Reviewer-facing companion repository for the TAS-AI paper.
Unless otherwise noted for third-party materials, the contents of this repository are released under the NIST data/work statement in LICENSE. In the United States, works of NIST employees are not subject to copyright protection under 17 U.S.C. 105.
This repo is intentionally narrower than the full working paper repository. It contains:
- the current manuscript and supplementary-information sources
- rendered manuscript outputs, with
docxas the preferred reviewer-facing format - the exact figure assets referenced by the current manuscript snapshot
- the paper-facing scripts used to generate figures and summaries
- curated archived data supporting the quantitative claims in the paper
It does not include local operational mailbox logs, prompt experiments, old manuscript versions, slides, or the full TAS-AI library source tree.
- paper/
- canonical manuscript source and outputs
- paper/figures/
- figures referenced by the current manuscript and SI
- paper/scripts/
- paper-facing figure/table generation scripts
- paper/data/
- curated archived benchmark, pilot, ablation, and sensitivity artifacts
- REPRODUCIBILITY.md
- current reproducibility guide copied from the working paper repo
- REVIEWER_GUIDE.md
- reviewer-facing provenance guide copied from the working paper repo
For reading, start with:
Source-of-record files in this repo are:
These correspond to the current v5 manuscript snapshot from the working paper repo, but are renamed here to canonical names so reviewers do not have to navigate version clutter.
The paper/data/ tree includes:
final/- benchmark and pilot summaries used for the main paper tables and figures
ablation_runs/- archived Section 5 ablation outputs, including the five-seed ghost-optic and bilayer reruns
- top-level coverage and sensitivity JSON files
- Laplace-coverage and reviewer-sensitivity artifacts used in the SI
See paper/data/README.md for more detail.
The paper/scripts/ directory contains the paper-facing scripts used to generate the current figures and supporting summaries. This includes the scripts for:
- the Figure 4 benchmark panels
- the Figure 1 workflow diagram
- the closed-loop pilot and overseer drivers
- the Section 5 audit-ablation harness
- the exchange-path analysis figure
- the PySpinW benchmark figure
- the Figure 7 hybrid-routing figure
- the ghost-optic schematic
- the bilayer ablation figure
- the pilot LLM multipanel figure
- the taper comparison and sensitivity summaries
The manuscript can be rebuilt with Pandoc using:
This repo is paper-facing. The reusable TAS-AI library code lives separately in:
Reviewers interested in the underlying library implementation should consult that repo, while reviewers focused on manuscript provenance should be able to work entirely within this one.