Public release page: Affine Fusion Control — the verdict a safety authority can re-derive by hand
A PUBLIC CASE page for the fusion study, built by a 5-phase multi-agent pass (harvest 170
figures across every source page -> three angle-drafts -> synthesis -> four adversarial
reviewers -> finalize) and then ground-truth-verified by hand before publishing:
- Every exact Greenwald integer re-run from source: ITER 1193661, SPARC 8523532, JET 977847,
DIII-D 1418177, W7-X NOT_APPLICABLE — matches the program output exactly.
- Every pinned figure confirmed in reproduce/validate.sh (digest f49b576e, 7016/140us/208/960,
5-of-5 terminals, byte-identical topology, HEADROOM/DETERMINISTIC booleans, PENDING marker,
beta_normalized=1.8 attribution) and the determinism corpus re-run: 400 + 2592 = 2992 verdicts.
- Clock selftest re-run: ticks=9999 skipped=0 gaps=0. CROSSOVER rows and metallib digest bd055be2
confirmed. Every cited file confirmed to exist; three overshooting line-ranges corrected.
- Conforms to the public-page discipline: no private .swift:NN source-line refs (the harness
caught them); provenance is filenames + the re-derivation commands in the reproduce block.
Grades every claim (VERIFIED/MEASURED/REPORTED/NOT_KNOWN), states its own limits (GPU loses ~2x,
tau_E NOT_KNOWN, operates no reactor, study is PENDING, Mac-only, no LICENSE), embeds the three
served screenshots, and links the study / guide / Study 33 / Replacement-Grade / Ontology.
Linked from the sidebar and a banner on the study page. Harness: 136 passed / 0 failed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Live operating-point court in the app — five real machines, WIN flips to MISS on screen
The running app showed the STREAMING law (verdict wall) but not the OPERATING-POINT
law. Now both courts are live in one window. A fifth panel grades five real published
machines every tick through the same FusionOperatingPointLaw the reproduce scripts
call — the UI computes no verdict, it reads the snapshot the control thread publishes.
Each machine sits at its own fraction of its OWN Greenwald limit; a shared integer
triangle drift walks them across the 0.85 operational margin. Because their bases
differ (ITER 74, SPARC 82, JET 88, DIII-D 96) they cross at different phases, so a
single frame shows genuine discrimination — ITER 72% WIN, SPARC 80% WIN, JET 86%
MISS, DIII-D 94% MISS — never a synchronized all-green/all-red. W7-X carries no
plasma current and stays NOT_APPLICABLE permanently: the thesis in one tile.
- New Panels/OperatingPointPanel.swift; ControlHost grades the five machines;
MachineVerdict rides the triple-buffered snapshot.
- Sweep math verified against the law via --grade: 80% -> WIN, 90% -> MISS
(binding greenwald), ip=0 -> NOT_APPLICABLE_NO_PLASMA_CURRENT.
- New captured screenshots: full app (five panels) + operating-point close-up.
- Study page: "five panels, both courts live", operating-point image embedded,
the ITER/SPARC/JET/DIII-D/W7-X frame described.
verify-all: VERIFY_ALL_PROVEN (7 gates, 0 warnings, GPU parity bit-exact).
Studies harness: 136 passed / 0 failed. no-float-outside-render still green —
the panel's CGFloat use is inside Canvas closures only; the sweep is pure integer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Mirror: cross-machine determinism digest
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Mirror: --grade mode + guide/study polish
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Mirror: real-machine grading, metallib pin, wall close-up
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Mirror: fusion study, researcher's guide, screenshots, palette gate
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>