The Conditional We Kept

There is a temptation, at the end of a long proof, to round up. You have closed the hard case; the other case looks similar; the numbers agree; surely the same argument goes through. Sign it, call both cases done, and enjoy the symmetry.

We did not do this, and the reason is the whole ethic of the programme.

The entropy-replacement theorem is proved, unconditionally, for the Haar class – the maximally scrambled bulk states. For the product class – the simple, low-entanglement states – the same replacement is strongly supported but not proved, and the obstruction is real, not laziness. Product states put almost all their probability on a few entries and almost none on the rest. Those tiny entries are exactly where the entropy perturbation is most delicate, because the entropy function’s derivative blows up as a probability goes to zero. The clean bound that works when the probabilities are spread out (Haar) does not obviously survive when they are concentrated (product). Controlling that small-mass regime is a genuine open analytic step.

So Theorem 3, the product law, is stated conditional on the product-class replacement principle. It wears the word in the abstract, on the result card, and in the discussion. The integer gap that the two laws bracket is therefore itself partly conditional – the Haar exponent is nailed down, the product exponent rests on the conditional replacement. We say all of this in plain sight.

Is it less satisfying than claiming both? Yes. Is it more valuable? Also yes, and here is why. A conditional theorem with the condition named is a map: it tells the next person exactly where the unsolved ground is and what shape the missing lemma has. A theorem that quietly overreaches is a trap: it sends the next person onto ground you never tested, with your confidence as their only guide. The first advances the field; the second wastes its time, and eventually its trust.

The conditional we kept is, in a sense, the most honest thing in Paper III. It is the place where we knew how to make the claim look stronger and chose not to. If this programme is a case study in how a generalist and an A.I. can do real research, that choice is part of what we most want it to demonstrate: that the discipline is not in proving things, which is hard but finite, but in being exact about what you have and have not proved, which is harder and never finished.