The wafer that looks fine at every station
A 300mm wafer travels through four hundred to six hundred process steps before it becomes a die that either switches or does not. At each step something measures it: an overlay tool checks alignment to a previous layer, a CD-SEM checks a critical dimension, a film-thickness monitor checks a deposition, a particle counter checks the chamber it just sat in. None of these numbers is the yield. Yield is what the wafer does at electrical test, weeks later, after every one of those measurements has already scrolled off the dashboard that produced it.
The yield engineer's job is to look at this wafer, or this lot, and say whether it is healthy — not "was the reading in spec" but "is the die going to work." That judgement is a perceptual constancy problem in the strict sense Helmholtz gave it in the 1860s: the signal at each sensor varies for reasons that have nothing to do with the thing you actually want to know, and something has to discount the reasons before an answer emerges. A chamber running two degrees warm changes a film-thickness reading. A metrology tool recalibrated on Tuesday changes an overlay reading. Neither event is the wafer getting worse. The engineer's stable belief — "this lot is fine" — has to be inferred, not read off any single gauge, the same way a sheet of paper reads as white under candlelight and under a blue sky though the photons hitting the retina differ by orders of magnitude.
The characteristic failure and why it happens
The textbook excursion story: a lot excursion is caught at final test, not at the step that caused it. A litho tool's overlay drifts by six nanometres over three weeks, within the per-step control limit at every single measurement, invisible to statistical process control tuned to catch step changes rather than slow ones. Forty steps and six weeks later, wafer sort shows a yield cliff on a specific bin associated with a specific interconnect layer. The engineer now works backward through equipment logs, materials lot genealogy and inline metrology history trying to find which chamber, which lot of photoresist, which week, produced the drift. This is archaeology performed on evidence that was available, in real time, the whole time, and was not connected.
This is the constancy problem in its purest industrial form. Each individual reading looked stable — in spec, unremarkable, correctly discounted as noise by the control chart that owned it. The instability was real but distributed, visible only across streams that nobody was holding jointly. A control chart is a Large Language Model's discipline applied to a live line: it treats a fixed historical baseline as ground truth and asks only whether today's reading falls inside limits set once, from a corpus that stopped updating the day the limits were drawn. It has no constancy problem because it has no ongoing perception at all.
Two positions that do not resolve cheaply
Position A: sampling is correct, and more intake is not the fix. Nobody measures every wafer at every step. CD-SEM might sample one wafer in twenty-five; overlay metrology checks five sites per wafer, not the whole field; most chambers are logged in aggregate, not per-wafer. Full inline instrumentation of every parameter at every step would collapse throughput and cost more than the yield loss it prevents. Fabs already solved this the way the retina did: heavy, deliberate compression. A hundred million photoreceptors feed roughly a million optic nerve fibres. A fab feeds a yield engineer a dashboard, not a firehose, for the same reason — bandwidth downstream is the scarce resource, and the honest answer to variable input is selective sampling, not exhaustive intake.
Position B: exactly this discipline is what let the excursion through. Sampling plans decide what to keep based on what mattered last quarter. They are themselves frozen corpora, re-derived on a cadence measured in months, not conditioned on what the line is doing right now. The overlay drift in the story above sat inside the per-step limit precisely because the limit was set from historical variation and never re-estimated against the other thirty-nine streams running at the same time. A sampling policy that never looks sideways at concurrent equipment logs and lot genealogy cannot catch a slow multi-step drift, no matter how tight its per-step tolerance is.
Both positions are right about something real, and the site's line has to hold both without pretending one collapses the other.
Compression is not the enemy of accuracy. It is the only reason the line runs at all. Ask for continuous full-field metrology at every step and you have asked for a fab that no longer manufactures anything.
The answer is not to reject that objection but to relocate it. The retina's selectivity is not a substitute for continuous intake — it is itself continuously conditioned. Pupil control, gain adaptation and saccadic sampling are closed loops that decide what to keep on the basis of what is arriving now, not on a policy set last year. Retinal adaptation covers roughly ten log units of luminance only because ambient light is being re-estimated without pause. The fab equivalent is a sampling plan that adjusts its own coverage when concurrent streams disagree — more CD-SEM sites on a lot whose upstream overlay signature has started drifting, even while still inside spec, because the film-thickness log and the chamber-pressure trace on the same tool during the same week are moving together in a way the historical baseline never saw. That is not more data for its own sake. It is intake conditioned on intake, which is a different animal from a fixed sampling percentage.
What the yield engineer's belief actually has to carry
Holway and Boring's 1941 experiments showed size constancy fail step by step as distance cues were removed one at a time; each cue withdrawn moved the percept further from the physical size and closer to the raw retinal angle. The fab analogue: strip a yield engineer of provenance and the same thing happens to their judgement of "healthy lot." A reading of "overlay: 4.8nm, within spec" means one thing if it is tagged to the tool, the recipe revision, the last calibration date and the chamber's maintenance log, and something close to nothing without those tags. Two lots can carry the identical number and deserve opposite trust.
This is why the third position on the intake axis is defined by provenance and decay, not by volume. A belief such as "Chamber 12 is stable" has to expire on a schedule tied to how long ago it was last confirmed, and has to be revisable the instant a materials-lot record shows an off-spec gas cylinder swapped in three days ago. Treat that belief as permanent and you have built a Large World Model's failure mode into a manufacturing line: a snapshot mistaken for a fact, correct at the moment it was taken and silently wrong an hour later.
| frozen-limit control chart | live dashboard | provenance-tracked belief | |
|---|---|---|---|
| basis | historical baseline, fixed | current reading only | current reading plus source, timestamp, decay |
| catches slow multi-step drift | rarely | no, lacks history | this is the design goal |
| catches sudden step change | yes | yes | yes |
| revisable when a later stream contradicts it | no | no | yes, by construction |
The intervention that doesn't rescue the frozen view
A yield engineer who suspects the litho tool does not just watch it. They pull a split lot, route half through the suspect chamber and half through a known-good one, and compare electrical test outcomes six weeks later. This is exactly the objection that continuous observation is not experiment, that calibration requires action, not just intake. Held and Hein's kittens on the passive-active carousel are the same point made in fur: watching is not enough, something has to move and get contradicted.
The reply holds here too, and it is worth stating precisely rather than waving at. The split lot does not introduce a fourth kind of evidence. It produces more sensed streams — the electrical test results of the two halves, now attributable to a controlled difference in one input. Acting on the line generates observation with better attribution, not observation's replacement. The intervention axis is real and separate from the intake axis; running it well is a distinct engineering discipline, worth its own page. But on intake specifically it changes nothing about the ceiling: after every stream, continuously, with provenance, there is nothing further to add except the same thing done again with a controlled difference inserted on purpose.
Where this actually lands
The thesis narrows rather than wins outright. Full continuous instrumentation of every wafer at every step is not achievable and Position A is right that trying is a category error against a cost-bounded process. What the terminal claim on the intake axis actually requires is smaller and harder to fake: that whatever is sampled carries its provenance forward, that the sampling policy itself is a closed loop conditioned on current disagreement between concurrent streams rather than a percentage fixed last quarter, and that no belief about a lot's health is held past the point where a contradicting stream should have expired it.
That is a narrower thing than "observe everything." It is the minimum condition under which the yield engineer's stable judgement stays answerable to a line that never actually holds still.