878 concepts (+71 net), 2,020 relationships (+326, -25, 25 reworded), 27
redefinitions, 67 alias sets, 33 background tags, 6 moves; guides 16 roles,
24 symptoms, 7 paths; gaps: attention-FFN disaggregation closed, verifiable
ML computation and reasoning-trace monitoring declared. Coverage judgments
updated for the new concepts and a blind held-out checklist added
(OSDI/NSDI/EuroSys/ASPLOS 2026, 202 papers, 22.3% covered, 96.5% homed).
Symptoms now declare the quantity they are measured by (`quantities`), so a
symptom's own metric no longer poses as a cause; the compiler, the StaffML
loader, and the explorer accept a check or lever that touches a cause or the
quantity and reject a quantity listed in another bucket.
README and CHANGELOG record the method, the numbers, and the open author
decisions.
The page moves from site/explore/ to site/taxonomy/, matching its title,
its nav label, and the identifier namespace the taxonomy already used.
/explore/ keeps unlisted noindex redirect stubs, with relative targets
so they resolve on both the live site and the dev preview subpath.
The binder now holds the folder and the public address in one place
(concept_graph.SITE_DIR and PAGE_URL), and the generated page data carries
the address so the on-page citation reads it instead of repeating it.
The dev preview rewrites links inside taxonomy/ like the other subsites.
The labels file now lists the 20 taxonomy areas in order (code, group,
guiding question, subjects) and each subject's definition. The committed
corpus summary is rebuilt, so every question carries its taxonomy tags;
its textbook links follow dev's renamed chapter folders.
Checks #11 (topic exists) and #14 (prerequisite DAG) and the release
hash read vault/taxonomy.yaml, which no longer exists, so both checks
passed without checking anything. They now read
schema/taxonomy_data.yaml. Turning #11 on found one question filed under
an unregistered topic (sequence-parallelism); it moves to
model-tensor-parallelism, and its taxonomy tags already name sequence
parallelism. Tests pin that both checks fail when they should.
Questions gain optional taxonomy_subject and taxonomy_concepts (identities
from shared/concepts). The model checks their format and vault check
--strict checks they exist, naming the replacement for a retired concept.
They join the content hash (CANON_VERSION 3), SQLite and D1
(questions.taxonomy_subject, question_concepts), and the site corpus, and
vault build writes the labels the site needs.
Removes content-authoring scripts, figure and data generators, and their working notes, which are not needed to build, check, or release the books and sites. The vault CLI drops the matching subcommands and doctor check; docs, comments, ignore rules, and the codegen hash are updated to match what remains.
- StaffML book references use numbered chapter URLs; the topic map and
tests follow. The draft validator writes its scorecard to the pipeline
directory, docs point at vault/README.md, and the architecture doc
drops dated session notes.
- Instructor course map and scale syllabus link to labs that exist, and
model compression links use the numbered chapter URL.
- Labs: remove per-lab planning files and the test that required them.
- MLPerf EDU: formatting and workflow fixes, dataset URLs, portable
evidence report paths, ignored run reports, and removed working notes.
- Kits social image points at an existing cover; slide deck header
trimmed; unused images, wheels, scripts, and duplicate modules removed.
- TinyTorch installer cleanup list and gitignore trimmed; retired build
scripts removed.
- kits, labs, and the MLSys·im docs: _quarto.yml was a symlink to
config/_quarto-html.yml. The website config now lives at _quarto.yml.
The kits PDF build copies config/_quarto-pdf.yml over it for the render
and restores it afterwards, and the CI PDF job copies instead of linking.
- Book builds: the binder CLI wrote books/_quarto.yml and books/index.qmd
as symlinks. ConfigManager.activate_config (formerly setup_symlink) now
writes copies, and the first line of _quarto.yml names its source, which
status, doctor, clean, and save_latex_log read. The mapped chapter
harness snapshots and restores both files by content. The Linux container
job copies index.qmd, as the Windows job already did.
- StaffML: releases/latest was a symlink to the current release directory.
releases/latest.txt now holds the version and is replaced atomically on
publish; vault doctor, release publish, and the paper macro script read
it.
A worktree that still holds the old books/_quarto.yml or index.qmd symlink
is safe: the next build removes the link before writing, so the source
config and volume index are never overwritten through it.
Moving the sources under books/ and the toolchain under bindery/ rewrote
slash-style paths but missed paths built with Path and os.path.join, test
fixtures, and a few regexes. Several checks and scripts have been pointing
at directories that no longer exist, most of them silently.
- binder check index scanned book/ and found nothing; it now scans the 151
book sources. The callout reference check read its prefix table from a
toolchain quarto/config folder that does not exist and reported itself
inert; it now reads books/config and scans 140 files. Both pass.
- The registry checks, the audit command, bib update, the EPUB baseline,
the table layout audit, release, and doctor looked for book/tools,
book/tests, book/binder, or book/.layout.
- Five call sites computed the repository root as books/../.., one level
above the repository.
- The StaffML vault looked for chapter sources under book/quarto/contents,
so its book cross-references silently fell back to slug titles and skipped
link-checking.
- The VS Code extension located the repository by book/binder and read
chapter lists from book/quarto/contents.
- Tests loaded Lua filters, build scripts, the LaTeX header, and a patched
module from paths that do not exist. The binder suite went from 12
collection errors, with no test running, to 342 passed; the 8 that still
fail are the mlsysim import path and the URL flattening script, both
fixed separately.
- test_content_structure and test_volume_bibliography matched config
entries by a contents/ prefix, so "every config path exists" checked
nothing. They now match the flat books/ paths. The structure generator
wrote books/STRUCTURE.md rather than books/shared/STRUCTURE.md, which is
regenerated here; the committed copy was missing the Volume III and IV
front matter the PDF configs list.
- extract_figures.py fell back to a hard-coded personal path.
- Five documents linked to ../book/README.md or ../quarto/config/, and the
SocratiQ bundle README described a symlink that does not exist.
A legacy quizzes script under _legacy/ still names a cwd-relative
quarto/config and is left as it is.
interviews/ held StaffML and nothing else, so the directory now carries the
product's name. Three names inside it repeated that prefix, and now line up
with vault/, vault-cli/, and paper/:
interviews/staffml/ -> staffml/app/
interviews/staffml-vault-worker/ -> staffml/vault-worker/
interviews/staffml-vault-types/ -> staffml/vault-types/
Every repository path moved with it: workflow paths: triggers and variable
fallbacks, pre-commit hooks, .gitignore, contributor tooling, vault-cli path
joins, the app's tsconfig mapping to the shared types, and the docs.
Left alone on purpose:
- the published /interviews/ redirect to /staffml/ and the gh-pages subsite
lists that protect it; those are URLs, not repository paths
- package and worker names (staffml-vault-worker, @staffml/vault-types) and
the vitest job name, which branch protection may key on
- question YAMLs and generated track JSON, which are content-addressed.
Three questions carry an older corruption, "recall @interviews/vault/
questions/...", that predates this change
- the applied D1 migration 0001 and the vault CHANGELOG, which are history
codegen-hashes.txt is re-recorded; the three pinned artifacts changed only in
path strings. vault-cli: 97 passed, codegen --check clean.
interviewer.py's __main__ loader was one parent short and pointed at a
corpus.json that never existed; it now resolves staffml/app/src/data/
corpus.json. tools/validate_playbook.py and generate_interview_flashcard.py
still expect per-track directories that no longer exist; they point at the
new root but remain dead.
The GitHub repository variables STAFFML_ROOT, VAULT_DIR, and VAULT_CLI_DIR
still name the old paths and must change when this reaches origin.
Vijay Janapa Reddi