# One reservoir, delayed sources, and a stellar birth population

This independent scientific Python package is separate from the browser Lab.
It contains original MIT code, fixed effective scalar inputs, and 96 original
synthetic measurements. It contains no original paper, observed-star rows,
observed-derived figure, App/Engine source, VICE executable or yield grid.

## Environment

Tested with Python 3.12, NumPy 2.3.5 and SciPy 1.17.1. Use an existing compatible
environment or prepare your own environment from requirements.txt. Nothing is
automatically installed. Optional extraction requires an existing system
`pdftotext`; it is unnecessary for all synthetic calculations.

Run from this unpacked directory. Outputs refuse overwriting existing files.
Each numeric producer is limited to 120 seconds. Default requests use no HTTP.
The scripts really solve the four histories; no preflight result constants
are substituted for producers.

## Default offline calculation

```sh
python science/verify_package.py
python science/make_records.py --save-every .005 --output results/records
python science/make_reference.py --records results/records --output results/reference
python science/validate.py --records results/records --reference results/reference --output results/validation.json
python science/compare.py --records results/records --output results/synthetic-comparison.json
python science/plot.py --records results/records --output results/figures
```

`make_records` computes the full chemical solution and keeps all fourteen
physical fields every 0.05 Gyr, plus every 0.005-Gyr integrated birth cohort.
The reference actually solves at 0.0025 Gyr with tighter tolerances. Its archived
NPZ retains all fine cohorts and computed per-field refinement maxima, and
hashes the actual full fine record. Full fine bulk arrays are not archived.
Positive GL4/GL8 observation integrals interpolate fine gas and element masses
before logarithms, using analytic formation weights. This is an independent
observation operator on the same chemical solution, not an extra galaxy model.
All NPZ arrays are lossless Float64 and are read without pickle.

The physical kernel uses prescribed SFH and inflow, R=0.4, trace metals,
instantaneous recycled material, a pre-enriched initial gas, and a well-mixed
wind. A stronger wind need not lower abundance at fixed time when SFH is
prescribed. The fixed Kroupa/VICE effective net yield convention retains
nonzero Ia magnesium. The normalized delay kernel spans 0.15--15 Gyr; stopping
at 8 Gyr does not renormalize its unfinished tail.

The earlier 0.02/0.01-Gyr birth-bin trial failed the unchanged 0.01 mean absolute
per-star log-density gate (maximum 0.0849854791574). It is not represented as a
pass. The finer 0.005/0.0025 finite checks pass for four histories, two samples,
three assumed correlations and separate one-axis +/-0.1-dex model shifts.
The public validator recomputes its own synthetic reference and analytic limits.
It does not depend on the author's private real catalogue.

## Optional fixed Hill source, kept on your computer

```sh
python science/fetch_hill.py --output reader-private/hill-v2.pdf
python science/extract_hill.py --pdf reader-private/hill-v2.pdf --output reader-private/hill-observations.json
python science/compare.py --records results/records --data reader-private/hill-observations.json --output reader-private/hill-comparison.json
python science/make_reference.py --records results/records --real-data reader-private/hill-observations.json --output reader-private/reference-with-hill
python science/validate.py --records results/records --reference reader-private/reference-with-hill --real-data reader-private/hill-observations.json --output reader-private/validation-with-hill.json
```

The named fixed URL is arXiv 1812.01486v2. The actual source must have 1,127,947
bytes and SHA-256 762b98b165ae7059ac8f93f2f8bf69006eee3bfffeb4c348bfcb5e8810da2f1f.
No retry, redirect, mirror or upload is automatic. To test from an existing cache,
replace only the first optional command with `python science/fetch_hill.py
--cache-file /absolute/path/hill-v2.pdf --output reader-private/hill-v2.pdf`.
That path performs zero HTTP and is not a fresh network-fetch test.

Both pdftotext layout/raw extractions of printed Table C5 must agree on 99
unique rows and all fields. Ninety-six have finite Fe/H, Mg/Fe and positive
marginal errors; three missing rows stay in the full local record. ID suffixes
divisible by five are withheld (real 80/16, synthetic 77/19). This partition is
not a coverage calibration.

The RGB luminosity/spatial/member selection has not been modeled. Birth-mass
weights are not a recovered RGB selection function. The full Hill covariance
and a numerical Hill-to-model solar zero-point conversion have not been
established. Rho=-0.7/0/+0.7 and separate model-axis shifts are sensitivity
assumptions, not measured error or solar corrections. Scores are normalized
conditional descriptive densities for four specified histories, not a fitted
Sculptor SFH, a calibrated likelihood, a posterior or unique ages. No r-process,
Eu production or spatial abundance scatter is simulated.

## What validation establishes

The validator reads actual records/reference hashes and checks field maxima,
positive early Mg/Fe support, observation refinement, analytic gas inventory,
element conservation, continuous DTD normalization/convolution and a changed
integration-order source integral. It includes independent prompt-wind and
positive conservative two-reservoir mixing limits and Gaussian normalization.
Domain/gas-exhaustion/non-positive covariance controls must refuse. A passing
exit applies to these finite checks only. Retain all outputs and failed trials.