The Game Behind the Leaderboard — a standalone Quarto statistical feature by Srihith Jarabana (SJ), with a playable last-round duel and a bounded synthetic study of strategic risk in ability selection.
SJ approved all five original voice passages, the two new mechanism paragraphs
and step 1 source/hosting review. The article and final Impeccable polish are
complete as a private local edition. Approval records are in
AUTHOR_REVIEW.md and DESIGN_REVIEW.md.
PROJECT_STATUS.json is the authority for approval status; no prose is silently
approved. The repository must remain private. SJ alone handles any eventual
public switch. No Pages, deployment, release or public preview is configured.
This project does not modify the portfolio site.
The current visual edition uses warm ivory and clay, with Commissioner headings and Asap reading/interface text. DESIGN.md records the built system. The contents include expandable evidence subsections on desktop and mobile; references have explicit readable colours.
Read the source audit for the supplied working-paper
versions, claim checks and access limits, and
build and hosting notes for the QMD/Python/HTML
structure and the proposed manufacturinggenius.jarabana.com subdomain.
Open _output/index.html after rendering. It embeds styles, scripts and figures
and uses native MathML, so it works offline without Python, R or an HTTP server.
A native-control duel lets the reader choose an action, change the observed gap
and correlation, and distinguish a random play from exact expected prize share.
A computed benchmark and explanatory prose remain available without JavaScript.
Validated on Windows 11 with Python 3.13.13, Quarto 1.9.38, and
Node 24.15.0. Exact report dependencies are in requirements-lock.txt;
data/environment.json records the final runtime versions. Python 3.11+ is the
package minimum; the resolved lock is validated on 3.13.13. Other platforms and
Python versions have not been run. The lock excludes the local editable path
and marks Windows-specific packages.
Install the Quarto CLI and Python/Node. From Windows PowerShell in the project folder:
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements-lock.txt
.\.venv\Scripts\python.exe -m pip install --no-deps -e .
powershell -ExecutionPolicy Bypass -File scripts/reproduce.ps1
The reproduction script stops on an error. It runs the Python checks, exports
browser reference states, compares JavaScript with Python, executes every frozen
study cell, generates all figures and renders index.qmd. It sets
QUARTO_PYTHON to this environment; no PowerShell activation is required.
On macOS/Linux, the equivalent explicit sequence is:
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-lock.txt
.venv/bin/python -m pip install --no-deps -e .
export QUARTO_PYTHON="$PWD/.venv/bin/python"
.venv/bin/python scripts/check.py
.venv/bin/python scripts/export_reference.py
node scripts/check_browser_math.cjs
.venv/bin/python scripts/run_study.py
.venv/bin/python scripts/make_figures.py
quarto render index.qmd
The standard-library introductory duel can still be checked without installing
report dependencies using python -m unittest discover -s tests -p test_duel.py
after setting PYTHONPATH=src. The complete suite requires NumPy.
The browser mathematical check uses only Node. The rendered-page checks use the pinned development dependency and a temporary browser profile:
npm ci
npx playwright install chromium firefox
npm run check:article
For an existing browser, set MG_BROWSER_EXE to its executable path instead of
installing Chromium. Validated in headless Edge and Playwright Firefox.
The check opens the local HTML file, blocks external network requests, exercises
the guided introduction, all 48 study selections, keyboard and numerical
controls, layouts from 320 to 1920 pixels and the no-JavaScript fallback.
It saves screenshots and JSON under artifacts/local/ (ignored). For Firefox
in PowerShell, clear any MG_BROWSER_EXE override, set $env:MG_BROWSER='firefox'
and run npm run check:article again. On POSIX use
env -u MG_BROWSER_EXE MG_BROWSER=firefox npm run check:article.
Native MathML and the offline font were checked in both engines. Safari,
a physical touch device and an actual screen-reader session were not tested.
The introductory players have equal ability; their observed lead is not skill. The separate study randomly labels persistent abilities +/-delta/2 and solves the last-round constant-sum game under complete information for players and unobserved ability for the evaluator. The evaluator ranks cumulative scores. Fresh independent standardized evaluation cannot affect selection.
The frozen grid has 48 cells and 10,000 independent contests per cell. Under these assumptions equilibrium accuracy cannot exceed both-safe accuracy. That sign is derived, independently integrated and checked. Equilibrium minus prescribed both-risky can have either sign. Selected ability and expected fresh performance are identities linked to selection accuracy, not independent findings. Winner overstatement is descriptive, not evidence of dishonesty. All observations are synthetic score points, not financial returns.
The exact solver is retained from the starter. It is for two-by-two zero-sum/constant-sum games, with unilateral-deviation verification. It must not be reused for general-sum personal costs without a different analysis.
| Path | Purpose |
|---|---|
index.qmd, _quarto.yml |
Main authored source and static HTML configuration |
manuscript/ |
Approved author passages and proposed Coverage blurb |
assets/duel.js, assets/duel.html |
Guided duel and free exploration |
assets/study.js, scripts/article_views.py |
Selector and static views of saved results |
tokens.css, assets/report.css, assets/fonts/ |
Warm palette, styling and bundled proportional fonts |
assets/navigation.js, PRODUCT.md, DESIGN.md |
Contents behavior and the current product/design constraints |
src/manufacturing_genius/duel.py |
Preserved exact game and solver |
src/manufacturing_genius/study.py |
State caching, simulations and exact benchmarks |
data/study_config.json |
Grid, seed streams, figure and precision plan frozen before results |
data/generated/study.json |
All policy estimates, paired differences, exact accuracy and audits |
data/generated/*.svg |
Original full grids and focused presentation panels |
data/reference/duel_benchmark.json |
Python reference for all slider combinations and boundaries |
scripts/, tests/ |
Analysis, reproduction, numerical and browser checks |
docs/MODEL_SPEC.md, docs/STUDY_DESIGN.md |
Scientific contract and finalized design |
docs/DESIGN_DECISIONS.md |
Design freeze and implementation rationale |
docs/SOURCES.md, references.bib |
Verified source metadata and access limits |
docs/VALIDATION.md |
Actual checks, limitations and clean-checkout record |
docs/AUTHOR_REVIEW.md, docs/DESIGN_REVIEW.md |
Completed original voice pass and two targeted new passages |
Generated JSON and SVG outputs are deliberately tracked. PNG previews, rendered HTML, environment folders and browser artifacts are ignored and regenerated. No external dataset, database, frontend framework or hosted service is needed.
SJ’s supplied opening/ending revisions and his explicit approval of the three
remaining passages are recorded in docs/AUTHOR_REVIEW.md. AUTHOR_APPROVED
records completion of that voice checkpoint. The source/hosting review is also
approved. SJ approved the two new explanatory paragraphs on 2026-10-26,
as recorded in docs/DESIGN_REVIEW.md. Final polish and browser validation
are complete; no author decisions remain pending.
The original five-passage review is not restarted. No author approval authorizes publication, deployment, DNS changes or a repository visibility change.