Spaces:
Running
Implementation, compute, and provenance
Implementation path
repro_campaign/run.py reads one committed config and imports every accepted claim
module. Any failed certificate, assumption audit, numerical gate, or control raises
and exits nonzero. The command and environment never vary between nodes.
uv run --frozen python -m repro_campaign.run
Python 3.12.12; locked formal-run packages include NumPy 2.5.1, SciPy 1.18.0, scikit-learn 1.9.0, SymPy 1.14.0, Torch 2.13.0+cpu, matplotlib 3.11.1, and marimo 0.23.15.
Experiment lineage
| Branch purpose | Scientific commit | Result | HF job time |
|---|---|---|---|
| Frozen judged toy baseline | 4120eb7 |
C1–C5 TOY, C6 BLOCKED | 26 s |
| C1 exact certificate + corrected calibration | 20bf5fe |
C1 VERIFIED | 53 s |
| C2–C3 arbitrary-eta/non-Gaussian | 82ebd6a |
C2/C3 VERIFIED | 59 s |
| C4 quantifiers + C5 contradiction | a6ace14 |
C4 VERIFIED, C5 FALSIFIED | 58 s |
| C6 GMM population route | 6bdd221 |
aligned partial evidence | 58 s |
| C6 Fashion feasibility | 970446a |
BLOCKED | 69 s |
| C6 CIFAR source contract | 6784a9e |
BLOCKED | 69 s |
| C6 falsification search | 19aa525 |
BLOCKED | 69 s |
Every long or multicore task ran on Hugging Face cpu-upgrade, estimated at eight
workers. The flavor advertises 8 vCPU; containers reported 64 logical/affinity CPUs,
so implementations explicitly bounded active parallelism to eight. Short static
syntax and JSON checks used one local CPU core for less than five minutes.
Source and protected baseline
Paper HTML retrieved 2026-07-28 with explicit User-Agent; SHA-256
472fb9e246aea7c5d1e643d5755f034f005d1f2bbb61f86371bae855f62765e3.
TeX source SHA-256
c25e50d4fe4c1a730754e525f3065754bc75232d0d3bb890f1a383e738b7bffa.
The exact judged Space 49401cddb554d5c3f7ae98d400567b8d6f10c028 was
downloaded before changes. Its 15 paths remain present; original current entrypoint
texts are copied under
the historical README mirror
with the protected manifest.