2026-09-02
Iain Pears wrote a novel called An Instance of the Fingerpost that tells a single story four separate times, through four narrators, none of whom can be fully trusted. Each one has their own blind spots, their own self-interest, their own honest confusion about what actually happened. No single account settles the matter. The title comes from Francis Bacon, who used the Latin phrase instantia crucis — the crucial instance — for the one piece of evidence that finally points a single direction when every other signpost has been ambiguous. Bacon’s idea wasn’t that any one witness becomes trustworthy. It’s that the right collection of witnesses, compared against each other, can settle what no one of them could settle alone.
Our reasoning machine is built around exactly that idea, and it’s where an earlier fact about the system — that correct readings of a problem all agree with each other, while wrong readings scatter each in their own particular way — finally gets put to work. If agreement is evidence of truth, then the way to get honest answers out of an imperfect reader is not to trust the reader. It’s to build several imperfect readers whose mistakes don’t overlap, and only speak when they all land in the same place.
The wall has layers, each one a different kind of witness.
First, the same problem gets read five times, with the sentences permuted — reordered, reshuffled — so that nothing about the underlying arithmetic changes but the surface presentation does. If the machine actually understood the problem, permuting harmless detail shouldn’t move the answer; if it was pattern-matching on sentence order or position, the readings will disagree with each other. All five readings have to land on the same diagram before this stage passes.
Second, the problem is checked against models trained separately, from different lineages and different widths — different starting points, different training runs, not just different random permutations of the same trained head. Two models that came from genuinely different training histories are unlikely to share the same blind spot, so if they land on the same graph anyway, that agreement means something the first stage’s agreement alone couldn’t guarantee. Many landscapes, one shape.
Third, behind both of those, an out-of-distribution check — we call it the mouth — asks a different question entirely: not “did the readers agree,” but “does this problem even live in territory the system has been trained to recognize at all.” A problem can pass the first two walls by accident if it happens to sit somewhere strange enough that every reader makes the same unusual mistake; the mouth exists to catch exactly that failure mode, by checking familiarity rather than agreement.
Independent witnesses, independent failure modes. Wrong readings scatter across all three checks in their own particular ways; only a correct reading survives being interrogated from three unrelated directions at once. Only unanimity crosses the wall — a fingerpost built, the way Bacon meant it, out of testimony rather than authority.
And when the wall doesn’t pass — when the machine can’t certify an answer — it doesn’t just fail silently. The symbolic solver at the core of the system, the one doing the exact logical search, has a particular kind of honesty built into its refusals. When a diagram turns out to be contradictory — when the constraints it’s been handed genuinely cannot all be true at once — the solver doesn’t just report failure. It can hand back a minimal unsatisfiable core: the smallest subset of those constraints that, on their own, already cannot be satisfied together. Not “no.” Not even “no, because of these fifteen things.” Just the smallest possible “no, and here is exactly the contradiction” — the precise handful of facts that don’t fit, stripped of everything else that was irrelevant to the failure.
That is the deeper point the fingerpost was always making. A single unreliable witness can’t be trusted on their own account, but a collection of them, compared honestly, can point one direction with real confidence — and even when the story falls apart entirely, falling apart is not the same as going silent. Even failure, read correctly, has a shape.