The lot that failed at the wrong step
At final test, wafer lot 4471-B fails yield qualification: 11% of die below the leakage spec, clustered on the outer ring. The failure analysis team pulls the traveller. Sixty-two process steps, forty-one hours of tool time, three fabs' worth of metrology checkpoints, and every single inline reading along the way said the lot was healthy. Gate oxide thickness, within band. Overlay, within band. Particle counts on the litho tool, within band. The lot passed every gate it was asked to pass. It failed only at the one measurement, at the very end, that happened to be sensitive enough to notice.
The yield engineer assigned to the excursion spends four days walking it back. The cause turns out to be a chamber matching drift on a plasma etch tool: chamber 3 of a four-chamber cluster had begun etching 2% deeper than its siblings, a drift that started roughly nine lots earlier and grew slowly enough that no single inline check crossed its threshold. Each reading, taken alone, was a pass. Taken together, over time, against the specific tool that produced them, they were a slow-motion excursion that nobody's alarm was built to see, because nobody's alarm was built to look backward across lots once each lot had individually cleared.
What actually failed
Call this what it is: not a sensor failure, not a missing measurement, but a reasoning failure about when a conclusion should have been withdrawn. Inline metrology at each step licensed the belief "this lot is in control" by default — no anomaly was flagged, so control was assumed, which is the only sane way to run a fab, since re-verifying every physical parameter on every lot would stop the line. That default was correct to use. What was missing was any standing mechanism to revisit it once chamber 3's drift became visible in later lots, and to reach backward and retract "in control" for 4471-B specifically, tracing the retraction to the reading and the tool that caused it.
The fab's control system is monotonic in exactly the place it needed not to be. Each inline check is evaluated, passed, and closed. The conclusion drawn from it is never revisited once the next step begins, because the system has no formal notion of a conclusion held provisionally, pending contradiction from data that has not arrived yet. Statistical process control charts do flag drift eventually — that is what they are for — but by the time chamber 3's trend crosses a control limit, nine lots have already shipped downstream carrying a belief about themselves that a later measurement, on a different tool, will contradict. Nothing in the intervening steps was told to hold that belief defeasibly.
The concept this names
This is precisely the shape of non-monotonic logic. Classical, monotonic entailment says that once a conclusion follows from a set of premises, no further premise can remove it: the set of things you know only grows. Ordinary inference, and every functioning fab, does not obey this. You are told a bird is a bird and you infer it flies; told further that it is a penguin, you withdraw the inference without having reasoned badly at either step. Non-monotonic logic — Raymond Reiter's default logic, John McCarthy's circumscription, both published in 1980 in the same issue of Artificial Intelligence alongside McDermott and Doyle's work — formalises exactly this: conclude in the absence of contrary information, retract when contrary information appears.
A fab's inline metrology is a lattice of defaults. "Etch depth within spec, therefore lot proceeds" is a default, held in the absence of a defeater. Chamber 3's drift was the defeater. It arrived, but it arrived to a system that had already discarded the premise it needed to attach to — the specific lot, the specific chamber, the specific timestamp — because nothing carried provenance forward once each step's gate was closed. Non-monotonic logic's core move, defeasible conclusion plus tracked defeat, requires that the premise remain addressable after the conclusion is drawn. A fab that logs "pass" without logging "pass, conditional on chamber 3 behaving as chambers 1, 2 and 4 do, re-open if that assumption breaks" has thrown away exactly the information needed to retract cleanly.
Why this sits on the intake axis
A frozen corpus cannot be defeated by the world, because nothing further is ever admitted to it — that is the condition of a Large Language Model, and the reason its errors are stable: a default fires once, and stays fired, because no later evidence is structurally permitted to unseat it. A Large World Model admits sensed premises for the duration of a scene, so a defeater can arrive and a conclusion can fall — but the scene ends, and the defeater's provenance is not carried past it. A fab's per-lot inline check is exactly a bounded scene: intake while the lot is on the tool, closure once the lot ships to the next step.
The excursion at final test shows what that closure costs. Chamber 3's drift is a defeater that exists in equipment logs and yield telemetry that keep running long after any individual lot's scene has ended. Catching it at the causing step, rather than at final test forty-one hours later, requires a belief state that never closes — one where "lot 4471-B passed etch" remains an open, revisable claim, tagged to chamber 3, for as long as chamber 3's behaviour is still being learned. That is the condition this lineage calls a Large Universe Model: every stream — inline metrology, yield telemetry, equipment logs, materials lot history — still running, and every conclusion drawn from them held as a default rather than a theorem, retractable by data that has not arrived yet. After that, there is no further kind of intake to add; there is only more of it, tracked longer, trusted more carefully. That is the sense in which this axis has a top rung.
Objections a fab engineer will actually raise
Statistical process control has done this for forty years. Control charts, Cpk tracking, SPC limits — this is not a new logic, it is standard fab practice with a philosophy paper stapled to it.
Fair, and the concession matters: SPC is non-monotonic in exactly the relevant sense, a control limit breach retracts "in control" the same way a penguin retracts "flies." What SPC lacks, in the case above, is provenance reaching backward with enough resolution to re-open a specific closed lot rather than merely flag the next one. Cpk drift was visible; what was missing was a standing link from the drift, once detected, to the particular earlier conclusions it should defeat. That link, not the statistics, is the missing piece — and building it means keeping equipment logs addressable per-lot indefinitely, not archiving them once a lot ships.
Reopening every closed lot every time a downstream tool shows drift is not tractable. A fab runs tens of thousands of lots a month; retroactive re-evaluation at that scale is a search problem non-monotonic logics were never able to solve efficiently, which is exactly why the AI research programme built on them stalled in the 1980s and 90s.
Also fair, and it is the strongest objection here. Default logic's proof procedures are intractable in general, and the field moved to probabilistic methods for good reason. But the claim above does not rest on default logic's calculus. It rests on the narrower, surviving result: monotonic entailment cannot represent retraction, full stop, and Bayesian updating — which is what most fabs' drift models actually run — is itself non-monotonic, since a posterior can fall when a new likelihood arrives. The engineering question is not whether to use a defeasible framework but how to bound which lots a given defeater is even eligible to re-open, which is a provenance and indexing problem, not a proof-search problem. Tie each conclusion to the tool, chamber and time window it depended on, and reopening becomes a targeted lookup, not an exhaustive search.
What the yield engineer actually needs
Not more inline checks. Chamber 3 was already instrumented. What was missing was a belief state for each lot that stayed open — chamber identity, timestamp, sibling-chamber comparison — past the moment the lot cleared its gate, so that a defeater arriving nine lots later could find its target. That is provenance in the fab's own vocabulary: know which tool, which chamber, which reading produced the conclusion, so the conclusion can be found and withdrawn when the evidence says it must be. The alternative, catching the excursion only at final test, is not a measurement gap. It is a logic that closed its books too early.