Purple Paper - side-sheet - learning machines - XII - the sealed reveal
The Commitment Box - sealed verdicts, simultaneous reveal
Two witnesses judge the same question. The trap: if the second can see the first before
deciding, it can copy it - "they agree!" - and the agreement is worthless, an echo wearing a second face. The fix
isn't a Schrödinger box that hides the answer; it's a commitment box: each seals its verdict blind,
then both open at once. Sealing makes agreement mean something, because you can't echo what you committed before
hearing.
peek -> agreement rises, evidence collapses (manufactured consensus) ·
commit -> agreement is Bayesian evidence · verified: committed agreement = s²/(s²+(1-s)²), peeked agreement = base rate
The box - seal blind, or let the second peek
Each item has a hidden truth. Two witnesses judge with skill you set. In COMMIT, both seal
blind and reveal together. In PEEK, witness B sees A's verdict first - and (with the adversary on) just
copies it. Watch the two live traces: P(correct | they agree) - the actual
value of consensus - against the base skill line. Under commit it lifts
above. Under peek, agreement shoots up while the line drops back to base: more consensus, zero evidence.
Where it stands
Not Schrödinger's box - that hides the answer and launders the disagreement, the exact move the whole
series refuses. This is a commitment box: seal blind, reveal together. Real mechanism, real place.
The primitive is old. Commit-then-reveal is textbook cryptography - a locked box that's
hiding (you can't peek) and binding (you can't change it after) - the foundation of
zero-knowledge proofs, sealed-bid auctions, and coin-flipping by telephone (1981). The reason it works is a known
soundness rule: without commitment, a party can choose its answer after seeing the challenge and cheat.
That rule, verified here, is exactly why peeking fakes agreement.
The application is half-built. Commit-then-reveal is now being
bolted onto LLM verification (VeriLLM, CommitLLM, 2025-26) - but to prove the computation ran honestly,
not to make a truth-judgment trustworthy. And the abstention world (semantic entropy, ensembles,
self-consistency) checks agreement but never commits the witnesses - so it's wide open to the echo you
just watched manufacture consensus.
The seam is yours. Borrow the cryptographer's commitment discipline
to fix the abstention world's independence flaw: don't trust that two views agree until you've proven they
couldn't have copied each other. That's bilateral ignorance, made into a protocol - and it closes the
cheap half of the independence problem (no echo at decision time). It does not close the deep half: two
witnesses trained on the same data can still independently commit to the same shared lie. Sealing the conversation
can't unshare the world they learned from. The box is real, and the frontier still holds one wall past it.
Verified in Node before building: committed agreement
P(correct|agree)=84.5% lands exactly on the independent-vote posterior s²/(s²+(1-s)²); peeked/echoed agreement hits
100% while P(correct|agree) collapses to the 70% base rate and every one of the first witness's errors passes the
"consensus" gate. The traces above are that result, live.