# Calibration round 2

Date: 2026-11-20. Round-1 files were not overwritten. No ledger writes were made.

## BSD: conductor at most 30

File: `data/bsd_curves_r2.json` (sha256 `bf843735dce1f33bafcde1a6146569972499512f8ceb2b63b382c2bd1ba70175`).

- Curves: 57 total. Round 1 had 28 (conductors 1 to 20); round 2 adds 29 curves at conductors 21 to 30 (21a1-21a6, 24a1-24a6, 26a1-26a3, 26b1-26b2, 27a1-27a4, 30a1-30a8).
- Rank split: 57 rank 0, 0 rank nonzero. No pattern break.
- All L(E,1) nonzero (smallest |L| 0.2538418608559107). No pattern break.
- Method: same as round 1, PARI/GP 2.19.1 `ellanalyticrank` at realprecision 30, conductor recomputed with `ellglobalred`, script `compute/bsd_curves_r2.gp` and `compute/bsd_curves_r2.py`.
- Runtime: PARI scan 0.5 s.

## Riemann: first 2,000 zeros

File: `data/riemann_zeros_r2.json` (sha256 `5f4b6f3d360e9aa2d873fb44034b1686ee14f7906270770989b559537cccbb9b`).

- Zeros checked: 2,000. First t 14.13472514173469379045725, last t (zero 2,000) 2515.28648292471288003819.
- Max real-part deviation after one mpmath Newton step (dps 30, free to move off the line): 0.0. No zero off the line.
- Max |zeta(0.5 + i t)| at the line: 6.5408e-26.
- Count check: von Mangoldt N(T) at T = 2515.28648 gives 1999.38 expected vs 2,000 found, S(T) = 0.62, in line with round 1 (S(T) = 0.58 at T = 1419.42). No missed zero indicated.
- The first 1,000 t values match round 1 exactly.

## Runtimes and what did not go as planned

The round-1 estimate (about 5 minutes total for the Riemann doubling) did not hold. The machine was saturated by sibling round-2 jobs (run_r2.py, navier_stokes_hunt_r2.py, p_vs_np_r2.py, Lean mathlib build; load average up to 8.38 on 1.5 shared CPUs), so wall times below are mostly contention, not compute.

- Riemann PARI scan: 907.3 s wall for the upper interval [1420, 2560] (1,042 zeros found), plus the reused round-1 scan of 44.46 s for [0, 1420]. Rough CPU cost of the upper scan was about 4 to 5 minutes.
- Riemann mpmath verify: 428.91 s wall (about 7 minutes CPU) for 2,000 zeros, run at nice 19.
- Total wall for the Riemann extension: about 1,336 s.

Two scan attempts failed before the split scan worked:

1. Full [0, 2560] scan with the default 8 MB PARI stack: `lfunzeros` stack overflow. Round 1 at [0, 1420] fit the default stack; T = 2560 does not.
2. Full [0, 2560] scan with parisize 256 MB: killed after 377 s wall with no output, while sibling jobs saturated the machine.
3. The working approach split the interval: round-1 t values reused for [0, 1420], and a fresh scan on [1420, 2560] with parisize 64 MB (`compute/riemann_scan_r2_upper.gp`, output `data/riemann_t_r2_upper.txt`). The verify script (`compute/riemann_verify_r2.py`) asserts continuity (first upper t 1420.4165 above the round-1 last t 1419.4225) and takes the first 2,000 of the combined 2,042 zeros.

Lesson for round 3: run the PARI scan when the machine is quiet, or keep the split-interval approach; do not raise parisize to 256 MB on a contended box and expect round-1 speed.

## Pattern breaks

None. Every BSD curve at conductor at most 30 is analytic rank 0 with L(E,1) nonzero, and every one of the first 2,000 Riemann zeros sits on the critical line by the same check round 1 used.
