data/ — the member-winter cache

member_winters.csv is written by load_cache(rebuild=True) in ../lensml.py (run from chapter 1, The member-winter table) and read by every chapter. Its first line is a # header recording the parameters it was derived with. The committed copy says derived by 02_applications.ipynb, this cookbook’s earlier single-notebook version, which produced it on 2026-10-20; the same derivation now lives in lensml.py, and a rebuild writes derived by lensml.py. The header records (SEASON, WINDOW, FRONT_RANGE, NINO34, n_members_per_half, RNG_SEED, access date). One row = one member-winter: member_id, forcing_variant, winter, nino34_son_K, prect_winter_mm_day — the raw SON Niño 3.4 box mean and the raw season-mean precipitation; the forced response is removed and the target defined in chapter 2, on training members only, so the cache carries no split-dependent quantity.

Shape of the committed copy, each value measured from the file itself (the command in the last column, run over the lines after the # header with csv.DictReader; chapter 1 prints the member and winter counts it reads at run time):

• Every chapter asserts the header against lensml.py (SEASON, WINDOW, FRONT_RANGE, NINO34, AUTUMN_MONTHS, and the access= date against ACCESS_DATE) before reading the table. A cache derived with other parameters stops the run with a message to rebuild, so a recorded decision (a changed SEASON, or a changed autumn) cannot be silently ignored. The committed header predates the AUTUMN_MONTHS field, so the file’s autumn is assumed, not read: the reader takes a header without it as AUTUMN_MONTHS=[9, 10, 11] (SON, the only autumn the derivation code had when the table was written) and prints a line saying so. A changed AUTUMN_MONTHS in lensml.py still fails against that assumption; a rebuilt header records the field, so the check then reads it from the file.

• The committed copy is the full table: 50 members per half, every member of both halves (the header records it, and chapter 1 prints the member and winter counts it actually read). It was produced by a full rebuild — a background job: the one full derivation took 33 min on a fast link (2026-10-20); on home broadband expect an hour or more (chapter 1’s chunking line computes the volume from the store: four 600-month chunks of about 133 MB, read whole, ≈ 531 MB per member-variable, three variables, ≈ 159 GB uncompressed for 100 members; the transfer is smaller) — and the access date was written into the header’s access= field. A rebuild is a new access: set ACCESS_DATE in lensml.py to the day of the rebuild first: load_cache(rebuild=True) writes it into the header and every figure’s citation; on every other read the header’s access= is asserted equal to ACCESS_DATE in lensml.py, so the two cannot drift.

• A quick sample (REBUILD = True in chapter 1 with a smaller N_MEMBERS_PER_HALF) runs the same code: the test hold-out and the fold count then follow the data (N_FOLDS = min(N_FOLDS_MAX, training members)), and the chapters print the counts they actually used and say when the fold re-draws collapse.

• Source and required citation (NSF NCAR’s Data Citation section, which the UCAR Data Terms bind users to): Grover, M., & Banihirwe, A. (2021). NCAR CESM2 Large Ensemble Cloud-Optimized Dataset. NSF NCAR CISL. doi:10.26024/y48t-q717; and Rodgers et al. 2021, doi:10.5194/esd-12-1393-2021. UCAR Data Terms of Use (CC BY 4.0). Access line: Community Earth System Model v2 Large Ensemble (CESM2 LENS) was accessed on 2026-10-20 from https://registry.opendata.aws/ncar-cesm2-lens. (the header’s access= field)

member_winters_ssp370.csv (chapter 7)¶

The same 100 members, boxes, DJF season and SON autumn under SSP3-7.0, winters 2016–2100 (DJF 2015 would need December 2014 from the historical store), derived from the ssp370 stores by derive_future_cache in ../lensml.py with the same derive_member_winters. Its header records the experiment, window, parameters, member count and access date (2026-10-29); load_future_cache asserts that the header contains FUTURE_RECORD (experiment, season, window, both boxes, autumn months) and that its access= date equals FUTURE_ACCESS_DATE in lensml.py. ssp370-test-protocol.md, beside it, is the protocol chapter 7 follows, verbatim. Chapter 7 is the only reader, under a protocol written, by the author’s account, before the file was derived (self-attested; no external timestamp). The protocol is kept exactly as written, so two of its phrases predate the published text: its title and chapter 7’s printed rule say “pre-registered”, meaning fixed in advance by the author, not lodged with a registry; and its “seven limits” refer to a draft of chapter 6 that day, not in the repository history (chapter 6 now lists eight). Its second line (Glen chose "2") is the author’s note of choosing a fresh test; the “2” numbers an option in a list that is not in this repository. Its “scored on 2026-10-20” omits the neural network’s second scoring of the test members, on 2026-10-29 (chapter 6). Chapter 7 compares the file’s SHA-256 with its hash as committed with the test (851d900), so any edit, even a typo fix, shows there as “differs”; an amendment belongs in a new, dated file beside it.