The public-facing writeup, reference pages, and the figures plus data they
render from. Read this when you want to know what lives where under docs/.
index.html: the interactive web essay. The base comes frompaper/build_paper.py, with hand-maintained navigation and interactive additions. The builder refuses to overwrite these; keep shared style and copy edits in both files and preserve the overwrite guard.paper/: essay and paper source:paper/src/(LaTeX + bib),paper/data/(the JSON each interactive figure reads),paper/build_paper.py(the builder).figures/: figure assets used by the essay: static PNGs plus interactive D3 versions (*.html+ their*.json), rendered through the sharedexplorer/d3charts.jsmodule. Generated byanalysis/fig_*.py.explorer/: a static, embeddable copy of the trace explorer.assets/: shared CSS, images, and the licensed Instrument Sans font for the hosted pages. Font files andfonts.cssare mirrored underhf-space/static/so each host can serve them locally.build/: the ecosystem-map builder and its output (a self-containedindex.html, generated, not served by Pages).METRICS.md,STUDIES.md: the measurement reference and worked case studies.
GitHub Pages serves the site from this directory; index.html is the landing
page.
Plot outputs resolve to this checkout's docs/figures/, regardless of the working
directory. The patch-size, trajectory-length, and step/token-cost scripts read
from a sibling bidirect-align-dev-traces checkout by default. Set BIDIRECT_ROOT
to use a different location, for example:
BIDIRECT_ROOT=/path/to/bidirect-align-dev-traces python analysis/fig_regression_length.pyThe paper cache reader accepts --cache-dir /path/to/trajectory-cache. These
external inputs are not bundled with procgrep.
python analysis/render_essay_figures.py renders the four static figures backed
by paper/data/figure_inputs.json. The saved values include source hashes and
observation units. The render step checks the bundled Instrument Sans font and
writes directly to the image paths used by the essay, with provenance beside
each PNG. The older four individual entry points call the same renderer.
The original extended-cohort inputs for the action/token, actions-per-resolution, patch-size, and trajectory-length figures are unavailable locally. Those four image assets remain historical until their inputs can be recovered; do not rebuild them on a different corpus as a styling change. Other interactive figures read the saved JSON in this checkout; their text uses the same bundled face.
python analysis/fig_induced_vocab_tree.py renders the saved historical vocabulary
in paper/data/induced_vocab_tree.json, without network access. That snapshot was
recovered using code 413ef3d08defc0e1da31043cfc2f71fb5a3199c3 and
SALT-NLP/SWE-chat@f66cca95b14caaa4177f7ed5eaa424608dadcffa (the only revision
predating the original figure). The original first-25-per-interface selection and
adapter rules retain 89 of 100 sessions with at least two actions. The derived
snapshot records selection IDs, transcript hashes, exclusions, merges and agent
presence; all displayed trees and presence marks match the historical image.
No GCP resources are needed for either rendering command.