Sampler budget#

How many HMC iterations the TEXAS calibration actually needs, and how many the inverse model needs — which turns out to be a different question with a different answer. This is Text S3 and Figs. S18–S19 of the manuscript: the published fits use 4 chains × (400 warm-up + 1,000 sampling) iterations, and every budget that clears the convergence gates agrees with a long reference run to within 0.07 SD for every parameter. The inverse reconstructions use 300 warm-up iterations with M = 300 forward draws for the global coretop maps and 500 warm-up iterations with M = 500 for the downcore records.

The page below reports, for each model, the cheapest warmup/sampling budget that clears four convergence gates, together with the evidence behind it: the full grid of R̂ failures, the choice between gating on all parameters or on the calibration parameters alone, and the inverse model’s sweep over budget and M scored against measured coretop SST.

Every number on it is read from the sweep’s own output by docs/_scripts/build_sampler_budget.py, so the page cannot drift from the run the way a hand-written summary does.

Open the sampler budget report full-page →

Reproducing it#

The sweep itself is scripts/paper/run_param_sensitivity.py, which does the sampling unattended and is resumable per fit:

python scripts/paper/run_param_sensitivity.py all      # forward grid + proxy refits
python scripts/paper/run_param_sensitivity.py part3    # inverse budget and M
python docs/_scripts/build_sampler_budget.py     # rebuild this page

notebooks/manuscripts/SI_code02a_model_param_sensitivity_test.ipynb runs the same analysis interactively and draws the figures; a test fails if the two configurations drift apart.

Why the page ships with its data#

data/revision1/groupA/param_sensitivity/*.csv is tracked through Git LFS, so a clone without LFS objects (or a Zenodo zip of the repository, which carries only the LFS pointers) does not have the grid. So the build script keeps a compact _static/sampler-budget.data.json alongside the HTML, holding exactly the numbers the page displays. Both are plain tracked files, which means:

  • the report is readable from any clone, with no data and no rebuild;

  • build_sampler_budget.py reproduces it byte-for-byte from the snapshot on a machine that has never run the sweep;

  • where the raw output is present, the script reads that instead and rewrites both files.

With neither source available the script leaves the committed page alone and exits cleanly, so a docs deploy can never blank the report just because the data lives on another machine.