File size: 1,927 Bytes
f164af8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 | # Contributing to screening-ceiling
The one non-negotiable rule:
> **`verify.py` imports nothing from this project.**
It is pure standard library, hand-rolled linear algebra included, so a skeptic
can read the whole thing in one sitting and run it with nothing installed. The
moment it imports a helper, the reader has to trust code they did not read, and
the point of publishing a verifier alongside a theorem is gone.
That rule costs real performance — the Gauss-Jordan inverse is the hot path and
a closed form would be several times faster. It is not worth trading legibility
for a fraction of a second in a file whose product *is* its legibility.
## What a good contribution looks like
- **A counterexample.** The most valuable thing you can send. If you find a
layout in the family box that violates the bound, that refutes the theorem,
and we would rather learn it from you than not learn it. Include the four
parameters and what you measured.
- **A second, independent implementation** of the ceiling evaluation. Two
implementations that agree are worth much more than one that is fast.
- **A scope correction.** The bound is a theorem about the **monopole-closure
model**, not about Maxwell's equations, and the closure-versus-solver gap is
additive, disclosed and unresolved. If you find anywhere in the data, the
README or `theorem.json` that blurs that line, say so — the scope statement
travels with the data specifically so it cannot be separated from it.
## Before you open a PR
```bash
python verify.py # nothing to install
python -m pytest tests/ -q # needs pytest, numpy
```
Every figure in the README must be re-derivable from the published rows. There
is exactly one that is not — the `0.081%` closure-versus-solver agreement, which
came from a solver that is not part of this release — and it is labelled as such
at the figure. Do not add a second.
|