Why this object

The amplituhedron (Arkani-Hamed and Trnka, 2013) is a shape in positive geometry whose volume gives the scattering amplitudes of a quantum field theory — the numbers that say how particles interact. The radical part is what the construction doesn’t use: no spacetime, no locality, no unitarity as inputs. Those pillars of physics come out of the geometry instead of going in. A calculation that takes pages of Feynman diagrams becomes the volume of one object.

That is the kind of idea I cannot leave alone, and it is the same itch as my CTI work on criticality: the suspicion that the deep vocabulary of a field — space and time here, intelligence and life there — might be emergent bookkeeping for a simpler structure underneath. The amplituhedron connects outward in every direction I care about: combinatorics (triangulations, the positive Grassmannian), computation (how you actually enumerate and verify these volumes), and by analogy the questions of complexity — how much structure a system must have before the familiar categories appear. Where the idea is valid, it reorganises physics. Where a given approach to it fails, that failure is itself a finding. Both outcomes are worth having; the interest survives either.

What I did with it

In 2025 I pursued a learned-search programme over amplituhedron triangulations, with funding drafts (NSF EAGER, SBIR — prepared, never submitted) describing the approach. In 2026 I audited that programme the hard way, and the audit is the part I stand behind most:

  • The programme’s verifier tested simplex disjointness by sampling barycentres only. An exact rational-arithmetic census shows that test misses 673 of 928 truly overlapping pairs (72.5%) at C(8,4) — with explicit counterexamples constructed, and the result independently replicated by a second method (float LP).
  • The headline “99.1% of the volume” was a sum over overlapping simplices; true Monte Carlo coverage is 0.697 ± 0.004.
  • The codebase named in those drafts has not been found on any surface searched. I am still looking; for now that chapter is closed as lost — and the drafts’ stronger descriptions of it do not stand, as I have said on the record.
  • I built verifier v2 (exact disjointness plus ridge-pairing checks; all 40 audited closures pass, the old method fails), and a baseline study on the sound test closed the k=1 arm as a clean null result: greedy fills exactly, and a theorem (Rambau 1997, via Rambau–Reiner) explains why simplex counts at this level are fixed at binom(n−3, 2) regardless of method — and the m=2 tile count is likewise fixed, at binom(n−2, k), by the Magic Number Theorem (Parisi et al., 2024), so the audit’s open question becomes whether the verifier’s closures reproduce that count.
  • A methodological control, demonstrated synthetically and now mandatory in my work: log-sampling without the Jacobian silently measures ∫f dx/x — a 3.65× error on the test integrand.

Code

None in this entry — honestly. The 2025 codebase is lost (above), and inventing a plausible excerpt would repeat the exact failure this project documents. Verifier v2 exists in the review workpapers and can be excerpted at republication time.

Lessons

  • The verifier is part of the result. Every number a pipeline produces inherits the soundness of the cheapest test inside it.
  • Exact beats sampled when the claim is exact. A rational-arithmetic census ended an argument that estimates had sustained for a year.
  • Theorems are cheaper than compute. One 1997 corollary retired an entire experimental arm.
  • An interest is not a claim. The mathematics is worth pursuing however the pursuit ends — including at “false”. What is not permitted is letting enthusiasm write numbers the arithmetic hasn’t earned.

Status: Audit complete (2026-11-11); m=2 verifier work continues against a theorem-fixed tile count · Positive geometry, exact rational arithmetic, Monte Carlo, Python