handoff: the resume point is publication now -- gate, article, email, arXiv

Every surface serves the same card, so the next session's work is the writeup as a Hub
community article, the email after it, and arXiv beside both. The header states the
2026-09-05 state (lock, D074 review, D075 review pass feeding no published number, the
D074 collision and how to avoid the next one), the gate lists the four digests that must
agree plus the deploy-vs-commit check the digest cannot see, and each step names what a
session may do and where it must stop. The Wave 2 'next action' section is retitled as
historical so nobody resumes from it.

Author: caiotheodoro

docs/HANDOFF.md

@@ -3,7 +3,126 @@
 The resume point. A fresh session should be able to continue from this file without
 re-deriving anything.
 
-**Last updated: 2026-09-04 — (D071/D072) H2 is measured and closed.** The raw-corpus adapter
+**Last updated: 2026-09-05 — delivery is current on every surface; what is left is
+publication.** Main is green (a `constraints.txt` lock pins CI's install; the Space's stats
+snapshot no longer fails on a test count or a float ulp). `MEASUREMENT_CARD.json`
+carries 18 claims and one unmet item (Result 2), verdict `NOT_VERIFIED`; the Hub dataset,
+the Hub model and the Space all serve that exact card (digest `3aacbe46…`), and the dataset's
+`results/` now ships both `h2_design_effect` files the H2 claim cites. D074 is the independent
+review of D071/D072 `WAVES.md` required: the verdict stands, every number reproduced from the
+response records, and the pre-registration turned out to state H2's threshold two ways (design
+effect ≥ 2, and "twice the width", which is 4 under its own definition; the result clears
+neither). D075 opened a third, salted, blind labelling pass over the 18 rater/judge
+disagreements plus an equal control arm; it is a separate `review` pass and **feeds no published
+number** — folding it needs its own entry with before and after. Two sessions wrote a D074
+within minutes of each other on 2026-09-05; the later one is now D075. That will recur while two
+sessions share one append-only file: check `grep -c '^## D0' docs/DECISIONS.md` against the
+last number before appending.
+
+**Still open, in order:** publication (the section below); the D075 fold decision, once the
+review pass is complete and `scripts/review_labels_cli.py report` says whether the controls
+moved; a second rater on `G200-ego` (`RED-TEAM.md` A1, still the first weakness on the front
+page, ~35 minutes of someone else's time); `judge_weights_rev` recorded into the result JSON on
+the next real judge run (D074 finding 4); `REVIEW.md` R2, the pseudo-cluster proxy on Build AI's
+own frames, now calibratable against real `worker_id`s (`RED-TEAM.md` A13); Result 2, blocked
+on two reasons the adapter did not touch.
+
+## Next session: publication
+
+Three steps. One is the critical path, one follows it, one runs beside it. `docs/private/`
+holds the outreach sequence and the email draft; it is gitignored and is never quoted here or in
+anything public (`AGENTS.md` rule 5).
+
+### Gate, before anything goes public
+
+Run all of it; publish only if every line is clean.
+
+```
+git status --short && git rev-list --left-right --count origin/main...main   # clean, 0 0
+gh run list --limit 1                                                          # ✓ on main
+make validate                                                                  # tests, mypy, lint
+make card && git diff --exit-code MEASUREMENT_CARD.json                        # committed card == fresh
+python3 -c "import json;print(json.load(open('MEASUREMENT_CARD.json'))['content_digest'])"
+python3 -c "import json;print(json.load(open('space/public/data/stats.json'))['generated_from']['card_digest'])"
+curl -sL https://caiotheodoro-vernier.static.hf.space/data/stats.json | python3 -c "import sys,json;print(json.load(sys.stdin)['generated_from']['card_digest'])"
+curl -sL https://huggingface.co/datasets/caiotheodoro/vernier/raw/main/MEASUREMENT_CARD.json | python3 -c "import sys,json;print(json.load(sys.stdin)['content_digest'])"
+```
+
+The four digests must be identical. The digest does not see UI-only changes, so also compare
+`git log -1 --format='%h %ci' -- space/` with the Space repo's `lastModified`
+(`curl -s https://huggingface.co/api/spaces/caiotheodoro/vernier`); as of this header the last
+Space commit (f5ef78a, 23:31 local) is four minutes newer than the last upload. If either is
+behind, `make space-data && make space`
+then `hf upload caiotheodoro/vernier space/dist . --type space`; if the dataset's is behind,
+`make hf-dataset` then `hf upload caiotheodoro/vernier hf/dataset . --type dataset` (both need
+`HF_TOKEN` from `.env`; `hf` is the 1.30 CLI at `~/.local/bin/hf`). If the other session has
+folded D075 into the primary labels in the meantime, every number downstream moves: run
+`make agreement`, `make distil`, `make card`, `make space-data`, republish, and only then
+continue — a writeup that cites numbers a later commit restates is the exact failure this
+project audits in its subject.
+
+Then the writeup's own rule: every figure in `docs/WRITEUP.md` is copied from the card or a
+`data/*.json` file, and the card wins on disagreement. Check the ones that have moved most
+recently by hand: the H2 table (1.25 / 1.62 / 1.27 and 1.31 / 1.66 / 1.29), the 2,144-vs-2,153
+worker count, the decision range (D001–D075), and the two cost figures ($17.63 from the two E2
+result files; ~$7.50 from D072, the one number in the writeup that cites a decision rather than a
+data file, because no result file carries it).
+
+### 1. The writeup, as a Hugging Face community article — critical path
+
+There is no CLI for this; it is the web editor at https://huggingface.co/new-blog, published
+under `caiotheodoro`. Body is `docs/WRITEUP.md` with two edits: drop the italic *"Draft for the
+Hugging Face community blog"* paragraph at the top (keep the sentence about where the numbers
+come from, reworded as a statement, not a draft note), and turn the bare repository paths
+(`docs/PRE-REGISTRATION.md`, `data/e2_full_n10000.json#H1`, …) into links to
+https://github.com/caiotheodoro/vernier at `main`, since on the Hub they resolve to nothing. The
+H1 title is the article title. Tags: `egocentric`, `dataset-audit`, `llm-as-judge`,
+`evaluation`, `build-ai`. Link the collection
+(https://huggingface.co/collections/caiotheodoro/vernier-same-judge-same-6pp-on-2-hands-both-releases-6a9b1bc9fac182a7a2d9c997)
+in the "Where everything is" list; if the editor offers to attach datasets and models, attach
+`caiotheodoro/vernier`, `caiotheodoro/vernier-rung1-probe`, and the two `builddotai/*-Evaluation`
+datasets so the article surfaces on their pages, which is the whole point of publishing on the
+Hub rather than anywhere else.
+
+After it is live: a `DECISIONS.md` entry (D076 unless the other session has taken it) recording
+the URL and the date; the URL added to `README.md`'s map, `llms.txt`, the dataset card
+(`scripts/export_hf_dataset.py`, then `make hf-dataset` and re-upload), and the collection.
+Nothing in the article is edited after that without a dated note in the same entry — it is a
+published result now, and the project's own standard for Build AI's card applies to it.
+
+### 3. The email — after 1, never before
+
+`docs/private/OUTREACH.md` is the sequence and the reasons; `docs/private/EMAIL-DRAFT.md` is the
+text. Five lines, link first. What a session can do: check every link in the draft resolves,
+check every figure in it against the card, confirm the article is live, and stop. Sending is the
+user's act, from the user's account, and is not automated. Two things the draft must reflect as
+of now: the article URL (it is the first link), and `docs/ETHICS.md` §4 as rewritten in D073 —
+24 downscaled stills are republished in the Space and come down on request; that is the one
+sentence in the email that is a courtesy to them rather than a result. D009's exception (private
+disclosure first, only for a live privacy or security exposure) has not been triggered; nothing
+found in this project is one. Record the send as a one-line `DECISIONS.md` entry with the date
+and nothing else — the content stays private.
+
+### 2. arXiv — beside the other two, gating nothing
+
+cs.CV, single author, no affiliation ("Independent"). An unaffiliated first-time submitter needs
+an endorser; ask with the article already public, since the ask is then "read this" rather than
+"trust me" (`docs/private/OUTREACH.md` says why in more words). Do not name a candidate in any
+public file. Build the submission from the writeup, not from a rewrite:
+
+```
+pandoc -s docs/WRITEUP.md -o arxiv/vernier.tex        # TeX source, which arXiv requires
+                                                       # for anything TeX-produced
+```
+
+Then a title, an abstract (the first two paragraphs of the writeup, tightened), CC BY 4.0, and
+cross-list `stat.AP` if it is offered. The arXiv version cites the article URL and the Git
+commit it was built from. Submit after 1 is live so v1 links to it; submit before 3 only if the
+endorsement arrives first — the email should carry the arXiv link if one exists, and does not
+wait for it.
+
+
+**Prior header — (D071/D072) H2 is measured and closed.** The raw-corpus adapter
 D065 declined to size turned out to be small: ffmpeg's `subfile` protocol extracts one frame
 from an h265 mp4 inside a tar over HTTP range requests, so no new dependency was needed, and a
 tar header is 512 bytes at a computable offset, so `scripts/build_corpus_manifest.py` indexed
@@ -252,7 +371,7 @@ A live E2 or E5 run now also leaves per-frame records behind, next to its checkp
 per-file negation. Forward-only -- the committed aggregates predate this and have no jsonl.
 `docs/DECISIONS.md` D069.
 
-## The next action
+## The next action, as it stood in Wave 2 (historical; the current one is "Next session: publication" above)
 
 **The Qwen3-VL judge is deployed and fully smoke-tested live end to end, image + logprobs
 included — real correct output.** `cloud/modal_qwen3vl.py` is live on Modal