Reproducing the project

Commands

make sync restores the Python environment from uv.lock. make check runs Black, isort, flake8, and pytest. make ci-docker runs the same checks using the standard python:3.14-slim image, without modifying the mounted checkout.

make fetch
make prepare AS_OF=2026-10-26
make compare
make publish
make site

AS_OF defaults to the current UTC date; supply it explicitly for reproducibility. DATA_DIR defaults to data/raw/cricsheet/odis. Preparation can also read the existing local data/odis YAML directory. Numeric filenames are Cricsheet match IDs; mixing two representations of the same ID is an error.

The website is rendered with Quarto 1.10.19 in CI. QUARTO can point to a local binary. The homepage includes the README automatically. Rendering does not acquire data, execute the exploratory notebooks, or refit models. A post-render check verifies local links and assets. Before rendering, saved publication checksums are verified. Format pages and the coverage table are generated automatically from the saved rankings and manifest. The publication is included in the repository so CI and readers can render it without fetching match data or estimating models.

make compare reads the prepared tables and verifies their hashes. It runs the specified three models on 2023–2025 annual holdouts, with separate gender fits and three- and five-year training windows. MODEL_ARGS can override years/windows for an explicitly exploratory run. See the specification.

make publish verifies saved analysis files and copies them into a publication with generated tables and download links. Changed analysis files fail validation before the preceding publication is replaced. Predictions, paired losses, tuning scores, coefficients, calibration, diagnostics, and source provenance are retained. Publication also ranks the saved player coefficients within each competition, model, window, and cutoff. It generates the ranking page and rankings.csv without refitting the model. Publication code hashes are separate from the preserved analysis code hashes.

Rankings by format

After preparing ODIs, acquire and prepare the other two international formats:

make fetch FORMAT=Test
make prepare FORMAT=Test AS_OF=2026-10-26
make fetch FORMAT=T20I
make prepare FORMAT=T20I AS_OF=2026-10-26 PREPARE_ARGS="--since 2021-10-27"
make rankings AS_OF=2026-10-26

FORMAT selects the official archive and format-specific default directories. The general --since preparation option filters by match start date and records excluded files with their hashes. Here it sets the T20 input to the five-year ranking window, excluding an older identity conflict documented in the specification. Update that boundary when changing the cutoff. Included matches still undergo strict identity checks. Ranking refuses prepared inputs that omit part of the requested window.

make rankings fits the six cohorts and replaces out/release/current/ only once all fits succeed. It saves an all-player CSV, six format/gender CSVs, reconstructible model parameters, preparation manifests, the specification, and code/output hashes. Ranking cutoffs and tuning split IDs are recorded. Models are fitted during this command; website rendering only checks and displays saved files. The CLI accepts --specification when running outside the source checkout.

Layout

Location	Contents
data/raw/cricsheet/odis/	Downloaded archive, JSON matches, and acquisition receipt
data/build/odis/	Validated match, participant, and identity tables with manifest
src/player_value/	Acquisition, preparation, CLI, and website checks
tests/	Synthetic fixtures and behavior tests
experiments/war/	Preserved exploratory notebooks and results
out/analysis/lineup/	Saved fits’ coefficients, forecasts, tuning, and evaluation
out/release/current/	Five-year rankings by format, model parameters, and provenance
out/release/lineup/	Verified research publication and generated summary
docs/	Website source and research specifications
out/site/	Rendered website

Additional models will keep their own results under out/analysis/ and verified publication tables under out/release/. The format-ranking command publishes rolling-window associations. Dynamic current-strength and WAR models remain future work.

Prepared tables

matches.csv has one row per match. Dates, gender, format, teams, venue, toss, outcome, delivery count, and lineup eligibility are preserved. team_a and team_b are alphabetical labels, not home/away designations. Boolean flags are 0/1. Empty cells mean missing or not applicable, not zero.

participants.csv has one row per match, team, and player ID, retaining the source’s player name. players.csv has one row per stable ID, the latest observed name, and the first and last appearance dates within the prepared sample.

manifest.json records the cutoff, source-file hashes and exclusions, counts by gender, preparation-code hash, output hashes, and the acquisition receipt when available. Existing local downloads without a receipt are labeled with a null receipt; preparation does not invent a retrieval date. A cutoff filters match dates but does not recreate the source data as they existed on that date.

Acquisition and preparation build a temporary directory before replacing their destination. Errors preserve the previous successful stage. If an interrupted installation leaves a .previous directory, inspect and resolve it before rerunning. Raw downloads, build products, and analysis intermediates are ignored by Git. Verified publication files under out/release/ are intended for version control.

Repository transition

This project continues outside-edge/war under the name outside-edge/player_value. The existing history is retained. The original artifacts have a checksum relocation record in experiments/relocations.csv; their numerical contents are unchanged. The local checkout directory can retain its existing name without affecting any commands. There is no dependency on ../side_order.

CI checks the package, builds its distribution, and renders the website. The Pages workflow deploys the website after checks pass on pushes to main, and can also be triggered manually. Scheduled rating updates and player-value releases remain gated on the model-validation criteria in Model comparisons.