rescore.py — re-apply the observation operator to runs already computed.
python rescore.py --dry-run # show what would change
python rescore.py # rewrite and publish the corrected records
WHY. The dose ladder scored the wrong dose and said it had not (b5e9a51): the
end point rounded past the final grid point, and a requested dose that was not
itself a grid point silently took the last point BELOW it, however far below.
On the exact arm `T4_D_4k` the "0.2 dpa" column carried 0.0665 dpa numbers -- 3x
low -- while `reached_all_scoring_doses` reported True. Every run computed
before that fix carries columns labelled with doses they were not taken at.
The trajectories are sound; only the extraction was wrong. So this re-reads
each run's saved solution and re-runs `run_ensemble.observe` under the current
code, rather than recomputing 17 hours of solver time.
SCOPE. A run can only be re-scored on the machine that produced it: the
trajectory lives in `output/<stamp>_<run_id>_.../plots/plot_data.pkl`, and
`output/` is gitignored (a 38 MB pickle per run, and one of live objects that
will not survive a class refactor -- see .gitignore). What crossed the wire is
the compact record, which is the thing being corrected. Run this on EVERY
machine that computed rows; each fixes its own and pushes.
