The lab book stays open

Keep the failures.
They’re part of the work.

What we read, what we implemented, what we measured, and what didn’t earn a place in the running search.

External critique: what changes, what needs evidence

Coverage, verification cost, leaderboard units and the proposed Huisman bias test.

The full synthesis: what remains useful on a Mac, what the literature establishes, and where our confidence ends.

Booker–Sutherland, cubic-field norms, Elkies, elliptic curves and newer approaches. No claim that we have found the globally best one.

Our held-out experiment failed its acceptance gate. Cost learning remains useful; discovery enrichment remains unproved.

A complementary region in the cubic field. Preliminary exact checks, no deployed worker and no demonstrated performance gain.

The exact finite domain, unit separation, class ownership, endpoint conventions and arithmetic bounds for the native campaign.

Task definitions, receipts, exact verification, randomized selection and honest overlap accounting.

GitHub bank-ins, bounded independent replay, epoch updates and the distinction between a report and verified work.

Archived implementations, validation results, literature notes and an inventory of what is retained locally.

Experiment	Measured result	Decision
Native execution optimizations	≈1.58× earlier release on paired workloads	Used in the Mac campaign
Extra normalized residue sieve	About one-third fewer exact tests; no CPU gain	Not deployed
Discovery-directed learning	No demonstrated held-out advantage; failed gate	Not promoted
Complementary field geometry	Small exact domain audit passed	Research proposal only
Public distributed worker	Separate cross-language verification suite	Volunteer experiment; new evidence needed

Benchmark results belong to their stated workload and implementation. They are not odds of finding a solution.