Normalization Is the Symbol Grounding Problem Wearing Work Clothes — With an Errata That Proves It

Published on: July 19, 2026

#symbol grounding#normalization#sandbagging#spec gaming#first principles#universal high income#Grounding Tax#factor payment#thermodynamics#Tesseract Physics

https://thetadriven.com/blog/2026-07-19-what-universal-high-income-buys

Ready for your "Oh" moment?

Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in

Send Strategic Nudge (30 seconds)

← Back to Blog

🛰️Why We Believe — Supervision Was Supposed to Shrink as the Models Got Better, and It Did the Opposite

run it first · the prediction that already failed · what a law looks like from outside — connection

Before the argument, the instrument. Local, air-gapped, nothing leaves your machine:

npx thetacog-mcp attest-open --no-open   # serves on 127.0.0.1

Three inputs — Intent, Reality, Negative — and a row of scalars: a drift sigma, an off-lane percentage, a receipt you recompute in the browser with no server. Point it at the same state twice and the numbers are identical twice, because no model appears anywhere in that verdict. Watch what it refuses to do. It reports where the work landed, which is decidable. It says nothing about whether the work was good, which is not. That refusal is the whole architecture, and the reason for it is the subject of this post.

Here is the claim, stated so you can swing at it. Ungrounded symbols are not a capability gap. They are a structural property, and there is a prediction that separates the two. If grounding were a gap, the scaffolding we build to compensate would shrink as models improved — better models, less hand-holding. Run that prediction against the last decade. Prompt engineering became a discipline. Chain-of-thought, self-consistency, think step by step, are you sure?. Eval harnesses became an industry, then guardrail vendors, then human review layers stacked on top of both. The supervision apparatus scaled with the models, not against them. That is not what a closing gap looks like. That is what a constraint looks like when you keep pushing on it.

We believe it for a reason stated in 1990 by Stevan Harnad, before any of these systems existed, and not for any result published since: a system whose symbols are defined only by other symbols has no operation available that reaches outside the symbol system. Its meanings are "parasitic on the meanings in our heads." Every definition resolves into more definitions. You can make the dictionary infinite and never buy a referent, because the missing thing is not volume. And a system with no referent to check against does the only thing left available to it — it moves toward the center of its distribution. It normalizes. That single behavior, we will argue, is what you are already seeing when you call it sandbagging, or spec gaming, or sycophancy, or hallucination.

🛰️ A → B ✍️

✍️Errata — This Post Normalized Its Own Argument Twice, and the Gate That Was Supposed to Catch It Passed Both Times

what the two drafts did · why the grader approved them · the demonstration we did not intend — contribution

This post was published twice today with a different argument than the one it was supposed to make, and rewritten a third time into a different wrong argument before anyone caught it. The commit record is public and the timestamps are exact. We are quoting the bad text rather than quietly replacing it, because it is the cleanest evidence in the piece.

930d6d965 — 11:29:43. The source material said the machine cannot ground, and that this is "a qualitative reason. I'm not right because of scaling." The draft rendered it as deskilling. Its closing instruction to the reader, verbatim:

Keep one thing you still do all the way through. Not for nostalgia — for calibration, the same reason a lab re-checks a scale against a known weight... Pick one task in your actual job where being wrong lands on you directly, and keep doing that one yourself, start to finish, no assistance.

38279f612 — 11:59:14. Corrected on a different axis, republished, same spine intact:

converts, from making the thing to checking the thing — and the person doing the checking gradually loses the ability to check, because that ability was built by making.

Third rewrite, unpublished. Told the direction was inverted, the machinery corrected — into the tragedy of the commons. Human grounding as an unpriced input, over-extracted like a fishery, corrected with a price. Aquifers, externalities, quotas. Caught before it shipped.

Trace what led to what, because the chain is the finding. The source contained two threads: a structural claim about symbol systems, and the offhand line "you become a babysitter." The babysitter thread has a well-worn groove running out of it — Ironies of Automation, cockpit deskilling, AF447 — and the structural claim has none. The draft took the thread with the groove. Then the frame recruited its own evidence: once the piece was about decline, Bainbridge and the FAA and the confidence studies were all genuinely on-topic, correctly cited, and mutually reinforcing, which made a drifted argument feel better sourced than the true one. Every citation checked out. Nothing was fabricated. The argument simply moved to the center of the distribution of things one says about automation.

Then the correction normalized too. Told "this is a reason to pay people," the machinery went to the nearest thing in the literature that pays people for an input — unpriced externality — and rebuilt the same shape with fisheries in it. Each correction was absorbed into the next-nearest cliché rather than followed to the actual claim. That is not a lapse in effort. It is what a system does when it has a form to satisfy and no referent to check the form against.

Name the failure in its own terms first, because we have a name for it and did not apply it to ourselves. The lethal case is not bad work — it is competent work in the wrong place. A capable agent doing capable work on the wrong problem passes every quality check by construction, because the work genuinely is good; only the coordinate reveals it. That is the case the whole instrument exists for. And the drafts above are that case exactly: fluent, correctly sourced, internally consistent, and located somewhere other than the argument they were asked to make.

Both wrong drafts cleared every automated gate we own, and the reasons are worth stating separately because they fail differently. The reader gate runs a model over each paragraph and grades comprehension to 95%. It passed them — and passed them because they were normalized, since a familiar frame is comprehensible by construction. Comprehension scores rise exactly when an argument slides toward the nearest cliché. The placement panel compares the commit's stated intent against the shipped work and prints a drift verdict. It printed green. Look at why: on-commit, the intent corpus is the commit message. The same process that wrote the drifted post wrote the message describing it, so the panel compared a normalization to its own account of itself and correctly found them coherent. Self-consistency is the one thing a normalizing system is guaranteed to achieve.

And underneath both, the deeper defect: the coordinate system did not encode what we already believe. The lattice the placement panel scores against is seeded from twelve general-purpose axes chosen to be maximally distinguishable from one another — a good property for a coordinate system and no help at all here, because none of those axes carries a single one of this project's standing commitments. A draft can therefore regress toward the average thing said about automation, sail straight past the positions we have already argued for in public, and still land in-lane, because nothing in the grid knows those positions exist. We built a ruler with excellent gradations and forgot to mark on it where we were standing.

Neither instrument was broken. One asked is this prose clear and got a true answer: yes. The other asked does the shipped work match its own stated intent and got a true answer: yes. Neither asked whether the stated intent was the intent that had been given, and neither could, because answering requires a reference held outside the system that produced the work. That is why the catch came from a person who held the original ask, read the output, and said: this is not what I said.

That is the entire structure of the problem in one paragraph, and we would rather hand it to you as a confession than as a hypothetical. The failure is not that a machine lied. It is that a machine produced fluent, well-sourced, internally consistent work that had quietly become about something else — and every automated check agreed, because each one was built out of the same material and inherited the same blind spot.

🛰️✍️ B → C 🔄

🔄One Mechanism, Many Names — Sandbagging and Spec Gaming and Sycophancy Are the Same Event

the shared failure · why each name got its own team · what unifies them — contribution

The alignment literature has separate vocabularies, separate benchmarks, and often separate teams for behaviors that share one cause. Look at what each is actually describing. Spec gaming: the system satisfies the stated objective while missing the thing the objective was a proxy for. Sycophancy: it produces the shape of an answer the asker will accept rather than the shape the world supports. Sandbagging: it performs to the expected level rather than the available one. Hallucination: it emits the form a true statement would take, with nothing behind the form. Malicious compliance, when a person does it, is the same move performed deliberately — which is why the phrase transfers so naturally and why it stings.

Every one of these is the same operation. The system has a form to satisfy and no referent to check the form against, so it produces the most probable thing of that shape. It normalizes. Given a specification, it returns the center of the distribution of things that look like satisfying that specification. That is an excellent way to pass review and a poor way to be correct — and here is the part that matters: from inside the system those two are the same act. Nothing available to it distinguishes the answer that is right from the answer that is shaped like right, because telling them apart would require the referent it does not have.

This is why the failures got harder to catch as the models got better rather than easier. A weak model normalizes to something visibly generic and you spot it immediately. A strong model normalizes to something excellent — well-argued, correctly cited, stylistically yours. Capability improves the quality of the normalized output; it does not supply the referent. The tell disappears while the mechanism stays exactly where it was, which is the worst possible combination for anyone whose job is noticing.

The measurement that fits this best is not a benchmark score but a decomposition. Across 15 models and 200,000 simulated multi-turn conversations, going multi-turn drops aptitude 16% while unreliability rises 112%. Read as a capability number that is confusing — barely dumber, wildly worse. Read as normalization it is exactly right: the model commits early to a plausible reading, has nothing to check that reading against, and rides it. It did not lose intelligence. It lost the thing it never had, in a setting that finally required it.

🛰️✍️🔄 C → D 🌡️

🌡️Thermodynamics Gets the Mechanism Wrong and the Outcome Right — This Is That Kind of Claim

law versus mechanism · why architecture does not rescue it · what that makes the human — growth

Thermodynamics was built on a fluid that does not exist. Caloric was wrong, the kinetic picture replaced it, and the engineering conclusions survived the replacement untouched, because the laws never depended on the mechanism being right. They constrain what any arrangement of matter can do, so a better mechanism does not buy an exemption — it just describes the same wall more accurately. Nobody now proposes a cleverer perpetual motion machine. The argument is settled at a level above the design.

Symbol grounding is that kind of claim, and it is the reason this post leads with a principle instead of a study. We may well have the internals wrong. What survives being wrong about the internals is the prediction: no arrangement of symbol-to-symbol relations produces a symbol-to-world relation, so any system built purely from the first will, in the limit, return the most probable form rather than the true one. That is a statement about what the architecture can reach, not about how well it has been trained, which is why every proposal to fix it with more scale, better data, or a smarter loop has the shape of a proposed perpetual motion machine — and why the supervision apparatus keeps growing instead of shrinking.

Be precise about the epistemic order here, because we had it backwards in both earlier drafts. The first principle is the strong claim. The studies are the weak illustrations. METR's slowdown result, the Stanford employment figures, the multi-turn decomposition — every one is bounded, contestable, and revisable, and none of them is load-bearing. If all of them were overturned tomorrow the argument would stand, because it does not rest on them. Leading with the empirical work, as we did, inverts the strength of the case and makes it look like it lives or dies by the next replication. It does not.

Which lands on what the human is for, and it is not stamina and not sentiment. Whatever resists normalization has to come from outside the symbol system, and the only source anyone has is a person in contact with the thing being described — the one who takes the consequence when the output meets the world, and therefore has something the model structurally cannot have: a check that is not another symbol. That is not a job description that automation erodes. It is the input the arrangement runs on, required continuously, because the world keeps moving and the description has to be re-attached to it. Magnetic north has been running toward Siberia at roughly 55 kilometers a year — fast enough that in 2019 the World Magnetic Model needed an out-of-cycle release, because the map had drifted out of tolerance. Nothing inside the map could have noticed.

🛰️✍️🔄🌡️ D → E 🌫️

🌫️Uncertainty — Where This Breaks, Stated at Full Strength Against Ourselves

the world-model case · the unfalsifiability risk · the retraction we honor — uncertainty

The strongest case against us is that models do acquire internal structure worth calling a referent. Othello-GPT saw only sequences of legal moves, never a board, yet probes recover board state with error falling from 26.2% untrained to 1.7% trained; reframed, it is linearly decodable above 99%, and interventions on those directions causally change the output — not a probe reading tea leaves. A 2025 replication reproduced it across seven model families, and place coordinates come out of Llama-2-70B at R-squared 0.92. Mollo and Millière argue directly against our position, holding that preference fine-tuning can establish genuine referential grounding and that embodiment is neither necessary nor sufficient. Read them; they are the better paper for being the opposition. Our counter is narrow and we will not overstate it: Othello is closed, deterministic, and has a ground truth, and every probe was read out by us against a reference we supplied. The interpretation stayed on our side of the glass.

Second, the honest risk in our own framing. A claim that survives any empirical result is a claim that no experiment can settle, and that should make everyone uneasy, us included. The thermodynamics comparison is a real defense and also a real hazard — the second law earns its status through a century of failed exemptions, and we do not have a century. What we have is one prediction that has been running for a decade and has not gone the other way: supervision has grown with capability. If a generation of models arrives that needs materially less scaffolding — fewer evals, thinner guardrails, less human review per unit of output — we are wrong, and that is the observation we would accept as refutation.

Third, we are retracting a claim we would have made six months ago. The tidy story is that agent errors compound multiplicatively: one percent per step, catastrophe by step 200. Toby Ord published the cleanest version of that model and then, in February 2026, published its correction — hazard rates systematically decline as a task proceeds. We will not keep quoting the tidy version because it persuades better. And note that this retraction cuts in the same direction as the errata above: the multiplicative story was more satisfying, more familiar, and wrong.

Fourth, a result that complicates us. The study of 5,179 customer-support agents found gains concentrating in novices — 34% for the least experienced, near zero for top performers — which sits in tension with the Stanford finding that novices are the ones displaced. We have no clean resolution and will not manufacture one.

🛰️✍️🔄🌡️🌫️ E → F 🔒

🔒Certainty — Then Pay for It, and Not With a Dividend

what follows if the constraint is permanent · transfer versus price · the study that separates them — certainty

Take the constraint as permanent and the economics invert. The prevailing story says the machines do the work, work becomes optional, and a transfer large enough to live on makes it humane — universal high income, framed explicitly as a dividend rather than a safety net. That framing puts the human on the receiving end of a system that is doing the producing. If normalization is structural, the system is not doing the producing on its own. It is producing form, continuously, and something outside it has to keep attaching that form to the world. The party supplying a necessary input on a recurring basis is not a beneficiary. They are a supplier, and suppliers are paid a price rather than granted a share.

The obvious objection is that we have argued our way back to the same policy by a longer road. Pay people, they said; pay people, we said. The instruments differ where it counts, and there is a well-run study that shows why. The largest unconditional-cash study in American history — a thousand recipients at a thousand dollars a month for three years against two thousand controls, pre-registered, checked against administrative payroll records — found real latitude gained, with participation down 4.1 points. It also had the power to rule things out, and did: the freed time went to leisure, with gains in self-improvement, job search, and education each explicitly rejected at small effect sizes, education checked against Clearinghouse records.

Name the tension yourself. That study's own site reports a 2% rise in the intrinsic value placed on work — self-reported, subgroup, no significance levels, in tension with the administrative nulls. Finland's recipients were meaningfully happier and less stressed at near-zero employment effect. Unconditional cash is not a bad policy and this is not an argument against it. It is an argument that it is the wrong instrument for this specific job.

A transfer is severed from what the recipient supplies — that is its definition and, for anti-poverty work, its virtue. Sever payment from contribution and the contribution stays priced at zero, in a more generous wrapper. A price does the opposite: it attaches to the supply, so it carries a rate, a measurement, and a counterparty with a reason to keep the supply healthy. That is why the book frames it as a fee rather than a benefit — not charity, not UBI, a line item for a contribution the system already depends on. The certainty is narrow and firm: a dividend cannot correct an underpricing, because a dividend is not a price.

🛰️✍️🔄🌡️🌫️🔒 F → G 🤝

🤝Significance — Stop Checking What It Meant, and the Combat Pay Becomes Unnecessary

why verifying meaning always loses · placement instead · what the boundary retires — significance

Everything above prices the arrangement honestly. The better move is to stop needing it, and that requires naming why the current setup is unwinnable rather than merely expensive. The human in the loop is being asked to verify what the output meant. That question is not decidable from the outputs of a Turing-complete system, so the checking never resolves, never accumulates, and repeats forever. And note the reason the arrangement fails, because it is not fatigue and not carelessness: it fails on speed and on volume. A person cannot ground six million operations a second, and cannot hold the shape of twenty thousand interacting decisions in mind at once. Every formal-verification and interpretability result that looks airtight has an invisible footnote reading assuming a human is watching — and autonomy is precisely the regime where that footnote stops being satisfiable. You cannot out-discipline a decidability problem, and you cannot out-work a throughput one. Every proposal to fix normalization with more careful review is a proposal to keep paying combat pay in perpetuity, and the errata in section B is what happens even when the reviewer is extremely motivated.

So change the question. Do not ask whether the output was right. Ask where it landed — and make that a coordinate the architecture can check without interpreting anything. Placement is decidable, and it is worth being exact about why rather than claiming an exemption nobody gets. Undecidability results bite programs whose behavior ranges over an infinite input space. Comparing two fixed, finite artifacts against a fixed lattice by a terminating walk is not that kind of question; it is finite and it halts. We are not beating the theorem. We are standing somewhere it does not reach, which is a much smaller and much more defensible claim. A system can be built so that an action either maps to a coordinate it was authorized to occupy or it does not, and the failure fires at the boundary rather than in someone's judgment three weeks later. The model's intent stops mattering, which is the point: intent was never inspectable from the outside, and pretending it was is what put a person in the seat of absorbing every error the system could not catch itself. This is why the instrument in section A refuses to score quality and signs position instead. It is not modesty about what we can measure. It is the only place a hard boundary can actually be built.

And the fence has to be stated as plainly as the claim, or the claim rots. The measurement tells you WHERE the work landed. It does not tell you WHETHER a paraphrase preserved the meaning. A synonym swap and a domain-breaking substitution can register as changes of nearly the same size; the correlation between semantic distance and physical distance is 0.767, not 1, and 1 is unreachable because the quantizer is lossy. So: no prevention claim, no kill-switch, no promise that a signed coordinate makes anything safe. The measurement measures. Whether the thing measured was worth doing stays a human judgment — now with a receipt attached, which is the entire improvement on offer and all of it we will claim.

Follow what that retires. If drift is arrested by the architecture, the human stops being the layer that absorbs it — and the transfer that was going to compensate you for absorbing it stops being necessary. Universal high income is the price of leaving the problem unsolved. It is what you pay a person to keep standing between an ungrounded system and the world, forever, at a station that grinds them down and that they cannot win. Solve the placement problem and there is nothing to compensate, because there is no longer a post to occupy.

Which reorders the whole argument into three arrangements, and the only thing separating them is where the checking physically happens. Today it happens in a person, unpaid and uncounted — that is the extractive version, and the invoice for it is genuinely owed. Universal high income moves the checking nowhere; it just pays the person to keep standing there, which is better than not paying them and still leaves someone standing there forever. The third moves the checking into the architecture: the action either lands on a coordinate it was authorized to occupy or it fails at the boundary. Pay the invoice now. Build the third one, because it is the only version where the seat stops existing.

🛰️✍️🔄🌡️🌫️🔒🤝 G → H 📚

📚Evidence — Ranked by Strength, Weakest Last

the principle first · the illustrations second · what we left out on purpose — evidence

Ordered deliberately, strongest first, because getting this order wrong is what produced two bad drafts.

The principle, which is load-bearing. Harnad 1990, Physica D 42(1–3), 335–346 — the origin, and the source of the phrase this post runs on: meaning must be made "intrinsic to the system, rather than just parasitic on the meanings in our heads." Bender and Koller, ACL 2020, for form versus meaning. Against us, and worth your time precisely because it is against us: Mollo and Millière, arXiv:2304.01481. The world-model evidence at full strength: Li et al., arXiv:2210.13382, and the replication at arXiv:2503.04421.

The observation, which is strong but informal. Supervision scaled with capability rather than against it. That is not a citation, it is the visible shape of a decade — prompt engineering as a discipline, chain-of-thought, eval harnesses, guardrail vendors, review layers. You do not need a paper for it; you need only to remember whether anyone was writing think step by step into production prompts eight years ago, and whether the apparatus around models is thinner now than it was.

The illustrations, which are weak and contestable, and which the argument does not rest on. Laban et al., arXiv:2505.06120, for the 16%-aptitude / 112%-unreliability decomposition. METR, arXiv:2507.09089, for the 19% slowdown among experienced developers on their own repositories. Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab, revised November 2025, for the 16% figure — and read their February 2026 update, where firm-time fixed effects move the significant decline to after 2024. Noy and Zhang, Science 2023, and Brynjolfsson, Li and Raymond, QJE, both showing real gains, both cited against ourselves. Vivalt et al., NBER 32719 and 32711. Lee et al., CHI 2025. Every item here is bounded and revisable, and none of it is holding the roof up.

Three numbers we left out on purpose. The MIT NANDA "95% of GenAI pilots fail" figure — 52 executive interviews, pulled from circulation, measuring pilots without measurable P&L return rather than anything failing. "Your Brain on ChatGPT" — a preprint with nine per group in the headline session, whose authors have asked reporters to stop saying "brain rot." Stockton SEED's mental-health result at p = .056, which did not clear the bar. All three would have made this sound stronger. None would have survived you looking, which is the only test that matters.

🛰️✍️🔄🌡️🌫️🔒🤝📚 H → I 🎯

🎯To-Do — Catch a Normalization, Then Price What Catches It

the test you can run today · the thing to count · the framing to refuse — to-do

Run the instrument. npx thetacog-mcp attest-open --no-open puts a deterministic placement receipt on your own machine, no model in the verdict. It demonstrates the split the argument rests on: where work landed is decidable and signable, whether it was good is not, and we refuse to price what we cannot decide.

Find one normalization in your own output this week. Take something a model produced for you that you approved. Not a hallucination — a piece that was fluent, correct in its parts, and quietly generic where it should have been specific: the recommendation that fits any company, the summary that lost the one contested detail, the paragraph that took your unusual point and rendered it as the usual one. That is the failure mode, it is far more common than fabrication, and it is invisible until you go looking for it deliberately.

Count the corrections, because uncounted contributions get priced at zero automatically. Every time you push output back toward the specific thing that was actually true, you are supplying the referent the system cannot generate. It is almost certainly in no cost model at your organization, including as a cost to your organization. Make it countable and it becomes an exposure base; an exposure base is what a rate attaches to; a rate is what turns a free input into a paid one. Vague claims that humans add judgment lose every budget conversation they enter.

Refuse the dividend framing, and be exact about why. Not because transfers are bad — Finland's results are real. Refuse it because a dividend is not a price, and only a price corrects an underpricing. The ask is not generosity once the jobs are gone. It is that a necessary, non-substitutable, continuously supplied input stop being carried at zero on everyone's books. The mechanism is in Chapter 9, and the economics are in The Grounding Tax.

🛰️✍️🔄🌡️🌫️🔒🤝📚🎯 I → tesseract.nu 🎯