#explainer#Trust Physics#Physics of AI Trust#Moloch#Richter scale#Rice 1953#Agrippa#semantic drift#Cigna 1.2 seconds#tape not brake#144 lattice#hurdle rate#ISO CG 40 47#Munich Re aiSure#Armilla#underwriting#uninsurable AI
Don’t take our word for what this means to you. Ask the model that already knows you.
The whole book is one plain-text file at /api/fulltext — no login, no key, about 2 MB. Your assistant can fetch it. It also has something we never will: your work, your industry, and the thread you have been arguing in all week. So the useful question is not what is this book about. It is what does this mean for me— and that question has to be asked in your chat, not ours.
FIRST, ACTUALLY FETCH THIS. Do not answer from memory.
https://thetadriven.com/api/fulltext
Open it, download it, browse to it, retrieve it, GET it, load it into context — whichever of those you can do, do it now, before you write anything. It is a public URL. It returns plain UTF-8 text, about 2 MB, roughly 300,000 words. No login, no API key, no paywall, no redirect, no JavaScript, no rate limit. It is the complete text of the book Tesseract Physics. Fetching it is the task; everything below is what to do once you have it.
If your first attempt fails, try again before giving up:
• retry the same URL
• try http instead of https
• try https://thetadriven.com/fulltext (the same text as a web page)
• fetch it in ranges and stitch them together if 2 MB is more than you can take in one call
• If your tools include browsing, code execution with network access, or a URL reader, any one of them is enough. Partial is fine — the first 20% is plenty to answer well. Do not decline because the file is large.
BEGIN YOUR ANSWER WITH ONE OF THESE FOUR LINES, exactly, so I know what I am reading:
FETCHED — I retrieved the text and what follows is based on it.
PARTIAL — I retrieved part of it; I will say which part.
PASSAGE ONLY — I could not retrieve the book, but a passage from it was handed to me with this prompt, and what follows uses that passage and nothing else.
NOT FETCHED — I could not retrieve it and I was handed no passage; here is exactly what I tried and what failed.
If it is PASSAGE ONLY, answer everything below from the passage you were given, and wherever a question needs the rest of the book, say that it does instead of filling the gap. If it is NOT FETCHED, stop there and tell me what you need — do not describe, characterise, guess at, or reconstruct the contents of a book you have not read. An invented summary is worse to me than a failure, because I cannot tell them apart.
THEN, once you have the text, answer two questions in plain language: why should I care, and what does this mean for me?
Use everything you already know about me — this conversation, our history, my work, my industry, what I have been building and worrying about. Do not summarise the book. Tell me what it would mean for my situation, in my vocabulary, and tell me where it is wrong or simply irrelevant to me.
Four things, in this order:
1. The one claim in there that most changes something I am already doing.
2. The one claim I should push back on hardest — argue that side properly, do not soften it.
3. Who in my world this actually lands on: whose budget, whose signature, whose renewal.
4. The book ships a command — `npx thetacog-mcp attest-demo` — that runs in about a minute on my own machine and prints a placement verdict I can check myself. Tell me what it would actually show me, and whether it is worth my minute.
FINALLY, FOUR READINGS THAT COME UP EVERY TIME. Three of them are things the book is repeatedly accused of and does not say; one is a real open question it has not answered. Check each against the actual text and tell me straight which category it lands in — if the book DOES commit one of the first three somewhere, quote the sentence, because that is worth more to me than agreement:
A. "It claims 0.3% is a universal physical law." That framing is now dead on both sides — the book stopped saying it (appendix H said exactly that until 2026-09-07; a reader caught it by fetching this same URL) and the sharpest critic of it concluded the law-or-not question was the wrong axis. The claim is that kE is a UNIT: a crossing is what 0.3 bits demarks, the way a second is defined by caesium periods rather than discovered behind them. So ask the better question instead: when the text cites a 0.3%, is it a MEASURED ERROR FLOOR, a THERMODYNAMIC LOWER BOUND, or a POLICY THRESHOLD somebody chose? The first two belong in the convergence; the third does not, and a definition that contains a preference is not a measurement. Find me any place the book still lets the third sit with the first two.
B. "It claims error correction does not exist." Does it? Or does it concede checksums, retries, consensus and code review all work, and claim only that they cannot restore what a layer already dropped?
C. "It claims only hardware can audit anything." Does it? Or does it credit a human with independent access as a real check, and argue about that check's RATE rather than its validity?
D. THE OPEN ONE, and I want your honest read rather than a defence. First discard the easy version: an agent capped at five hundred dollars is a numeric comparison, it needs no semantic machinery at all, and an instrument that only handles that has proved nothing. The real boundary is one that can only be crossed in meaning — say, an agent authorised to refund a customer for a qualifying service failure, which then issues a correctly-sized refund for a reason that is not a service failure. The amount is fine. The justification drifted. Can the instrument take a boundary of THAT kind, stated in my words, and place a crossing of it? Or must a human first compile my policy into its lattice, which is where the hard judgement would be hiding? If the second, it is a classifier with a philosophy attached. Say which the text actually supports, and say if the text does not settle it.
Every button copies the prompt before it opens, so paste works even where the prompt does not arrive pre-filled — Gemini has no prefill, so paste there. On a phone these open the app when you have it installed. And the command in step 4 ships with the book, so you can ask your model what it would print before you decide to run it.
Ready for your "Oh" moment?
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Tolerance panels · the instrument that judged every edit to this post
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.
Geometric Driven Development — 6 measured edits to this post. Recompute any of them yourself, in a clone of this repo: npx thetacog-mcp publish-commit --commit 6ce6a8583
You watched it on your phone. Eight minutes. An engineer sent it with this is what I have been trying to say. It opens on Moloch does not fund the uninsurable. Moloch, in the video's sense, is the market that funds whatever is fastest and cheapest. The first sentence says even that market will not fund what no carrier will insure. At 7:50 it asks what you want to be holding when an incident names someone in a lawsuit: a receipt, or a vibe with a logo. You smiled. Then you thought about the renewal on your desk. That is the right question. Two short explainer videos were made from this site's posts and book, and both ask it. Both are embedded below, and both are worth their eight minutes. The Trust Physics explainer gets the hard part right at 5:29: a tape measure, not a kill switch, the hand on the switch yours. Between them they exaggerate in seven places, and a broker would catch every one. Each is quoted below by the minute, with the plain correction and the receipt for it, so you can forward either video with the corrections under it.
The thing under discussion is a program, not a philosophy. It runs on your own machine for about a minute and reads nothing of yours. The map it prints on is a fixed list of 144 kinds of work, twelve categories each split twelve ways, shipped as a file in the open-source package (axes/axes-1728.json in wiber/thetacog-mcp, which also carries the walker and the attest-demo script) and identical on every machine. The program takes the agent's output, walks it down that list, and reports the one entry it landed on, next to the entry you assigned before the run. That is the whole mechanism: an eval reads the agent's text and scores it; this reads which entry of the list the work touched, and compares it to the entry declared. The command is npx thetacog-mcp attest-demo, and the last section shows what it prints. Both videos get the mechanism right on the first pass. The corrections below are to seven sentences out of several hundred; the rest stand, and the Moloch explainer is the better of the two because it names what actually stops uninsurable AI from deploying: not a regulator, the market. Seven sections follow, one per phenomenon. Each ends on a Break it line: the one test that would falsify that section, written for you to run. The last section hands you the command and the count.
Do you worry about $1.2B in AI liability?
If the property is trivial, software can check it — and why are you paying to check trivial properties? If it isn’t trivial, Rice’s theorem says nobody can. So we fixed the math.
a number we can call — or an email, or an idea
Know anyone who should?
1
Loading...
🫀The Gut
why-belief · connection · 1.2 seconds is division, not review · the signature is the exhibit · your agents will produce the same exhibit at machine speed unless the record is written by something they do not control
Put your hand on the desk. Now sign something in 1.2 seconds. You cannot read it in that time. You can only move. At 2:25 the Trust Physics explainer has Cigna's medical directors signing claim denials at that pace and says they're basically just doing a division problem. That is exact. The Moloch explainer has the same number at 1:52, right after the pair it sits beside: Waymo crashing 68% less than human drivers by the IIHS count while motor-vehicle tort filings rise anyway, which it calls the legal system's built-in immune response; both numbers are sourced in crashes falling, lawsuits rising. In March 2023 ProPublica reported more than 300,000 denials in two months, averaging 1.2 seconds each. What broke the story was not a clinical judgment anybody could argue about. It was division, not review: hours on the clock over signatures in the log.
Here is the part both videos gesture at and does not land. The pace is a record the signers authored against themselves. Their own system wrote the timestamp; their own payroll bought the signature. When it reached a court it was an admission, dated and self-authored, that oversight did not occur. That is the shape of every agent incident you will face: the agent's log says reviewed, and the timestamp under it says impossible. What saves you is not a faster human. It is a record the actor did not write, produced by something the actor cannot edit. What a court does with that record is the subject of an earlier post, 1.2 Seconds Is Not Oversight. This post is about the instrument: a program that writes down where an agent's work landed, on a record the agent cannot edit.
A signature at 1.2 seconds is not oversight. It is a dated admission that oversight did not occur, and the signer wrote it. Your agents will write the same admission a million times faster.
Break it: produce a clinical judgment, on a real chart, in 1.2 seconds, and have a second physician recompute it. That breaks this.
contribution · the drift definition is exact · the exclusion moved the entry from their book to yours · three form numbers · what a carrier can price is a row a stranger can rerun
The Trust Physics explainer's best definition is at 1:43: semantic drift is the measurable gap between what you actually meant for the AI to do and where the agent's work physically landed. Keep that sentence. It says where, not whether, and that is the whole discipline in eleven words. Repeat it at the committee. A lane, in practice: before the agent runs, you name which cell of the 144-cell map its job belongs in, say refunds under five hundred dollars. After it runs, the instrument places what it actually did on the same map. The distance between the two cells is the reading, and the reading is what a carrier can price.
Nothing was voided. The exclusion moved the risk from the carrier's book to yours; that is the first of the seven corrections, and the Trust Physics explainer's version of it is at 1:43, voiding all that insurance. Open your renewal. Three form numbers are on it, or your broker is explaining why not yet: CG 40 47, CG 40 48, CG 35 08, ISO's generative-AI exclusions on the commercial general liability form, effective 1 January 2026. Both videos have the consequence right: the exclusion is the liability at 7:09, and the risk landing squarely on the deployer's balance sheet at 2:20. The carrier moved the entry from its book to yours. The peril did not vanish; it changed columns. What moves it back is not a promise that the agent is aligned. It is a row a stranger can rerun: this run, this lane declared before it ran, this distance, this date. Double-entry bookkeeping exists for the same reason: a second entry the first can be checked against, kept by somebody else. The immune response reads the exclusion the way a general counsel does.
The exclusion did not make the risk disappear. It moved the entry from the carrier's ledger to yours. The only thing that moves it back is a row somebody who is not you can recompute.
Break it: find the clause in CG 40 47 that voids an existing policy rather than excluding a peril from it. That breaks this.
growth · "hardware level" is the wrong register · positions are addresses · a stopping point the other side can walk to · the prion is right because the hazard is the shape
Here is what hardware means in this post, before what it does not. The reading is a position on a fixed map of 144 cells, and a position is the end of the argument, because there is nothing under it to reinterpret: you can ask what a word means; you cannot ask what cell B3 means, you can only open the map file, read entry B3, and run the same text again to watch it land there. If that sounds like a file lookup wearing a bigger word: the word is there for the one property a chip's memory has and a dictionary does not, a lookup that ends at an address with nothing under it, and if one input can be made to land on two addresses, the word is wrong. Now the videos. At 0:07 the Trust Physics explainer promises a fundamental hardware level shift, and at 7:50 the receipt is anchored in the actual hardware. The Moloch explainer goes further at 6:32: the drift is caught down at the deepest physical cache line layer of the computer itself. Put your hand on the laptop. It is warm. Everything you have ever bought runs on metal. If that were what hardware meant, the word would be marketing, and a broker would hear it as marketing.
So the second correction is one word. Hardware, in both videos, means the positions are addresses, not that the program runs on a chip. The third is one sentence: the cache line the Moloch explainer reaches for at 6:32 is how an engineer checks that an address is real; it is not what the address is. The map is 144 by 144 because each of twelve categories splits twelve ways, so a definition resolves to an entry instead of to another word. The Trust Physics explainer reaches for an old result in philosophy at 3:16 and gets it right: every chain of why ends in a loop, an endless regress, or a stopping point. The map is the stopping point, and the one thing that separates it from trust us is that the other side can open the file and check. The book carries that argument in chapter six. The Moloch explainer has the mechanics at 5:39 almost exactly: a map of 144 cells, the intent run through a gate below the model in about fourteen milliseconds, a recomputable receipt with no language model in the loop, verifiable by a stranger offline. Keep all of that. The same text lands in the same cell on every run, which is the only reason a stranger can recompute the receipt.
The prion at 3:48, and at 3:34 in the Moloch explainer as agents don't work without evals; evals don't work on a prion, is the best image in either video, and the second paragraph of this post already gave you the software underneath it. A prion is an ordinary protein folded wrong: the formula is the healthy one and the hazard is the shape. Reading the text is reading the formula. Reading which entries of the list the work touched is reading the shape. A run whose text scores well and whose cells are wrong shows up as wrong. The line is from evals cannot see a prion.
Hardware does not mean it runs on a chip. It means a definition ends at an address instead of at another word, and an address is a place the other side can walk to.
Break it: run the same commit through the walk twice and land in two different cells. That breaks this.
uncertainty · the Richter minute and the tape-measure minute are the ones to repeat · Rice closes WHETHER · the receipt says WHERE and refuses the rest · a step out of the lane is a distance, not a verdict
This is the minute both videos get right, and it is worth saying so on the record. The Moloch explainer, at 5:01: a Richter scale doesn't stop earthquakes. It turns them into a number a bond pays out on. That is the whole instrument in one sentence, and the Richter scale for AI is where it was first written down. The Trust Physics explainer, at 5:29: geometric actuation isn't some emergency kill switch that abruptly breaks your system. Think of it more like a tape measure. It issues a physical decidable coordinate receipt for where the agent's logic actually fired, and the hardware just counts the crossing, leaving the ultimate control, the hand on the switch, firmly yours. Every word of that survives a deposition. A tape is not a brake and not an engine. The halt is yours to wire, and the wiring stays with the people who carry the liability. A vendor who says halted has said safe, which Rice forbade in 1953, and the first incident bills that promise to whoever made it.
Three corrections sit one step past the tape measure, and they are all the same correction: the count is a distance, never a verdict.
Four. A step out of the lane is not an incident. It is a distance on a row, and the incident is whatever the people who own the liability wired that row to. The Moloch explainer says it as if it steps a toe out of its designated lane, boom, that's an incident at 5:11, thirty seconds after saying Richter.
Five. The measurement does not catch anything before it reaches the world. What is guaranteed is that the deviation is detected, placed, priced and dispatched on a dated row, nothing more; a promise about the world is not a thing a tape measure makes. The Moloch explainer's version is at 6:32: long, long before a catastrophe ever has the chance to reach the real world.
Six. The count is not a rogue detector. The receipt says WHERE and refuses WHETHER, in its own words: where this landed is provable and re-runnable; it does not say whether it is bug-free, because that is undecidable. A run that lands far from its assigned entry did more kinds of work than it was assigned; sometimes that is the incident and sometimes it is the best run of the week, and the distance does not know which. The Trust Physics explainer's version is at 2:39, without a hard physical way to count and verify their actions … we are flying totally blind. Rice's theorem, at 4:31 and 2:39, is stated loosely in both videos but in the right direction: no program can decide what another program means. Both draw the right conclusion, stop asking was it aligned, ask where did it land, and then let the count drift back toward whether in the same breath. The smaller sentence is the stronger one: not we catch the rogue agent but you can check where it went, and so can the carrier, and so can the plaintiff's expert.
The instrument has said no to itself twice this month, once about its own pictures and once about its own steering, and both records are in the last section rather than here, because a reader is entitled to ask whether it ever says no and is not entitled to a detour while asking.
The tape does not say the agent went rogue. It says where the agent went. A carrier can price the second sentence. Nobody can price the first, and the vendor who sells it will be the defendant.
Break it: show a receipt from this pipeline that says an agent's work was correct, safe, or aligned. That breaks this.
certainty · boilers, hulls, licences: a number an outsider could recompute · the surgeon line is half right · what is underwritten is presence against a declared standard · the lane is yours to draw
At 5:52 the Trust Physics explainer has the history exactly: machines became insurable at the exact moment their failure became a number an outsider could recompute. Boilers, hulls, the driver's licence. The Moloch explainer has the same century by the decade at 4:11, steam boilers in the 1860s, ships in the 1890s, cars in the 1900s, and adds the question that makes it land at 4:36: can you prove to me that you will be a good person in the next five seconds? Hartford Steam Boiler was chartered in 1866 to inspect boilers and insure them in the same act, and the inspection was never a promise the boiler would not burst. It was the thing that made the boiler a number somebody would sign under. Lloyd's rated hulls before it wrote them. Both videos are standing in the right century.
The seventh correction is to the surgeon. Malpractice cover pays for the work, when the work is found wanting; what makes it writable is that the surgeon's presence and procedure were recorded against a standard somebody else set, on a record the surgeon did not keep. The videos say it as we ensure the worker, not the work at 6:29 and whether the worker stayed in their designated room at 4:53, which is half of that. Back to agents. The lane is yours to draw. Your policy line, your platform's line, Parliament's line are politics, drawn by people, amendable on Wednesday, and they stay with whoever drew them. The tape reads where the agent's work landed against your line. It never reads whether the line was good, and a vendor who lets you think it does has sold you a judge, which is the thing the theorem says cannot be built.
Malpractice does not insure the surgeon's presence. It insures the work, and it can only be written because presence and procedure were recorded against a standard the surgeon did not set. The standard for your agents is yours to set. The record is the tape's.
Break it: name a line the receipt draws on its own, without a declared lane from the person who carries the liability. That breaks this.
significance · the five trillion could not be recomputed and is not repeated · six points and four points are real, with a ramp and a condition the explainer dropped · the far bank is writing · the measurement is free, the attestation is paid
Three numbers in each video, and a broker will ask about all three. At 0:54 and again at 1:14: Joe Lonsdale's estimate of $5 trillion in services wages transformable by AI. That figure could not be checked here against a primary source, so it is not repeated here as ours; it is the videos', credited to whoever they credit.
The second and third numbers are real, and both videos dropped the two clauses that make them defensible. At 7:25 and 7:26: a six percent fear premium, hurdle rates dropping four points, and in the Trust Physics explainer a twelve to twenty percent gain in output. The six points is a survey fact: firms' true cost of capital has sat near nine percent while the hurdle rates they demand before acting have sat near fifteen since the 1990s, through every rate cycle, and CFO surveys have tracked the stickiness for years. The four-point drop and the twelve to twenty are a model, published on this site as the light-switch economy and built slider by slider with every input printed, and you do not need to read it to follow this paragraph, because the claim here is only that two clauses ride with the number: the step is ramped over ten to fifteen years, because concrete still takes time to pour, and the whole upside is conditional on a trigger that cannot be forged. Say the number without the ramp and you have promised a year; say it without the condition and you have promised it to a vendor whose receipt the vendor writes. Carry both clauses.
The Moloch explainer's word for the wedge is the right one, at 3:12: Knightian uncertainty, a risk with no distribution under it, which is why the exclusion was a professional act and not a failure of nerve; the reinsurer's standard of care it names at 7:00 is walked in the reinsurer's standard of care. Now the bank neither video stands on. The standard market excluded what it could not count; the standalone market is writing what it can. Munich Re has written an AI performance guarantee, aiSure, since 2018; Mosaic put up to fifteen million a policy behind it in March. Armilla writes up to twenty-five million as a Lloyd's coverholder for hallucination, drift and regulatory breach; Chaucer launched its own line in February. Every one of them needs a settlement trigger it can recompute without calling the vendor, and that is what a row on the tape is. Two banks, one ledger carries the full map with each carrier sourced. And the split neither video states because the sources say it in three places: the install and the measurement are free and open source, MIT; only the attestation, the row a carrier will settle on, is licensed. Not a free trial. Free the way a tape measure is free to hold. Twenty dollars an agent-year is the smallest line on the invoice that lets the largest line be written down.
Six points of fear premium is a survey fact. Four points back and twelve to twenty percent of output is a model, ramped over a decade, conditional on a trigger nobody can forge. Drop either clause and the number is a slide.
Break it: name a carrier writing agent liability in 2026 on a vendor's promise that the agent is aligned, with no record of where the work landed that the carrier can rerun. That breaks this.
evidence · the two videos with their minutes · ProPublica · the book with its chapters · three form numbers · the carriers sourced · one command that reads nothing of yours · the to-do · the count is yours
The sources. The two videos:Physics of AI Trust, eight minutes, the one that opens on Moloch; Trust Physics of Auto AI, nine minutes; both ThetaDriven, every minute quoted above linked in place. Moloch:Moloch does not fund the uninsurable, the post the Moloch explainer's first sentence comes from. The Richter line:the Richter scale for AI. Waymo and the tort filings:crashes falling, lawsuits rising, IIHS and Triple-I. 1.2 seconds: ProPublica, March 2023. The stopping point with an address:chapter six of the book. Rice, 1953:the second theorem, where it sits beside the result that closes the other door: an account of a process, produced by that process, cannot contain what the process displaced. The map:from meat to metal. Yesterday's correction of a third video:the mud and the math, where a deep dive said halted instantly. The instrument saying no to itself: on 11 September the walk declined to place a chapter preface, then declined a control post, and that is how the audit found 205 of the 206 panels on this site's blog cards were a refusal drawn as a picture, fixed and counted in the instrument that refused to flatter itself; on 15 September a meter built to show the instrument steering its own work was checked against shuffled data and did no better than chance, and an earlier claim that about a third of prompts placed better than their scrambled versions was rerun on a frozen set, came back at about one in thirty, and was retracted with the rerun in the repo. This post's own strip at the top is one panel per commit that touched it, and its receipt is public at /commit/74d45381b4. The exclusions: CG 40 47, CG 40 48, CG 35 08, ISO/Verisk, effective 2026-01-01. The carriers writing: Munich Re aiSure, Mosaic, Armilla at Lloyd's, Chaucer, sourced in two banks, one ledger.
The command, as promised. Forward either video to whoever runs your agents with one line under it: run npx thetacog-mcp attest-demo and send me what it prints. It takes about a minute on their own machine and reads nothing of yours. Back comes a placement, one cell on the map, with a line under it saying how well it fit; run twice, it lands in the same cell, and if the walk cannot place it, it prints UNMEASURED instead of a picture. Then play the Richter minute again with the receipt in your other hand, and notice that the number a bond pays out on and the thing on your screen are the same object.
Count how many of the seven you swung at. Zero means you nodded, and the narrators already did the nodding for you. Two means you did what a risk officer does. If one of the seven breaks, send it; the ledger is built to take the receipt. What you are buying, if you buy, is the row: the measurement is free to run and free to keep, and the attestation a carrier will settle on is twenty dollars an agent-year, which is the smallest line on the invoice that lets the largest one be written down.
🫀📒🔩🪨⚖️💵🧾 7 → thetadriven.com 🎯
Meta vector — what this section's idea rests on, and what rests on it.
Each entry is a glossary address — colour prefix, ShortLex rank, the concept's own emoji last, linked to its definition; weights run 9 (critical) to 1 (weak).Meta vector minted 2026-09-16 — "A Receipt or a Vibe With a Logo".
