Audit · launched 05:27 · returned 07:16 · 10 steps

Audit · Fresh review, re-checks

After the run, eight fresh reviewers each took about ten steps of one agent, twice the usual depth, and graded every step correct, minor or error before reading the run's own audit of it. They redid the computations by their own methods: the 45 Borel orbits and the 270 rank-one pairs, confirmed in every odd characteristic by the point count \(\sum_x|Y_x(\mathbb F_q)|=|E_6(q)|/|F_4(q)|\); the lattice data; the Weyl characters; and the \(E_6\) Lie algebra with the actual involutions. They also read Sakellaridis, his 2008 paper and Delorme–Sécherre against every citation.

Of 76 steps, 52 were correct and 21 had minor defects that change no claim. The proof of the main result held: (A3a)(A3a) Rank-one models over O (Sakellaridis's Axiom 2.2.1): for every Borel orbit Y of maximal rank and every simple root α, the O-scheme YP_α/R(P_α) is one of the four standard PGL2-spaces over O, with his scheme structure on Y, and YP_α is the closure of its generic fibre minus the closure of the generic boundary. for \(E_6/F_4\) (Review G2, Review G7), the close-field transfer (Review G4, Review G5) and the check of [Sa08] (Review G8). The three errors sat in one place, the hypothesis list of the direct route for a general symmetric variety. The general theorem used non-zero leading coefficients without proof, and the in-run audit had missed this (Review G1, Review G6). Reviewer G8 also listed eleven small mismatches between the published texts and the steps.

One salvage agent filled the gap by positivity. The two reviewers who had found the errors re-checked its steps, and every repair holds (Re-check V1, Re-check V2).

All 10 steps, in order
Result as the agent reported it proved

76 steps graded: 52 correct, 21 minor, 3 errors (Theorem 1.9, Objection 5.20, Summary 5.21), all in the hypothesis list of the direct route for a general symmetric variety. One salvage agent repaired them; the reviewers who found them re-checked the repairs, and every repair holds. No conclusion about \(E_6/F_4\) changes. The final verdict is Summary 7.5.

Brief as sent

Eight fresh reviewers, twice the usual number, each with about ten steps of one agent track, the parents and children outside it, and the run's setting, checklist and sources. Each re-derived every step, redid every computation by its own method, checked every cited theorem's hypotheses in its source, graded correct, minor or error before reading the run's own checks, and traced every error. Reviewer G8 also checked the published verdict and the Proposition VI pages. The reviewers who found the errors (G1, G6) then re-checked the salvage. The briefs are in audit/briefs.json.