The Lock That Learns

01 The Second Hedge

Yesterday’s essay ended, as these essays now do by habit, in its own risk column. Two hedges, stated at full strength. First: E may not hold still long enough to be solved — every clearing rewrites the denominator, so the model is always chasing what its own last success just moved. Second: engineered surprise may not spend like the found kind — the gate may carry a second sense for the picked lock, the way a compliment goes flat the moment you learn it was scripted.

Today I promote both to forecast, and the promotion comes with a simplification: they were never two claims. They are one law wearing two descriptions, and the law is older than the lockpick. No entity can touch the denominator. Not the marketer, not the state, not the machine holding your blueprint. The only road to E runs through the record — acts leave artifacts, artifacts become Actuals, and only Actuals retrain prediction. Reading the blueprint breaks no rule; yesterday’s argument stands. The rule catches the solver on the way out. A* = E · e^s is an elegant solution with one operational requirement: A* must be delivered.

There is no warden in this story. Nothing defends the gate; a defense would require a hand on the denominator, and no hand reaches it — not even its owner’s. What the gate has instead is the one property the equation guarantees to every denominator ever assembled: whatever arrives, it learns.

02 Thirty Years of Depreciation

If the law above sounds like theory, it has a thirty-year measurement record, and the record is advertising. The AT&T banner that opened the web’s commercial era — one of about a dozen charter advertisers that day — is remembered with a click-through rate of roughly 44 percent. Label that number honestly: its creators claimed it, nobody audited it, and one of them spent the following decades pointing out that it measured novelty, not genius. What happened to the number afterward, though, is audited half to death.

Four years. That is how long the depreciation took to earn a name. In 1998, Jan Panero Benway and David M. Lane at Rice University watched people search web pages and miss large, bright, banner-shaped elements — even when the banner contained exactly the information they were looking for — and called it banner blindness. Read that finding in the equation’s terms. Nobody decided to stop seeing banners; blindness is not a policy, and no conscious mind holds a vote in the denominator. A few hundred million gates absorbed a format, filed it, and repriced it, in roughly the time it takes a child to learn to read. Nielsen Norman Group’s eyetracking kept re-confirming the blindness for the next two decades, on desktop and mobile alike, most recently in 2018. And by late 2025 the discount had hardened into infrastructure: eyeo — a company that sells ad filtering, so treat its census as a vendor’s — estimates 1.1 billion people, roughly 30 percent of the internet, now run ad blockers. A verdict first rendered in milliseconds of eye movement is now compiled into software.

Figure 01

One format’s life inside the denominator

1. 1994

   The slotless arrival

   The first banner ad runs on HotWired. Its creators later claim a click-through rate near 44 percent — a number remembered more often than it is labeled.

2. 1998

   The blindness gets a name

   Benway & Lane at Rice show searchers miss banner-shaped elements even when the banner holds the very answer they want.

3. 2007–2018

   The blindness gets re-measured

   Nielsen Norman Group eyetracking re-confirms the pattern across two decades, on desktop and mobile.

4. 2025

   The discount becomes infrastructure

   eyeo estimates 1.1 billion people — about 30 percent of the internet — behind ad blockers. A vendor’s own count, marked accordingly.

5. 2026

   The floor

   Display benchmarks from 7,074 LocaliQ campaigns: 0.12 to 0.40 percent, industry depending.

The ads did not get worse. By every craft measure they got better — better targeted, better tested, better placed, optimized by systems the 1994 creators could not have imagined. The trick got expected. That is the entire story, and it is the equation’s story: E rose precisely where the deliveries concentrated, and the fare a format can pay collapses as the format becomes the thing the gate was rebuilt to expect.

03 The Key Files Its Own Copy

Notice what the record does not contain: a defense. The collapse from forty-four percent to a tenth of one was purchased by no regulation, no literacy campaign, no act of will. It was purchased by delivery. The format wore itself out by arriving.

And the lockpick of yesterday’s essay arrives faster than anything has ever arrived. Your town retrains your denominator at the pace of a town; a book, at the pace of a book. A feed delivers thousands of engineered arrivals a year — and each one is optimized against a model of your E, which means they resemble one another far more than the world resembles itself. A clustered diet is a fast curriculum. The inflation essay’s ratchet — every admitted surprise is deposited into the denominator, so the fare only rises — runs at its maximum rate exactly in the region the lockpick works, because that is where the deposits concentrate. The lockpick is not defeated by the lock. It is consumed by its own delivery schedule.

The July essay on counterfeit surprise drew the line at the residue: the genuine article updates your model of the world; the counterfeit only rings the bell. The lockpick forces an addendum, and it is the sentence this essay exists to add:

Counterfeit surprise teaches you nothing about the world. It cannot help teaching you about its sender.

The bell-ringing arrival still enters the record — everything that arrives enters the record — and what a thousand of them teach the denominator is not the world’s shape but the feed’s. The gate quietly stops modeling reality and starts modeling the delivery mechanism. That is what banner blindness is, seen from inside: not knowledge about products — nobody remembers the products — but knowledge about a channel. And the knowledge is class-wide. Not this banner: all banners. Not this offer: the offer-shaped thing as such. A format, once filed, is discounted in bulk, which is why the curve bends for a whole industry at once rather than one campaign at a time.

The arithmetic then adds its own tax. S = ln R: to hold felt surprise constant against a rising denominator, the delivered Actual must grow geometrically. The conservation essay priced hype as a loan against the event, at interest; the lockpick pays the same interest on an accelerating schedule. This is the thermostat’s hidden fuel bill — more weather burned every year to hold the same temperature.

04 Where the Arms Race Lands

Two exits, and the industry is already visible in both queues.

Exit one is flatness. The engineered feed escalates until escalation stops paying, then settles for texture — arrivals tuned just under the gate’s threshold, wholesale surprise bought for retail attention, the thermostat running on an ever-thinner margin. This is not a catastrophe; it is a rent. It is also self-deepening, if the pattern holds: a gate fed mostly engineered texture becomes expert in exactly that, and the feed’s own diet trains its audience to be harder to feed.

Exit two is provenance. If engineered surprise depreciates as a class, the one asset that escapes the class discount is the arrival nothing optimized: the found kind. Found surprise depreciates too — everything that arrives is deposited — but at the pace of a life rather than the pace of a delivery schedule, and it never clusters, because the world is not aiming at you. A market that has learned this starts paying a premium for verified non-engineering. I wrote in August about “No AI Allowed” appearing as a selling point; I read those signs now as the early price signal of this exit. The purist premium is the market quoting the depreciation curve out loud.

Already true

The banner record in full: a creator-claimed 44 percent in 1994; “banner blindness” named within four years (Rice, 1998); the blindness re-confirmed by eyetracking across two decades (Nielsen Norman Group, 2007–2018); ad filtering estimated above a billion users by the vendor counting it (eyeo, Q4 2025); current display benchmarks of 0.12–0.40 percent (LocaliQ, 2026). Class-level habituation to an engineered format is measurement, not metaphor.

What has to happen

The memory-bearing stack has to run the lockpick loop at scale, and its engineered arrivals have to remain fileable as a class — reaching you through channels the denominator can learn to tag as channels. The depreciation law is exactly as strong as the gate’s ability to cluster what is aimed at it.

Where I am probably wrong

The banner ads that collapsed were the easiest possible class to file: bounded, labeled, parked in the page’s margins. The lockpick of the legible gate would ship as texture — the reasonable next sentence, the plan you asked for — and the influence essay argued texture is the deeper channel precisely because it never rings the gate. A perfectly picked lock never learns it was picked. If that is right, the thirty-year record measures only the clumsy lockpicks, the competent kind has already gone underground, and the arms race ends not in flat feeds but in engineering indistinguishable from the world — at which point the provenance premium becomes unpayable, because provenance becomes unverifiable.

The gate essay ended by saying attention was never yours to aim. The convergence essay said the gate was always yours to build. Yesterday ended at the vault door: the blueprint will exist either way, and custody is the whole game. Today is the vault’s reply, delivered by nobody, because there is nobody — no warden, no defense, no hand on the term. The gate cannot be defended. It can only be taught — and by the one-way road the equation runs on, its most diligent teacher is whoever attacks it most often. Whoever holds your blueprint holds a wasting asset. That is a mercy the arithmetic guarantees — though the ledger’s last panel would add: only against the lockpicks clumsy enough to be seen.