program-structure-certificates / RELEASE_SCOPE.txt
system's picture
system HF Staff
Mirror GitHub f31f2ceadd54
fd74a4d verified
Raw History Blame Contribute Delete
1.55 kB
PROGRAM STRUCTURE CERTIFICATES | RESEARCH ARTIFACT v1.0.0
The archive covers 120 synthetic task families, 480 paired base-count
variants, 960 arm-specific certificates, 1,008 saved retrospective archive
tasks, 7,680 wider evaluation targets, and 1,920 normalized model outputs.
All trajectories retain generated text, generated and prompt token IDs, and
the deterministic selected-program status. The base-three, four, and six
records retain rendered prompts. The base-five records retain reconstructed
user messages, prompt-text hashes, and the pinned model chat templates.
The CPU entry points verify the file manifest, replay the Figure 1 example,
recheck the certificates and reference labels, and reproduce the count-panel
paired effects and family-level outcomes. The full replay uses Python 3.11+
and NumPy. It reads saved model outputs with zero model or GPU calls.
The paper also reports earlier discovery, corrected Phi, attention-validation,
timed symbolic search, and three-model prefix analyses. For these analyses,
the artifact inventory separates paper tables and protocols from underlying
historical records marked not_in_package. The original resampling arrays are
also listed there. The original adjusted intervals in the first two result
tables use their historical resampling schedule. The count-panel replay
recovers the Qwen and Olmo base-five point estimates.
See docs/reproduction.md for commands and expected output, docs/result-map.csv
for paper-to-artifact links, and DATA_PROVENANCE.md for source and reuse terms.