docs(paper): research paper + presenter pack, with six figures (supersedes #14)

[ksdisch]

Write-up of the closed v1 chain (M0–M4) from the 18 committed result JSONs in results/. Review only — please do not merge without your sign-off.

Supersedes #14. That PR was tables-only by rule, not by choice: the research-paper skill forbade figures outright. It has since been fixed (claude-config d95f620), so this run let the recorded data earn its figures. I have not touched #14 — its disposition is your call.

What's new vs #14

Six figures, plus a committed script that draws them:

	Figure	The shape a table hides
1	gate contrasts	all five pre-committed contrasts and their interval widths on one axis
2	window sweep	the late cliff on a floor, and 0.5B's visibly raised floor
3	dose grid	a dimmer not a step; binary and mass channels falling together
4	the 12 × 12 matrix	the dark diagonal on a near-white grid — what M3's brief calls "the killer figure", finally drawn. silver's fragile column vs its harmless row is visible at a glance
5	M4 floor reads	why the verdict says AS-SCORED ONLY: one lower bound clears 0.5, another doesn't
6	collateral asymmetry	primes cluster tight and high; probes are bimodal with a tail to 0.5

Figure 4 replaces the brief's 12 × 12 ASCII grid entirely — each cell is annotated with its recorded hits/n, so the figure is its own table view.

What the figure script may and may not compute

docs/paper/figures.py is deterministic and headless, reads only the committed JSONs, and writes only the six PNGs beside it. It computes nothing beyond hits / n where a file records the pair rather than the rate; every interval drawn is a recorded wilson_95 or newcombe_* endpoint. It never smooths, interpolates, fits, re-bins, or pools arms the runners did not pool. Specifically:

• No line joins the marks on the sweep or dose figures — the values between two grid points were never measured.
• The interpolated half-mute λs (0.23 / 0.29 / 0.36) are quoted in prose, labelled as interpolations, and plotted nowhere.
• It prints every plotted number with its source key, so the figures check against the paper's tables without opening a PNG.
• Re-running produces byte-identical PNGs (verified).
• matplotlib is injected per-run — uv run --with matplotlib docs/paper/figures.py. pyproject.toml is unchanged.

Constraints honoured

• No new measurements. No runner, gate, ablation or anchor script was executed; nothing in results/ was written or overwritten. uv run pytest was not run either — the 396-test count is quoted from HANDOFF.md rather than claimed as re-verified.
• No fabricated numbers. Every load-bearing statistic was re-checked by parsing its named source file (a throwaway script over all 18 JSONs; all checks pass). Two claims that looked derived — "126 of 132 ordered pairs" and "633 of the 2,340" — are recorded verbatim in M3-BRIEF.md and DECISIONS.md.
• Honesty framing intact. Every null, the AS-SCORED ONLY qualifier inside the verdict string, both re-attributions (silver's column vs row; category-block composition), the M1 coverage bound, and the full owned-deviations ledger are carried through. CLAUDE.md's stale "v1 chain (M0–M3) is closed" was overridden by the briefs and result files — the paper covers M4.

Two things for your review

1. Length. The paper's prose runs ~6,600 words against the skill's ~3,000–5,000 guideline. I cut what didn't cost content; the remainder is five milestones, six figure captions carrying required disclosure, and an owned-deviations ledger I'm not willing to thin. Say the word if you want it inside the band and I'll propose specific cuts.
2. Figure 6's denominators differ per mark (12 deletions × a concept's gated items, or 11 × for a subset concept). Stated in the caption and on every labelled mark — worth a look to confirm it reads clearly.

🤖 Generated with Claude Code

https://claude.ai/code/session_016PZepEZ6FAuLU51giFk31i