 · the index · everything, in one place
the free, MIT repo — the standardized format to log what your AI is doing, runnable today, alone, for nothing
the ten-frame brief as a file — the one to attach to a board pack
the builder’s door — run it free in forty seconds, then hand it up or license it yourself
the director’s door — Article 14, Caremark, and what a receipt is and is not for
what a receipt actually looks like — one encircled panel per commit, ours included, unflattering ones left in
check a signed receipt in the browser — signature plus full recompute, nothing installed
the advisory quote — prices a liability already held, opens as a quote and never as a bill
the pilot Letter of Intent — non-binding, two pages, 250–1,000 agents for 60–90 days; clause 4 is the one that says what we will not claim
the mechanism underneath, for whoever asks for it — nothing above depends on reading it
who is liable — the entity structure and licence flow, staged one thing at a time
the adoption board — every metric with its own "what counts", our own unflattering reading, and the three hardening documents that gate everything downstream
the competence-coordinate thesis — software cannot verify software (Rice); a decidable, hardware-signed gate now exists
the Smith / Rice spine — where a governed duty maps to a measurable boundary
the disclosed, capped escalation ladder, and the institutional case (executive shield, kill switch, dual-track)
the advisory invoice itself, and the financial mechanics behind it
$20 per agent-year, metered at 365 days or 10,000 attestations, whichever comes first — the only payment path
the instrument
Tesseract Physics — Fire Together, Ground Together — the argument in full, plus the receipts it points to
why the word for certainty misunderstands a system whose context keeps changing
why the diagnosis names the condition, never the person
There is no obligation attached to anything on this page. If it accompanied an advisory invoice, that invoice is advice and not a legal instrument — nothing is owed. The only real transaction anywhere is the per-agent licence at /pricing.
