What does 'if it's debatable, it's not insurable' mean?

Insurance prices a countable event. If whether the event happened is a matter of expert opinion rather than record, an underwriter has nothing to attach a number to, and the risk stays unwritten. The sentence is about the boundary being arguable, not about the risk being large.

Are my AI agent logs enough to prove what the agent did?

A log records that a write happened. It does not record whether the write was inside the boundary the deployer declared beforehand, because the log is emitted by the same process whose conduct is in question. That is why logs get argued about rather than relied on.

Can software prove its own AI agents stayed in their lane?

Not in general. Rice’s theorem (1953) establishes that every non-trivial semantic property of a program is undecidable, so no checker settles whether a run was good. What is decidable is WHERE a run landed against a declaration sealed before it ran, which is a placement, not a verdict on quality.

What does the $20 entry point include?

One agent-year licence under US Patent App. 19/637,714 — metered at 365 days or 10,000 attestations, whichever comes first — plus the manual, Tesseract Physics: Fire Together, Ground Together. Installing and running the gate itself is free and always has been.

iamfim.com — the sentence on the sticker

If it’s debatable, it’s not insurable.

An underwriter cannot price what two experts can argue about.

You don’t have to trust the sentence. Run the thing it describes.

Forty seconds, on your own machine, nothing uploaded and nothing to sign up for. It reads a commit, walks it against a declaration sealed before the work ran, and prints a coordinate and a sigma. Run it twice on the same commit; the coordinate comes back the same, because the verdict is a pure function of the commit and no model sits in its path.

npx -y thetacog-mcp@latest attest-demo

The repository is open and stays open: github.com/wiber/thetacog-mcp. Open source is why you can check us. It is not the product — the building code is free to read, and the inspection stamp is not.

Why the sentence holds

Insurance needs a countable event: fortuitous, bounded, and pooled across enough similar units to be credible. A countable event needs a boundary that is not a matter of opinion. Where the boundary is arguable, there is no number to attach a premium to — so if it is debatable, it is not insurable, and the risk simply goes unwritten. That is not a rule anyone here invented; it is the condition every line of business already runs on.

Now hold your agent stack against it. Your agents are writing to production right now. Your logs record that it happened. Nothing you hold records whether it was allowed — because the log is an emission from inside the same process whose conduct is in question, and an account produced by a process cannot contain what that process displaced. Put two competent experts in a room with your logs after an incident and they will argue, in good faith, for as long as you pay them. That argument is the uninsurable part. Not the agent. The argument.

The move is the one insurance has always made, and it is deliberately unambitious: nobody claims a policy stops a fire. Underwriting made fire a number. We make no claim about whether your agents behave — that question is undecidable and we say so on every receipt. We claim the deviation from what you declared is detected, placed, priced and dispatched, which is checkable, and which the sentence above says is the whole prerequisite.

Two doors, and they are not the same person

“What can I do?”

You build the agents. Run the gate free, see what a receipt actually contains, then decide whose budget this is — yours, or the one upstairs.

The builder’s door →

“Am I personally liable?”

You signed for the fleet. The standard is not ours and predates every vendor in this market — and you do not need the agent stopped, you need to be able to prove you could have checked.

The director’s door →

The entry point

$20 an agent-year, and the manual comes with it.

Installing and running the gate is free and stays free — that is the lighthouse. Paying is what makes a run count. One licence per agent you run under US Patent App. 19/637,714, billed annually, metered like a lease: 365 days or 10,000 attestations, whichever comes first. This repository burns about 56 attestations a day, so a hard-working agent runs through roughly two agent-years in a calendar year. We publish that number because ours is the only fleet we can honestly measure.

The licence — one agent-year, $20, at the same flat rate a carrier pays for the same instrument. Both sides of the trade buy the identical thing, which is what makes it a measurement rather than a vendor’s opinion.
The manual — Tesseract Physics: Fire Together, Ground Together, the full book, so the officers who have to defend the record can read why it is built the way it is before they need it. Download the manual (EPUB) · read it on the web
The command — npx -y thetacog-mcp@latest attest-demo — the same one on the sticker, and the reason nobody has to be sold anything to check the claim.

License an agent — $20/yr →The full terms →

At fleet scale the entry is the Governance Harness rather than a handful of units — a five-agent trial proves nothing about a five-thousand-agent fleet, and we would rather say that now than six months in. The $20 unit is how a builder starts and how a fleet scales out; it is not a pilot.

What’s on the record

Raw ingredients, not a conclusion. They are all checkable without us.

The origin of the problem is the origin of the industry. Modern liability insurance was assembled at Lloyd’s Coffee House in the 1680s for exactly this gap: capital would not fund voyages it could not verify, and a captain’s word about his own conduct was the thing under dispute. The ship’s log — kept to a standard, not to the captain’s convenience — is what closed it.
Rice, 1953. Every non-trivial semantic property of a program is undecidable. No checker, and no larger model, settles whether a run was good. That door is closed by a theorem, not by an engineering budget.
The oversight standard is old and it is not ours. In re Caremark (Del. Ch. 1996) put directors on the hook for failing to implement any reporting system at all; Stone v. Ritter (2006) set the bar at bad faith rather than negligence; Marchand v. Barnhill (2019) showed the courts willing to enforce it. A receipt is one input to a board process and never the process itself.
The lag is the part worth arguing about. Nineteen years separated the Model T from compulsory motor cover in Massachusetts. Nobody here claims to know this market’s number.
Longer form: Moloch does not fund the uninsurable · You never insure the catastrophe · the definitions

What this page claims, and what it does not. WHERE a run landed against a declaration sealed before it ran is provable and re-runnable — recompute any verdict yourself with the command above. WHETHER the work was bug-free is undecidable (Rice), and we do not claim it. We are not an insurer, we bear no risk, we underwrite nothing, and a licence does not make anyone insurable — whether a carrier writes against this record is that carrier’s decision and none of our business. What is sold is the record.