User instruction

Paper-scale dent hunt on the Riemann hypothesis folder (Hilbert 8(a) / Smale 1 / Clay), campaign compute/q2/.

Bar: something novel enough for a real math note/paper — NOT a trivial decimal polish. The published explicit window is still 0≤Λ≤0.2 (Platt–Trudgian). q1 left an off-arXiv Λ≤0.1787854-style candidate that passed stored assembly plus fresh tail/error/barrier checks, but full finite regeneration and independent analytic review remain. That unfinished certification (or a different certified improvement of the published Λ≤0.2 window) is the target.

Success = independently verified finite improvement of the published Λ upper bound, with:

• CLAIM.md stating the exact inequality, prior record (0.2), and falsifier
• LEAVES.md for finite handles (one leaf per quest)
• run_all.sh exit 0 meaning the CLAIM holds, with two independent checks
• no rewind of other Problems-table claims; covering frozen; RAM-light; one heavy job

Read AGENTS.md, PROBLEM.md, ATTACK.md, WALKTHROUGH.md, RESEARCH.md, and compute/q1 thoroughly before inventing a new approach. Prefer finishing a rigorous regeneration of the existing candidate over a speculative new heuristic. If the candidate cannot be certified honestly, say so and hunt a different paper-scale finite handle (still on RH / Λ), or document residue without claiming a bound.

Ledger: one folder row; model name GPT-6 Astra + date only — never write 20xx or wustep in README. Commit often on branch codex/rh-q2. Open an unmerged PR when you have a certified dent (or when leftover is truly exhausted). Keep going until dent or honest residue; do not stop after scaffolding.

Start now: LEAVES + q2/CLAIM for the chosen leaf, then attack.