Twenty minutes — the artifact your DORA attestations need
Standards are not care — The T.J. Hooper, 1932 — which is the thread we pulled on: “everyone uses evals” is lagging custom, not the standard. So I built the thing that makes the standard enforceable. It’s a signed, recomputable record of whether an AI service stayed inside its impact tolerance — re-runnable by your auditor or reinsurer, unfakeable by software, the artifact your DORA attestations are meant to rest on. And no access to anything of yours: I run it on a spec and a sample you name, you watch the record get made, and you keep it.
last updated 2026-06-24
▶ Listen on the walk — the whole page, ~4 min
Action items
☐1. Send a timenow
You’re already up for a few questions — just say when, and I’ll call you then. Twenty minutes, no prep on your side.
Everything below is just context for the call — skip it; the only thing I need now is a time.
What we need from you — and what we don’t
You never have to produce a work product, a memo, or a review.
The one thing worth more than any of that: tell me where a hardware-grounded boundary does — and does NOT — map onto the impact-tolerance frameworks you live in (DORA, ISO TC 292). Your pushback is the calibration; finding the hole is the contribution.
What else to research
You already named the first two. Here’s the rest of the trail — take it and run with it; this is yours to use in any room.
✓
The T.J. Hooper (1932) — Judge Learned Hand — you have this
Custom is not the standard of care: "a whole calling may have unduly lagged in the adoption of new and available devices."
✓
Rice’s theorem (1953) — you have this
Why a verifier that is itself software cannot decide the semantic properties of the software it runs inside — the formal reason "evals" can’t close the gap.
→
DORA + ISO TC 292 — enforceable impact tolerances
Your frameworks already demand them; the boundary is how "enforceable" stops being aspirational.
→
Affirmative AI-insurance products (Munich Re aiSelf, Armilla, Mosaic)
They price model-performance drift, not the domain-drift that takes a service out of its lane. That gap is the white space.
→
The Guardian / Builder sources of meaning
Protecting what matters vs imposing will — and why care (disciplined presence) is the hardest, most active work in the room, not the passive one.
Roles, goals & assumptions
The rest is for the call — skip it. None of it needs a reply, and the only thing I need now is the time; it’s here so the twenty minutes isn’t cold.
why we believe this is true
The standard of care just moved: a recomputable, hardware-grounded boundary now exists that makes an AI service’s impact tolerance enforceable rather than asserted — and against roughly $10 trillion a year in cyber damages, that is the difference between a risk a board can price and one an insurer simply withdraws from.
Connection — where you already stand
You certify that mission-critical services survive their worst day, and you’ve watched paper compliance fail the moment a real shock arrives. A person who can read the gap between a checklist and actual structural integrity can do more with a recomputable boundary than anyone selling the checklist.
Contribution — what this lets you give
A board, a regulator, or an underwriter a way to actually price and transfer AI risk instead of excluding it — institutional continuity backed by a boundary that holds, not an eval that reports after the fact.
Growth — the new capability
A deterministic boundary that lives below the software: the moment an agent drifts out of its competence lane it triggers a physical revocation, not a failed test. The missing infrastructure under operational resilience, and almost no one is holding it yet.
Uncertainty — the open variable
Where a hardware-grounded boundary maps cleanly onto enforceable impact tolerances, and where it does not. We think it closes the gap evals can’t; you know the frameworks well enough to find where that’s too strong. That’s the twenty minutes.
Certainty — what does not depend on any of that
The drift verdict is decidable, signed, and reproduces byte for byte on any machine — verifiable today, no trust required.
Significance — who you become
The one who saw the standard of care had already moved — before the industry’s lagging custom caught up.
A few things we take as given
The first move is just a time — never a commitment.
We need your read on resilience, not a work product.
The exposure that matters is the unmodelable, uninsurable kind that makes carriers withdraw — not ordinary cyber loss.
The disciplined presence you steer with is what the boundary is meant to free up, not replace.
Evidence
Custom is not the standard of care
The T.J. Hooper (1932): two tugboats sank without radios; Judge Learned Hand ruled an industry’s lagging custom is no defense once an available device exists. "Everyone uses evals" is that tugboat.
The T.J. Hooper, 60 F.2d 737 (2d Cir. 1932)
Why software can’t audit software
Rice’s theorem (1953): no system can decide the semantic properties of another from inside the same failure domain. An eval IS software — it shares the agent’s failure domain.
Rice 1953
The scale of the exposure
Roughly $10 trillion a year in global cyber damages — the order of magnitude that makes this unmodelable risk an institutional crisis, and the reason carriers withdraw rather than price it.
Cybersecurity Ventures (annual global cybercrime cost)
And none of this is to replace the part of the job that actually matters — the disciplined presence you steer with, the thing that holds a room. It’s to make the floor enforceable so your people can get back to it.
