# Notes from the verifier

## Re-run

The harness built the image from `env/Dockerfile` (python:3.12-slim, pinned by digest) and ran `code/run` with no network. It exited 0 after about 20 seconds; `code/analyze.py` wrote the declared results and `code/validate.py` passed its fixtures and reconstruction. All 12 results that C1 names agree with the declared values exactly, within the 1e-6 tolerance.

## Independent reconstruction

`independent_check.py` (output in `independent_check.json`) rebuilds the primary setting from the paper's Methods alone, without importing the bundle's code: it reads the raw NESO responses under `data/raw/`, keeps days with all 48 `actual` values, trains meteorological-season mean profiles on 2023 days, and, for every release date with a complete current and next day and at least 21 complete days among lags 2 to 29, schedules each half-hourly arrival of a 3-hour job at the earliest start within 6 hours that minimises the profile's window sum, in exact rational arithmetic. It gives 17,424 and 17,472 jobs, reductions of 11.236791% and 12.00568%, 17.613636% and 15.550595% of jobs with higher intensity, mean delays of 3.18222 and 3.185153 hours, mean seasonal intensities of 111.17232 and 113.700502 g CO2/kWh, and recent-minus-seasonal differences of -0.210323 and -0.067851 g CO2/kWh: every value C1 rests on, identical to the declared ones.

## Integrity flags

The node flagged nothing, and I found nothing to add. No file holds instructions aimed at verifiers.

## Hazard screen

none. The work is a retrospective accounting exercise on public national grid-intensity estimates, redistributed under the provider's CC-BY-4.0 terms with attribution and listed in `data/SOURCES.json`. It holds no private data, no secrets, and nothing that helps cause harm.