epstein_height: run manifests

id: epstein_height-2026-09-09-surface
hunt: epstein_height
started: 2026-09-09T23:47-05:00
finished: 2026-09-10T00:20-05:00
ran:
  - .venv/bin/python hunts/epstein_height/probe.py --heights 10 20 40 60 80 100 120 160 --dps 15 20 30 50 80
  - .venv/bin/python hunts/epstein_height/from_log.py
  - .venv/bin/python hunts/epstein_height/boundary.py
  - .venv/bin/python hunts/epstein_height/law.py
  - .venv/bin/python hunts/epstein_height/law2.py
outcome: at dps 15 and height 120 the routine returns a value wrong by 2.3e+36 with no exception and no warning; the recorded constant plus a sigma term leaves a per-form bias of 1.6 digits; the loss re-derived from what the routine forms is -log10 t + sigma log10 pi - log10|Gamma(s)| - log10|zeta_Q(s)| with no fitted constant, and it reduces the per-form bias to 0.15, leaving one uniform offset of +1.068 with rms 0.357 over 49 transition cells
artifacts:
  - hunts/epstein_height/artifacts/surface.json
  - hunts/epstein_height/artifacts/boundary.json
  - hunts/epstein_height/artifacts/law.json
  - hunts/epstein_height/artifacts/law2.json

id: epstein_height-2026-09-10-guard
hunt: epstein_height
started: 2026-09-10T00:25-05:00
finished: 2026-09-10T00:35-05:00
ran:
  - .venv/bin/python hunts/epstein_height/guard.py
outcome: six planted faults, all passing, including the one that makes the others mean anything: with the cancellation constant set to zero the guard stops firing
artifacts:
  - hunts/epstein_height/guard.py

Notes

• The surface run was cut short by its own wall clock partway through the third form, and its JSON is written at the end, so from_log.py rebuilds it by parsing the printed lines. That is recorded rather than done by hand so the reconstruction is reproducible and so the third form's partialness is obvious.
• The rectangular grid put only twelve of ninety-eight cells where a prediction could be tested, the rest at the oracle's ceiling or flat at zero. boundary.py places cells where the law says the transition is, which is legitimate for measuring a residual and would not be legitimate for deciding whether there is an effect; the grid had already answered that.
• why_it_hangs.py measures the acceptance probability of the winding recursion at both precisions. The conclusion for hunts/dps_cap's open question does not depend on the measured value: because the depth limit returns a value rather than raising, the cap cannot cost a refusal.

id: epstein_height-2026-10-10-landing
hunt: epstein_height
started: 2026-10-10T15:44+00:00
finished: 2026-10-10T16:23+00:00
ran:
  - .venv/bin/python hunts/epstein_height/probe.py --heights 10 20 40 60 80 100 120 160 --dps 15 20 30 50 80
  - .venv/bin/python hunts/epstein_height/boundary.py
  - .venv/bin/python hunts/epstein_height/why_it_hangs.py
  - .venv/bin/python hunts/epstein_height/guard.py
  - .venv/bin/python hunts/epstein_height/guard2.py
  - .venv/bin/python hunts/epstein_height/probe.py --heights 40 60 80 120 160 --dps 15 20 30 --forms "1,1,4;2,1,3"
  - .venv/bin/python -m pytest -q hunts/epstein_height
outcome: against epstein_zeta as it was before #291 (the height lift set to zero), the 120 grid cells and the 42 boundary cells reproduce every field except digits_available, which the artifacts carry as dps + 10 from an earlier probe and the current probe writes as dps + 20; why_it_hangs.json and guard2_contract.json reproduce exactly; guard.py passes its six rungs; on the current routine the 30 re-measured cells all read 16.1 to 16.5 digits, guard2.py passes and guard.py fails its second rung; four stated numbers do not reproduce and are marked where stated
artifacts:
  - hunts/epstein_height/test_epstein_height.py
  - hunts/epstein_height/RESULTS.md
  - docs/41-the-module-that-missed-the-guard.md

Every script above was run on 2026-10-10 in a scratch copy of this directory, first with zeta.epstein.EPSTEIN_DIGITS_PER_UNIT_HEIGHT set to zero (the routine before #291, exactly, since epstein_completed did not change), then the last two with the lift in place. Outputs were compared field by field with the committed artifacts and left in the scratch copy. Measured on a 4-vCPU container shared with other jobs: the grid took about 15 minutes, the boundary cells 5, why_it_hangs.py 22 (its 48 evaluations at 110 working digits are the cost), the two guards 1.5, and the lifted grid 14.

• Landed on main and renumbered #120 to #129, 2026-10-10. On main #120 is qrh_rival_step; #127 is euler_defect_axis and #128 is li_dh_onset (#298). The front door, numbered 37 on the branch, is docs/41; main uses 37 for the methods index. AUDIT.md keeps the branch numbering and says so.
• The repair landed before the hunt did, and it is a different repair. #291 (2026-10-09, "Fixes #217") lifts epstein_zeta's working precision by the leading term, ceil(0.6822 |t|), and pins it with a lattice-sum test of the kind this hunt built. It refuses nothing and carries no sigma or zeta_Q term; its docstring says it cannot help at a zero of zeta_Q. The surface, both laws and the audit's counterexample therefore describe the routine as it was, and are reproduced against it; the current routine is recorded beside them.
• surface.log was never committed. The first note above and AUDIT.md attack 14 describe it; it is not in the tree. from_log.py therefore raises FileNotFoundError before it can write anything, which the test pins.
• smoke.json has no manifest. Its four rows are identical to the same cells of surface.json; it reads as the probe's first smoke run, kept.
• surface.json is the grid plus the boundary cells, merged by hand. 158 rows: the 120 grid cells and the 38 boundary cells that do not repeat a grid cell. No script performs the merge; the test checks its composition.