Affine Fusion Control

Affine Fusion Control — the local, exact, floating-point-free disruption court

Page class: STUDY (app realization of Study 33). Every figure below is produced by a program in reproduce/ or by the app itself, dated where it is a timing. Grades per the ontology: MEASURED, PROJECTION, REPORTED, NOT_KNOWN. Read the limits section first — it says exactly what this is and is not.

Public release: Affine Fusion Control — the plasma operating-point verdict a safety authority can re-derive by hand — the measured, claim-graded case, every figure re-derivable from source.

---

What this is, in one honest paragraph

A fully-compiled Swift 6.4 macOS app that grades a fusion reactor's telemetry against a frozen, exact-integer disruption law — no floating point, no network, no model, no GPU on the decision path, no slow start. It runs 65,536 agents on a 1 kHz control clock on one laptop. It is not connected to a reactor and holds no real machine data; it grades presented telemetry and presented operating points. What it demonstrates is an architecture: a control verdict that is an exact integer a third party can re-derive, in a domain the industry currently settles with floating-point surrogates that return a confident number for every input including the ones they were never trained on.

What you can hold

The app, running — 65,536 agents at 1 kHz, five panels, both courts live:

• THE LAW, FROZEN — the six constants, rendered from the law's own source, not transcribed into the view. The panel's own footer says so. This is the app's claim to being a court rather than a dashboard.

• VERDICT WALL — every agent as one cell, coloured by terminal (NOMINAL green, MITIGATE amber, REFUSED red). The instrument is visibly discriminating — four distinct verdicts on screen at once, not one alarm:

  

  In this shot: NOMINAL 54,272 · MITIGATE 8,192 · REFUSED-envelope 2,048 · REFUSED-malformed 1,024 (sums to 65,536). The regular bands are the demo's channel classes; a real feed produces whatever the sensors do.

• OPERATING-POINT COURT — the second law, live on five real published machines, graded every tick through the same FusionOperatingPointLaw the reproduce scripts call. Each machine sits at its own fraction of its own Greenwald limit; a shared drift walks them across the 0.85 line, so a single frame shows genuine discrimination rather than a synchronized alarm:

  

  In this frame: ITER 72 % WIN · SPARC 80 % WIN · JET 86 % MISS · DIII-D 94 % MISS · W7-X N/A. The stellarator carries no plasma current, so it never enters the density court at all — NOT_APPLICABLE_NO_PLASMA_CURRENT, permanently, which is the thesis in one tile (a trained surrogate divides by ~0 and extrapolates; this court refuses to invent an n_G that does not exist).

• CADENCE — the panel where the app tells on itself: the latency histogram, skippedTicks, p50/p99. Red the instant a deadline is missed.

• CHANNEL SCOPE — one channel's window against the ±6000 envelope and the growth trigger.

The colours match the browser client in clients/fusion-court/ — the two live in different languages (Swift RGB, HTML hex), so a gate (Tools/palette-parity.sh, PALETTE_PARITY_PROVEN) checks every build that they still agree. The claim that they cannot drift is enforced, not remembered.

---

The law — three ways it can refuse, and why refusal is the point

The disruption law is abs, subtract, compare, and a run counter — no multiply, no divide, no float. It returns one of four terminals, and the two REFUSALS are the whole difference from a trained surrogate:

terminal	meaning
NOMINAL	inside the envelope, no growing mode
MITIGATE	a mode grew past the trigger for the required consecutive windows
REFUSED_OUT_OF_ENVELOPE	the signal left the range the law was frozen against — the court does not rule on data it never saw
REFUSED_MALFORMED	the sample is outside the ADC domain entirely — not admissible data

A statistical interpolator returns a plausible number for every input, including inputs outside anything in its training set — and cannot tell you which of its answers it was entitled to give. This law returns a verdict only where it has one, and names its refusal everywhere else. That is an invariant approach to "what will happen," versus a model that can only replay what it was shown.

The verdict law has one home (app/FusionCourt/Sources/FusionLaw/) — enforced, because it was once forked three ways and the copies disagreed on real inputs (a refusal had silently become a pass). That correction is documented in full on Study 33.

The operating-point court — three exact inequalities, five terminals, all reachable

Beyond the streaming disruption law, the substrate carries an exact operating-point court: is a presented reactor configuration inside its stability limits?

• Greenwald density: 100·ne·a²·π_num < 85·Ip·1e6·π_den
• Troyon beta: β_N ×1000 ≤ 2380
• q_min: q_min ×1000 ≥ 2000

All in Int128 (Swift 6.4 native) — the Greenwald product overflows Int64 at a 20 m minor radius and fits Int128 with room to spare, so no arbitrary-precision fallback is needed. π is irrational and this is a court: the caller may declare its own π_num/π_den for an exact verdict, or the law evaluates at both 333/106 and 355/113 and, when they agree, reports the verdict as π-independent — a stronger claim than any float computation can make.

All five terminals are reachable, proven not asserted (reproduce/fusion-operating-court.swift, 11 tests):

OPERATING POINT                       VERDICT                           BINDS
ITER 15MA baseline (nominal)          WIN                               -
ITER pushed over Greenwald density    MISS                              greenwald
ITER over Troyon beta limit           MISS                              troyon
ITER under q_min floor                MISS                              qMin
W7-X stellarator (currentless, Ip=0)  NOT_APPLICABLE_NO_PLASMA_CURRENT  -
on the Greenwald line (pi-bracket)    NOT_MEASURED_PI_BRACKET           -
malformed submission (Ip < 0)         REFUSED_NONPHYSICAL               -
COURT TERMINALS REACHED: 5 of 5

The stellarator row is the thesis in one line. A currentless machine has no plasma current, so the Greenwald limit — which is a statement about a current-carrying plasma — is undeterminable. The court returns NOT_APPLICABLE and names the open branch rather than dividing by ~0 and extrapolating a confident answer, which is exactly what a surrogate trained on tokamaks would do at a stellarator.

Any reactor topology — the law never changes, only the layout does

reproduce/fusion-topology-agnostic.swift proves it at the law level: tokamak, stellarator, and spheromak are three genuinely different layouts —

topology	field periods	toroidal circuit	nodes	edges
tokamak	1	yes	512	1,504
stellarator	5	yes	2,560	7,520
spheromak	1	no	481	960

— and the same operating envelope grades to byte-identical verdict signatures across all three. The verdict moves only when a physics input changes (Ip 15 MA → 0 flips WIN → NOT_APPLICABLE), never on the topology descriptor. Negative control, and it can fail: the spheromak, having no toroidal circuit, emits zero toroidal-closure edges — a toroidal disagreement is not even representable in its layout, and the check prints FAIL if that count is ever > 0. That is "we run on any reactor topology" made checkable rather than asserted.

The court on real published machines

The operating-point court is not only exercised on synthetic points — it grades real machine geometries whose parameters are public (reproduce/fusion-real-machines.swift). Using only each machine's published plasma current and minor radius, it computes that machine's own Greenwald density limit and places the 0.85 boundary:

machine	Ip	a	Greenwald limit (computed)	at 0.80×	at 0.90×
ITER	15 MA	2.0 m	1.19×10²⁰ m⁻³	WIN	MISS (greenwald)
SPARC	8.7 MA	0.57 m	8.52×10²⁰ m⁻³	WIN	MISS (greenwald)
JET	4.8 MA	1.25 m	0.98×10²⁰ m⁻³	WIN	MISS (greenwald)
DIII-D	2.0 MA	0.67 m	1.42×10²⁰ m⁻³	WIN	MISS (greenwald)
W7-X (stellarator)	0 (currentless)	0.53 m	—	NOT_APPLICABLE	—

(REPORTED — Ip and a from iter.org, Creely et al. 2020, EUROfusion, General Atomics, IPP Greifswald.) The computed limits match the published Greenwald densities for each machine — SPARC's is high precisely because its minor radius is small and its current large, exactly as the literature reports. The court reproduces real machine physics from two public numbers, places the density boundary correctly on every one, and refuses W7-X because a currentless machine has no Greenwald limit to place. That is the difference between an exact law and a surrogate: the law is right for a reason a physicist can check, not confident for a reason no one can.

The GPU crossover — measured, and it goes the way the founder said

The honest question: does this law belong on a GPU? Measured on an M4 Max (40 GPU cores), --selftest-crossover, both sides bit-exact to the CPU golden:

agents	CPU (12 core)	GPU (round-trip)	winner
4,096	232 µs	1,037 µs	CPU
65,536	1,966 µs	4,349 µs	CPU
1,048,576	29,301 µs	58,969 µs	CPU

(One representative run. The winner and bit-exact parity are what hold on any machine and are pinned in the test suite; the exact microseconds vary run to run and are not a claim.)

The GPU never wins up to a million agents, ~2× slower everywhere, and bit-for-bit correct throughout. A kernel launch is a fixed ~1 ms; the law has no multiply for a GPU's float lanes to accelerate. "We needed it for floating point, but we are not that now" — measured, not assumed. The full account, including the two broken benchmarks that flattered the GPU before the instrument was fixed, is in the app's CROSSOVER.md.

The limits — stated plainly, because a court states its own

• No reactor. No real machine data. Every trace here is synthetic; every operating point is presented. The app grades telemetry; it does not read a tokamak. (NOT a measurement of any device.)
• Determinism is checkable, not asserted. A single sha256 over ~3,000 verdicts (reproduce/fusion-determinism-digest.swift) is identical on every machine — f49b576e… — because every verdict is an exact integer comparison with no floating point anywhere. Run it on your hardware; a different digest means the law diverged, which is the failure this architecture exists to make impossible.
• Timings are dated, not constant. The latency and crossover numbers are this machine on this date. The harness pins the claims that must hold on any machine (5-of-5 terminals, byte-identical cross-topology verdicts, bit-exact GPU parity), never the wall-clock microseconds.
• The disruption thresholds are illustrative and frozen for demonstration. A real deployment re-freezes them against its own device before grading anything. What is not illustrative is the architecture: exact, re-derivable, refusing where it has no answer.
• Mac-only. This is the local app, deliberately not the build that ships to the Affine.Earth cells.

Reproduce every claim

New to this: the fusion researcher's guide — clone, build, run the app, and re-derive every number above, with the exact commands and what each one proves.

---

Related: Study 33 — the fusion control verdict court · Fusion researcher's guide · The replacement grade · The ontology