Published on: July 2, 2026
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Send Strategic Nudge (30 seconds)Green in-lane · amber a little out · red drift. Every panel is a real commit, byte-identical on recompute. Tap any panel to open its shareable receipt.
Everyone else walking into a room full of risk officers is pitching "smarter AI" or an AGI timeline. We walk in speaking the room's native tongue: frequency, severity, volatility, spread. Not because it's a clever costume — because an autonomous agent really is a volatile position on your book, and the room already has instruments for pricing exactly that shape of problem. So instead of a demo, we bring a question: "What's your AI's breach rate right now — the one you'd actually put in a filing?"
Most vendors don't have an answer, because they don't have a ledger. We do. So we asked our own ledger the question three different ways, live, while writing this post — and got three different, defensible numbers. That's not us being sloppy. That's the actuarial joke, and if you sit with it for a second, it's the entire pitch.
Here's the experiment, run against data/pmu/measure-history.ndjson — the append-only ledger every attestation writes to — with nothing staged, nothing cherry-picked:
Instrument 1 — the sealed premium engine (calibration-premium.mjs) treats a breach as a strike crossing: driftPct above 4.668% is the loss event, exactly like a put option going in the money. Out of 172 priceable attestations, 44 crossed the strike:
$ node scripts/pmu/calibration-premium.mjs
breaches 44 (empirical breach rate 25.6%)
p̂ 95% Wilson CI [19.6%, 32.6%] half-width 6.5%
σ semantic vol 2.04 (stddev of driftPct = Black-Scholes σ)
PREMIUM 397.0 units
Instrument 2 — the unsealed lens proxy (the same signal our own prompt-time hook computes on every message you've been reading in this session) takes a rougher, wider window — every commit, not just clean attestations, flagged the instant offPct crosses 15%:
📈 series: vega 7.94 · breach 57.5% (all-commits offPct>15 — not the sealed premium rate) · n120
Instrument 3 — the variance rail (variance-option.mjs) doesn't ask "did it breach" at all — it prices the turbulence of the lane itself, the way a variance swap prices realized volatility instead of direction:
$ node scripts/pmu/variance-option.mjs
fair variance K_var 3.675 (fair vol 1.917)
spot realized var 2.462 (latest window)
read AT/BELOW fair — lane is calm (short pays)
Three questions, three numbers: 25.6% with a real confidence interval, tight enough to write a policy. 57.5% on a wider, noisier, all-commits window that was never meant to be sealed. 2.46 on a variance scale that isn't a percentage at all. None of them is wrong. They're not measuring the same thing, any more than a 10-year default probability and a one-week credit spread are measuring the same thing on the same bond.
The divergence is the actuarial skill. "What's the breach rate" is not one question — it's frequency-selection, and frequency-selection depends on the window, the seal, and what counts as an event. An instrument that always gives you one tidy number regardless of which question you asked isn't more honest. It's hiding the modeling choice from you.
Take the number we'd actually stand behind in a filing — the sealed 25.6%, roughly one attestation in four leaving its authorized lane — and put it next to a human doing comparable work. A clerical worker on routine claims processing runs a baseline error rate around 1–3%; at 25.6% that person is fired by lunch. A senior underwriter's fatal-error rate — the kind that exposes the firm to unhedged risk — sits under 0.5%. The closer human analog is a junior engineer mid-rabbit-hole: scope creep, "let me just refactor this while I'm in here," somewhere in the 15–20% range on unsupervised work. An autonomous agent, today, without a lane-keeping instrument watching it, runs hotter than that junior engineer and doesn't get tired, doesn't blush, and doesn't ask for a second pair of eyes before it ships.
That's the punchline for the actuary: the asset has the throughput of a machine and the operational discipline of someone who's never once had a manager look over their shoulder. You don't fix that by hiring a better model. You fix it by putting a shoulder there.
A human handed a vague instruction fills the gap with judgment and stays roughly on task — call their volatility to ambiguity close to zero. Their volatility only spikes for biological reasons: fatigue, a bad morning, a Friday afternoon. An agent's volatility runs the other way. Zero biological drift, but a single ambiguous word in a prompt can rewrite the whole trajectory — not a small nudge, a different destination.
We don't collapse that into one Vega number, because we measured it three ways and the ways disagree on purpose. The sealed engine's σ (2.04) is the standard deviation of how far driftPct wanders once it wanders — how bad a breach is, given that one happens. The rail's vol-of-vol (1.063) is the volatility of the volatility — how much the lane's own turbulence itself is shifting week to week. The lens proxy's rougher 7.94 is a same-day early-warning number, deliberately noisy so it can fire before the sealed engine has enough clean rows to be confident. Three Greeks, three jobs. A single "AI volatility score" that flattens all three into one badge is the thing that should make a risk committee nervous, not reassured.
Here is the part we didn't plan for and are printing verbatim, because "don't take our word for it" has to survive pointing at us. Running the sealed premium engine to source the numbers above returned a warning we don't get to edit out:
$ node scripts/pmu/calibration-premium.mjs
seal ⚠ ROOT MISMATCH — a priced row was altered, inserted, or removed
— run: node scripts/pmu/ledger-attest.mjs
Somewhere between the last clean attestation and this afternoon, a row in our own ledger got touched outside the sealed path — the ed25519 root no longer matches. The 25.6% and the 397.0-unit premium above are still what the current rows say, but the chain-of-custody claim — "no one edited history after the fact" — is, right now, false for us. We are not quietly re-sealing it before you read this. We're telling you it broke, in the same post where we're asking you to trust the number it produced, because a seal that only gets checked when it's convenient isn't a seal. ledger-attest.mjs is the re-run; until it comes back clean, treat the premium above as directionally right and not yet the number that goes in a filing.
This is the whole discipline in one accident: the instrument that would catch us lying to you is the same instrument, and it just caught something. We'd rather you see it break in public than trust a number we can't currently re-derive ourselves.
Here's the line you get to walk away with, the one that has to survive being said out loud at a bar: "Our own agents leave their authorized lane on roughly one attestation in four — a real confidence interval, not a vibe — and we'll quote you a volatility on top of that frequency, the way you'd quote volatility on any position. We can't stop the agent from drifting; that's Rice's theorem, not a roadmap item. What we built tells you the instant it happens, prices how often it happens, and just told on itself when its own seal broke." That's not a pitch about smarter AI. It's a pitch about a position you can finally see the Greeks on.
Do this: run node scripts/pmu/calibration-premium.mjs or node scripts/pmu/variance-option.mjs yourself against a public attestation ledger and get the same numbers we did — including, if you're unlucky the way we just were, a seal warning that isn't ours to hide from you either. If you carry risk for a living, the standing conversation is at thetadriven.com/dinner. For the mechanism underneath all three instruments, start at Are you out of your pixel?, or read One Ruler, Two Markets on why the same measurement prices as both a premium and a variance swap.