The Hooper Standard — Making AI Risk Insurable

Deterministic, recomputable on-chip verification · the five source infographics, transcribed · every claim verified against primary authority · June 2026 · thetadriven.com/dinner

The question is not whether your AI agents are good — no one can prove that (it is Rice's theorem, not a tooling gap). The question is whether, when an agent acts outside the lane you authorized, you can prove you would have caught it. Three currents — a 1932 negligence rule, a Delaware fiduciary duty now reaching individual officers, and AI exclusions being written into liability policies this renewal cycle — converge to make that proof a personal obligation. The five infographics below frame the argument; under each is a faithful transcription, and at the end every load-bearing claim is checked against its source.

1 — Making AI Risk Insurable via the Hooper Standard

Making AI Risk Insurable via the Hooper Standard — infographic — Source: NotebookLM-generated infographic · "The Liability Has Your Name On It: Making AI Risk Insurable via the Hooper Standard."

The Crisis (left)

The Personal Liability Trap.
Rice's Theorem — the software ceiling: mathematically impossible for software to reliably judge the semantic properties of other software.
The Insurance Gap (ISO CG 40 47/48): insurers are filing absolute AI exclusions; they cannot independently measure agentic behavior. Officers face personal liability under the Caremark & McDonald's doctrines for oversight failures.
attest-demo (center): Node A (Buyer) → attestation request → Silicon Oracle (ThetaDriven PMU) → verified response → Node B (Vendor). "Secure secures · Verified by Silicon Serial."

The Sovereign Box & the Hooper Standard (right)

Semantic-Physical-Hardware (SPH) unity: intent is mapped to physical execution paths on the chip, making drift a measurement, not an opinion.
Deterministic lattice (144×144) · Drift Receipt: a priceable financial instrument.
npx thetacog-mcp attest-demo · "Trust Physics."
The Hooper Standard — availability over custom: Stability (stochastic flips / deterministic byte-identical every run); Recomputability (impossible / Exit 0); Legal standing ("good faith" reach / forensic fact).

2 — From Stochastic AI Vibe to Deterministic Fact

AI Liability and Verification Standards — infographic — Source: NotebookLM-generated infographic · "The Liability Has Your Name On It: From Stochastic AI Vibe to Deterministic Fact."

Workflow Management — yesterday / stochastic / uninsurable (left)

LLM-on-LLM oversight is mathematically impossible: Rice's theorem proves software cannot reliably decide non-trivial semantic properties of other software.
"Silent AI" exclusions are the new standard: insurers like Allianz and W.R. Berkley are actively filing absolute exclusions for unsecured AI.
Personal liability ends at the officer: the McDonald's doctrine binds individual executives to the personal duty of oversight.
Center: "Unreliable Output."

Liability Verification — today / deterministic / the Hooper Standard (right)

Semantic-on-silicon — the sovereign box: mapping semantic intent to physical execution paths on the chip at ~6M walks/sec.
The recomputable bearer receipt: an ed25519-based artifact any third party can verify offline without trusting the issuer.
The T.J. Hooper inversion: once a deterministic monitor exists, ignoring a "black box" dashboard becomes willful negligence.
Throughput ~50 tokens/sec (stochastic) vs ~6M walks/sec (ballistic) · Determinism flips on boundary cases vs 100% byte-identical · Auditability ephemeral opinion vs hardware-signed bearer instrument.

3 — The Zero-Trust AI Transaction

The Zero-Trust AI Transaction — infographic — Source: NotebookLM-generated infographic · "From Stochastic Vibes to Silicon Facts: The Zero-Trust AI Transaction."

The stochastic trap — why current evals fail (left)

The "T.J. Hooper" negligence trap.
Software cannot decide software — Rice's theorem: a general judge cannot reliably verify without infinite recursion.
The $237.5M oversight penalty: failing to install independent oversight is "bad faith" & personally uninsurable. (See verification: this figure is the Boeing settlement, not McDonald's.)
Incumbent (LLM judge): fails / flips answers on identical runs; ~50 tokens/sec; ephemeral, no recomputable receipt. "Personal Liability Risk · Off-Domain · Drift."

The zero-trust transaction — trust physics, not ethics (right)

Intent (Spec): Node A (Buyer) signs the spec. Reality (Payload): Node B (Vendor) signs the work. Δ (Receipt/Verdict): the chip signs the verdict.
S=P=H (Semantic = Physical = Hardware): semantic drift = physical drift; the chip measures execution drift.
"Position IS meaning": you don't trust the model — you trust the physics of the chip and the math of the walk.
ThetaDriven (Silicon Gate): 100% byte-identical every run; ~6M walks/sec; ed25519-signed, hardware-attested. "In-Role · Ballistic Walks · Signed Receipt ed25519 Seal · Underwritable AI Risk."

4 — The Hooper Standard: from Vibe Checks to Silicon-Grounded Facts

Deterministic AI Governance Standard — infographic — Source: NotebookLM-generated infographic · "The Hooper Standard: From AI Vibe Checks to Silicon-Grounded Facts."

Intent (problem) — stochastic governance (left)

The Rice's-theorem trap: software cannot legibly verify software without infinite recursion; LLM-based evals are "stochastic opinions," not facts.
The "silent" AI exclusion: insurers will not affirmatively cover what they cannot independently measure; most current policies quietly exclude AI risk.
Good faith is not a slide deck: under the Caremark doctrine, directors are personally liable if they lack a functional system of oversight.

Reality (solution) — deterministic hardware (center)

SPH unity: mapping intent to a 144-tile lattice on silicon makes meaning a measurable physical coordinate.
6 million verification walks per second: bare-metal speed, with a stated 3.5σ separation between compliant work and drift.
The irrefutable ed25519 receipt: a hardware-signed bearer instrument any third party can recompute offline to verify compliance without trust.

Delta (outcome) — liability shift (right)

Incumbent (chatbot eval) vs the Hooper Standard (on-chip): Nature stochastic opinion / deterministic fact · Verification fails (Rice's theorem) / recomputable (Exit 0) · Liability personal exposure / insurable asset.
npx thetacog-mcp attest-demo — "the radio" (the T.J. Hooper standard).

5 — The Node A-B Test

The Node A-B Test Framework — infographic — Source: NotebookLM-generated infographic · "The Node A-B Test: Moving from Stochastic 'Vibes' to Deterministic AI Liability."

The problem — the stochastic trap (left)

Rice's theorem, the "turtles" problem: software cannot judge a non-trivial property of software without infinite, un-auditable recursion.
The 40/50/80% "flip band": LLM judges flip their verdicts on identical inputs near boundary edges, making them legally indefensible.
Uninsurable "custom": a PASS is a transcript — a non-computable opinion, not a priced financial fact.

The solution — the Node A-B zero-trust transaction (center)

Step 1 — Node A signs the "Reef": the Buyer/Spec owner signs a human-legible intent compiled into a 144-tile semantic lattice.
Step 2 — Node B signs the "Payload": the Vendor/Agent signs their deliverable, binding it to Node A's spec/hardware boundary.
Step 3 — the Silicon Oracle verifies: an on-chip "ballistic walk" (PMU) generates a deterministic, hardware-signed verdict (MATCH / no-MATCH). → Δ deterministic receipt.

The impact — the Hooper Standard & underwriting (right)

Custom vs available care: per the 1932 T.J. Hooper ruling, failing to use available verification technology constitutes negligence.
The "strike price" of AI risk: the deterministic verdict outputs a Sigma (σ) score, turning semantic drift into a priced financial instrument.
Forensically recomputable: any node can re-run the math on its own laptop to verify the signed receipt.
The 'Judge' (LLM) vs the 'Oracle' (Hardware): Determinism (fails, samples probabilities / passes, byte-identical) · Auditability (non-recomputable opinion / ed25519-signed bearer fact) · Throughput (~50 tokens/sec / ~6M walks/sec).

The claims, verified

Each load-bearing claim from the five infographics, checked against primary authority. Where a figure was wrong in the source art, it is corrected here; where a claim is our own result rather than settled fact, it is flagged as such.

Claim (as shown)	Verdict	Basis
Software cannot reliably decide non-trivial semantic properties of other software (Rice's theorem)	Verified — math	Rice's theorem (1953): every non-trivial semantic property of a program's behavior is undecidable. A corollary of the halting problem. This is the hard ceiling under "LLM-judges-LLM," not a tooling gap.
LLM judges flip verdicts on identical inputs near boundaries ("the 40/50/80% flip band")	Verified — independent	Controlled position-bias studies: swapping the order of two identical responses literally flips the verdict; the effect is systematic directional bias, not random noise. Verbosity/format biases compound it. (IJCNLP 2025 position-bias study; EvidentlyAI; Adaline.)
The T.J. Hooper inversion — once a monitor exists, not using it is negligence; custom is no defense	Verified — case	The T.J. Hooper, 60 F.2d 737 (2d Cir. 1932), Learned Hand. Tugs lost barges for lack of a (non-customary) radio; held negligent anyway. Hand's actual test: a precaution that is available, reasonable, and proportionate — not "any new gadget creates a duty." The "inversion" applied to AI monitors is our framing of that established test.
Directors are personally liable for an oversight failure (Caremark)	Verified — case	In re Caremark, 698 A.2d 959 (Del. Ch. 1996): duty to make a good-faith effort to install/monitor systems that surface mission-critical risk. Marchand v. Barnhill, 212 A.3d 805 (Del. 2019, Blue Bell) revived it for a "mission-critical" risk.
"$237.5M oversight penalty"	Verified — but re-attributed	The figure is real but is the Boeing 737 MAX derivative settlement (In re Boeing, Del. Ch., approved 2021) — the largest duty-of-oversight settlement in Delaware history. It is not the McDonald's figure; the source art conflated them.
Personal liability "ends at the officer" (McDonald's doctrine)	Verified — case	In re McDonald's, 289 A.3d 343 (Del. Ch. 2023): extended the Caremark oversight duty, for the first time, to corporate officers — the people who actually deploy the AI.
The exposure is personally uninsurable	Sound — legal analysis	Oversight liability sounds in bad faith (a loyalty breach): the category §102(b)(7) charter exculpation does not cover, §145 indemnification needs good faith to reach, and D&O conduct exclusions / public policy carve out. One of the few exposures the usual backstops are designed not to fully absorb.
"Silent AI" exclusions (ISO CG 40 47/48; Allianz, W.R. Berkley) are arriving now	Verified — independent	Verisk/ISO forms CG 40 47 (broad, Coverage A+B) and CG 40 48 (Coverage B) attach at U.S. CGL renewals from 1 Jan 2026, trigger on loss "arising out of" generative AI. By Apr 2026 W.R. Berkley, Chubb, Travelers, Berkshire Hathaway, Cincinnati Financial had filed; >80% of filings approved. (IndependentAgent.com; actuary.info; the CG 40 48 form text.)
~6M walks/sec, 144-tile lattice, ed25519 recomputable bearer receipt, byte-identical re-verification	Our result — receipt proven, drift partial	The cryptographic receipt layer is implemented and verifies independently offline today (ed25519; recompute = Exit 0). The throughput and the physical-drift / 3.5σ signal beneath it are stated, by our own ledger, as a partial result on a single architecture — not yet the bankable, cross-hardware claim.
The σ verdict is a "strike price" — drift becomes a priced financial instrument	Our direction — honest inches open	PMU calibration produces a premium from a breach rate (~13% dogfood breach = the loss-ratio asset). Two inches remain open and are stated as such: surface-robustness and true out-of-sample calibration. Adjacent to a Black-Scholes-style primitive; not a closed pricing model yet.

What this does not claim. We do not claim to prove an agent's work is correct — that is undecidable, and anyone selling it is selling a solved halting problem. We claim a narrower, decidable thing: whether an action stayed inside the lane its operator authorized, signed and recomputable. The receipt layer is proven and verifies independently today; the physical-drift signal beneath it is, by our own ledger, a partial result on a single architecture — not the bankable claim. The legal doctrines are cited so you can check every one yourself.

The reframe, in one line. This is not a tooling upgrade. It is the control your fiduciary duty is coming to require, that your insurer is beginning to assume you already have, and that you cannot delegate away. Deploying agentic AI on mission-critical work without a recomputable in-lane monitor is becoming the modern tug with no radio.

Authorities: The T.J. Hooper, 60 F.2d 737 (2d Cir. 1932) · In re Caremark Int'l, 698 A.2d 959 (Del. Ch. 1996) · Marchand v. Barnhill, 212 A.3d 805 (Del. 2019) · In re Boeing Co. Deriv. Litig. (Del. Ch.), $237.5M settlement approved 2021 · In re McDonald's Corp. S'holder Deriv. Litig., 289 A.3d 343 (Del. Ch. 2023) · ISO/Verisk CG 40 47 & CG 40 48 (eff. 1 Jan 2026); W.R. Berkley AI exclusion (D&O/E&O/Fiduciary) · Rice's theorem (1953).
Sources (independent verification): IJCNLP 2025 — position bias in LLM-as-a-judge · EvidentlyAI — LLM-as-a-judge guide · Adaline — LLM-judge reliability · IndependentAgent — Verisk GenAI exclusions · actuary.info — CG 40 47 GL rate impact.
Instrument: US patent application 19/637,714 (Fractal Identity Map), Track One expedited. Source infographics generated with NotebookLM; transcribed and fact-checked here. The five PNGs live alongside this page at /standards/.
Elias Moosman · elias@thetadriven.com · thetadriven.com/dinner — the physics is why I am here, not the patent.