Study 34 Observer Invariant Verdict

Study 34 — The Observer-Invariant Verdict: why an exact law, not a bigger computer, makes the safety verdict the same for everyone

Page class: STUDY / PUBLIC CASE. Every figure here is printed by a program in reproduce/ and, where marked VERIFIED, pinned in reproduce/validate.sh. Grades follow the ontology: VERIFIED, MEASURED, REPORTED, NOT_KNOWN, and — for readings that interpret the measured rows rather than being measurements — ARGUMENT, never smuggled in as a measurement. Read the limits section; it says plainly what is proven and what is not.

One operating point, two answers — the exact vQbit court refuses honestly where a floating-point surrogate contradicts itself

We need fusion

Fusion is the energy the planet is short of hope for. The plasma limits that decide a safe discharge from a disruption have been written down for decades; what has been missing is underneath them — a control verdict the world can actually trust and share: a safety decision that is the same for the engineer in Princeton, the regulator in Cadarache, and a student in Lagos, and that any of them can re-derive without being given permission. Where that verdict is carried by a floating-point machine-learning surrogate, this study measures the cost — and demonstrates the alternative that does not carry it. REPORTED (the surrogate framing is the accepted state of the art; the cost below is measured).

Affine Fusion computes the verdict in exact integers

The Affine Fusion Control court decides a plasma operating point with exact integer arithmetic on the vQbit substrate — the substrate's primitive is a packed rational carried unevaluated, so there is no floating point anywhere upstream of a verdict. That is not a stylistic choice. It is the reason the verdict is observer-invariant: the same inputs produce the same integer on every machine, forever, and the answer can be checked by hand.

The claim of this study, stated once and then proven:

A safety verdict must be the same for everyone who checks it. The floating-point way of computing this exact verdict is not — it changes with the precision and with the value of π the implementation happened to round to — while the exact vQbit court returns one verdict every machine reproduces, or a named refusal. For the layer that licenses a reactor and holds an operator accountable, exactness is not an optimisation; it is the requirement. — ARGUMENT, read from the measured rows below.

What the two programs measure

Two programs carry the proof, and both are replayable with no account and no key.

1 — One corpus, byte-identical on every machine. reproduce/fusion-determinism-digest.swift grades a fixed corpus of 2,992 verdicts (400 streaming traces plus a 2,592-point operating-point grid) and prints one sha256 over all of them. It is the same string on any hardware, because every verdict is an exact integer comparison:

f49b576e073835bcab17bee10fe0eee1938774643d900b8ffe1a583b159ab3d7

VERIFIED — pinned in reproduce/validate.sh. Run it on your machine; a different digest would mean the law diverged, which is the failure this architecture makes impossible.

2 — The floating-point verdict of the same law is observer-dependent. reproduce/fusion-exact-vs-float.swift grades the Greenwald density limit around the 0.85 safety threshold for four real published machines, the exact way and the floating-point way, and counts where they part company. The exact court commits to π only as an interval, 333/106 < π < 355/113, and returns a verdict only when it holds across the whole interval; where it does not, it returns NOT_MEASURED_PI_BRACKET — the answer sits closer to the limit than π has been pinned. A float implementation cannot do that: it must substitute one rounded π and return a confident answer.

machine	operating points the exact court refuses (π-bracket)	two defensible rational π's that contradict
ITER	27	27
SPARC	61	61
JET	22	22
DIII-D	32	32
total	142	142

MEASURED. The exhibit, one operating point with three answers — ITER, exactly at its 0.85 Greenwald safety threshold (n_e = 1,014,613 × 10¹⁴ m⁻³):

   pi = 355/113  ->  MISS   (disruption-limit exceeded)
   pi = 333/106  ->  WIN    (inside the limit)
   the exact vQbit court  ->  NOT_MEASURED_PI_BRACKET   (not to this precision)

Read the float32 flips = 0 column precisely: this is not a claim that floating point is imprecise. Single precision resolves this clean inequality fine, and the divergence is not driven by hardware — it is driven by the value of π the implementation rounds to (355/113 versus 333/106). The failure is deeper than accuracy: the float verdict is not invariant. Two people, both computing the identical safety test with defensible constants, reach opposite conclusions about the same plasma, and neither can show the other why. The exact court declines to pick, on all 142, and names why: NOT_MEASURED_PI_BRACKET.

Why an observer-invariant verdict matters

ARGUMENT — the three readings below interpret the measured rows above; they are not themselves measurements.

For a researcher: a verdict that is byte-identical across machines is composable. You can build a fleet, a cross-machine database, a decade of shots on it and the conclusions still reconcile — because there is one answer, not one-per-platform. A float verdict accumulates unreconcilable disagreement at the rate you scale it (the over-scaling law); no larger GPU cluster and no quantum sampler fixes that, because the problem is the kind of arithmetic, not the amount of it.
For a politician or a regulator: a licensed fusion facility has to put its machine-protection logic through a safety case. "The model was confident" is not a safety case; "this verdict is an exact integer you can re-derive after an incident" is. The exact court makes the safety authority — not the vendor — the party who can check the verdict. That is accountability that exists in the mathematics, not in a promise.
For anyone looking for hope: the barrier this study measures is not a missing machine and not a bigger budget — it is a control layer, and the exact one already exists and runs on a laptop (see Affine Fusion Control and Study 33), with no network and no accelerator, and you can run it yourself and get the same integer everyone else gets. Fusion being hard is not the same as fusion being out of reach. The part that was software is answerable, and here is the answer, in the open. REPORTED (the throughput figures are measured on the parent pages, not on this one).

The limits — a court states its own

The π-bracket band is narrow — 0.0007%–0.0023% of the Greenwald limit across the four machines (0.0023% at ITER, the exhibit machine) — and sits below plasma-density measurement uncertainty (a few %). So this is a demonstration of kind, not a claim that ITER's disruptions turn on the sixth digit of π. A tighter bracket would shrink the band; removing π entirely — the affine density invariant Φ_q/det(Λ) in Study 33 — removes it. The guarantee either way is interval-valid, or refuse, and the affine form has no interval to bracket. The stakes are compositional — a verdict that is observer-dependent at any width cannot be the shared, auditable authority a planet-scale safety case needs. ARGUMENT.
This court operates no reactor. Every trace is synthetic and every operating point is presented; there is no actuation and no diagnostic feed. The only real numbers are the four machines' published current and minor radius. REPORTED.
The bracket-width figure was itself corrected in the making of this study. The law's comment and an earlier page called the 333/106 … 355/113 bracket "~2.7×10⁻⁷ wide." It is not: 355/113 − 333/106 = 1/11978 ≈ 8.35×10⁻⁵. The 2.7×10⁻⁷ is the error of 355/113 alone, not the width of the interval. The integer identity is printed by the proof program and pinned in the harness. VERIFIED.
The parent study is PENDING. Study 33 carries STUDY33_FUSION_CONTROL_VERDICT_PENDING: a law graded on synthetic traces, with no device behind it, has not earned more. This study proves a property of the verdict, not a fusion result. VERIFIED.

Rights — source-available, not open-source

This study, its programs, and the Affine Fusion Control app are published source-available: the source is visible so that anyone can inspect it and re-derive every figure on this page. That visibility grants no rights. The repository carries no LICENSE, which under default copyright means all rights are reserved. No right is given or intended — by this publication, or by the source being readable — to use, run, or deploy it for any purpose other than re-deriving the figures published here, nor to reproduce beyond reading, modify, or build derivative works from any of it. Any use beyond reading and re-deriving the published results requires a separate written licensing agreement with the authors. Re-deriving the numbers is exactly the point; anything past that is by agreement only.

Reproduce it yourself

No account. No key. No permission needed to read it and re-derive the figures — that is what "re-derivable by anyone" means, and it is distinct from a right to use it for anything else (see Rights, above).

git clone https://github.com/gaiaftcl-sudo/uum8dSolarResearch.git
cd uum8dSolarResearch

# the same corpus, one sha256, identical on every machine:
bash reproduce/validate.sh                    # every published figure, digest-pinned

# the observer-dependence of the float verdict, machine by machine:
#   (staged and run the way validate.sh runs every reproduce script)
#   -> 142 two-pi contradictions · float32 flips 0 · exact court refuses all 142
#   -> PROOF_EXACT_VERDICT_IS_OBSERVER_INVARIANT

Affine Fusion Control — the app the verdict runs in, five panels, both courts live
Affine Fusion Control — public release — the measured, claim-graded public case
Study 33 — the fusion control verdict court — the charter and the three-way fork it retired
Study 35 — the safety brain that forgets — the sibling axis: this study proves the verdict is the same for every observer; Study 35 proves it does not decay over time, across machines, or against a new event
The full-grade replacement — the over-scaling law, and the vendor-to-public argument across 50 domains
Ontology — the evidence grades and the one-way-flow rule this page obeys

Rights — source-available, all rights reserved. This wiki and its repository are published for public inspection and to let anyone re-derive the figures. They carry no LICENSE; under default copyright, all rights are reserved. No right is given or intended to use, run, or deploy it for any purpose other than re-deriving the published figures, nor to modify or build on it — any other use requires a written licensing agreement with the authors. · Affine.Earth · zero float · zero shear

🧬 CURES — read in this order

Each step is the reason the next one exists. Nothing here is medical advice, and no page calls any medicine safe or unsafe.

1 · Why an exact safety screen at all

Cures Without the Gatekeeper — the medicine front door: six real written medicines, one screen anyone can re-run
The library admission law — what may enter, and the 71 arms that prove it refuses. The primary artefact.

2 · The three libraries, which grow rather than close

The Library of Compound Cures — exact off-target maps for the medicines the registry publishes
The Library of Proteins — 80,080 generated sequences, novel chemical matter, graded honestly
The Library of Material Systems — what a system is, what was measured, where the law lives. C-007 absolute: no recipes

3 · The maps — every place a molecule could act, counted

The off-target atlas — every nucleic-acid medicine the registry publishes a sequence for: WHERE it can pair
The order of the bases — WHETHER THAT BURDEN IS UNUSUAL: 472 strands ranked against sixteen rearrangements of their own bases
Where else could this guide cut? — the whole human genome, counted
Designed, or forced by its own bases? — every clinical CRISPR guide, with its own composition as the control
What a public genome deposit will tell you — and four ways it will mislead a health tool first
Study 45 — which of nine billion answers a laboratory can act on — a safety review of AlphaGenome Atlas, measured live on 1,200 real variants at two genes. The headline score separates every one. The detailed tracks do not: splice-site usage hands back 950 of every 1,000 values shared with another variant at HBB and 998 at CFTR, and the shared values pile up in the quiet band where a bench clears a variant

4 · One medicine at a time

Zilganersen — the first treatment for Alexander disease, screened on the real approved sequence
A drug an AI designed — rentosertib for pulmonary fibrosis, and exactly what our instruments reach
CAR-T, halted — the verdict a regulator could re-derive
N-of-1 antisense — the only safety net at a population of one
VERVE-102 — the off-target lattice a stranger can re-derive
PM359 — prime editing, certified before anyone is dosed
Del-Zota — the one safety question that can be made exact

5 · What keeps a disease alive, and what moves it

Study 26 — master regulator bonds — 17 tumour types, 7,673 tumours; eleven compound pairs where no single agent among 20,308 cleared any
Study 20 — Rife frequency — light and frequency, measured rather than dismissed
Study 37 — five molecules — 37,910 "validated discoveries", 5 distinct molecules; why per-item validation cannot see a corpus-level defect
Are the generated cures new? — 80,080 peptides against the human proteome
Study 16 — disease type · Study 17 — chemistry InChIKey · Study 14 — protein lattice

⚖️ How to read any page here

No language model in this stack — what the answers here are made of: measured 2026-09-12, no cell runs a model process, opens a model port or holds an unmasked model unit, and a gate refuses their return
Run any study in your browser — all ninety programs open on your own device, forty-nine run there, and the run tells you whether it printed the sealed bytes
The ontology — grades, terminals, controls, and what each page may say
Zero Float · Zero Shear — the method in one page
Ask someone you trust to check this — what to hand a sceptic
Readers’ guide · Program index — all 42 studies · White paper · Roadmap
The full-grade replacement — 49 retired instruments, 4 verticals
The exactness seam — the business case
Build a study — Falcon walkthrough — how to add one yourself

🔬 The method — exact against float, domain by domain

The same move every time: take a domain where a floating-point model is the accepted instrument, compute the same quantity in exact integers, and seal the cases where the two render opposite verdicts. The subject under grading is always the instrument, never the phenomenon.

Study 48 — the atom already has an address — silicon dimers 3.840 Å apart, the smallest commanded scale on the board: a length carried in single precision mis-addresses its first atom at step 8,783; an address cannot
Study 49 — the phase code never needs π — a phase-only modulator takes 256 codes per pixel; the code is a ratio of integers
Study 50 — CMS raw data from the LHC, read exactly — CMS's 2011 collision bytes streamed from CERN Open Data into the Affine IDE and read in exact integers, every collision a hologram you can turn: 138 of 3,564 bunch slots carry 93,110 of 120,742 collisions, and in 3,854 the event record reads its slot exactly 3 lower than the pixel boards · public release
Study 55 — IceCube: the light in the ice, hit by hit — IceCube's calibrated hits read byte for byte: 4 published files, 9,749 events, 2,289,821 hits, a census seal per file
Study 47 — translation shear: the meaning that survives a language — LAW FROZEN · LIVE CLAIM, measured 2026-09-11 and again fleet-wide 2026-09-12: translation as an exact coordinate transform, charts derived in memory at every start from the raw rows of a pinned public weight file and never written down; one lattice digest on 9/9 cells, zero drift, every refusal named. The generative comparison arm is ABSENT — there is no generative translator in the stack
Study 34 — the observer-invariant verdict — why a safety verdict needs an exact law, not a bigger computer
Study 35 — the safety brain that forgets — deaf in 8.4 seconds, forgets across machines, disagrees with itself
Study 36 — the language game of Fermat's Last Theorem — guess and shear, or project
Study 40 — the number the simulation throws away — their ICO result computed as a fraction; in float the effect returns 0 at every width, and an effect returned as zero cannot be searched for
Study 41 — fifty years of solving the wrong problem — the ordering was never about time, it was about arithmetic; 177× the work and 2,400× the wrong guesses to return the answer the machine already had
Study 42 — The Exact Contract — 2.7M flood settlements in Int128 cents; the step exists and the rigidity does not
Study 29 — continuous-model shear
The lattice holds · Impact study — continuum dead · Death of continuous shear
Fourier Phantom — Anima FNO vs 11+12+13 · Stellar dynamo kill shot
QCD: freedom is dilation · UUM-8D vs IUT — WIN
Peer-review bundle · Conjecture alignment

⚡ Fusion — the energy case

We need fusion — the verdict every machine can check
Affine Fusion Control — the local exact-integer court · public release
Fusion researcher's guide
Study 33 — the fusion control verdict court

🌍 The planet, and the sky

Every season, fifty tonnes — the biosphere-safety case
The forcing nobody measures · Impact study — the SpaceX trajectory
Study 31 — the biosphere joint ledger — LIVE on the court, 9/9 cells
Study 28 — the wet-bulb threshold court — Act 1 sealed
Study 32 — the taxi-out floor court
Where humans actually yield — the fatigue curves, and where the rules already agree
Study 30 — sovereign edge pod · Manufacture contracts

🏛 Markets, money and risk

The detector that flags the whole market — a manipulation geometry in exact integers, and the regulator's own indicator scored against a legitimate quoter
Study 43 — almost every order is cancelled, and that is normal — nine sessions, three operators, two continents: 935 to 998 of every 1,000 orders that ended, ended without trading. A check that flags almost everything is a denominator, not a detector — and the stock you pick moves it further than the exchange does
Study 44 — nine billion answers, four billion ways to say them — AlphaGenome Atlas ships 9 billion predictions in single-precision floats, which hold 4.28 billion distinct values: 52 of every 100 variants MUST share a score with another. Agreement and exhaustion look identical on the wire
Study 38 — the loss-reserve triangle — a reserve is an exact rational; 481 of 482 verdicts identical in both arithmetics; the sixteen-billion figure comes from an unchecked premise
Study 39 — the actuarial domain — life, pensions, multi-state and aggregation; the margin is 8 significant digits at its tightest

⚛️ Run a court yourself

Run any study in your browser — the ▶ badge beside a program name opens it in the Studio, already built and carrying its inputs, and runs it on your machine with nothing sent back
Explore the live courts
MCP user guide — all 51 tools · Deterministic no-float courts for LLMs
Court Client — generic wasm IDE for every court · Court-client checkpoint
Coding Court — the verdict IS the artifact
Zed — the coding agent, for developers — set Zed 1.20.2 up on https://affine.earth/v1, no language model anywhere; what a turn does, the wire, the autonomous closure
Zed — Minecraft comes to life — the two-person interaction, sealed: it asks, cites, clones a sibling with a value you supply, verifies by replay; the court flips REFUSED_UNKNOWN_BUDGET → WIN
Zed — the agent that teaches the whole domain — architecture, protocols, server management and git, each answered from lines it read and instruments it ran; five closures PROVEN, and the cattle question answered with a counter the fleet did not have
Math Court on Glama · Math Court user guide · Example app — entire court
Quantum algorithms inventory · Shor witness certifier
MCP clients (public)
Glama connector
Look in the UI (no visitor data)

📒 Program ledger — every study by lifecycle

A study appears here under the state its evidence has earned, and above under the question it answers. The two are different filings of the same work, on purpose.

✅ LAW FROZEN · DATA SEALED

Study 06 — explosion vs earthquake · Study 07 — Sgr A* raw visibilities
Study 11 — Ehrhart volume · Study 12 — parallel repetition · Study 13 — Connes rigidity
Study 14 — protein lattice · Study 16 — disease type · Study 17 — chemistry InChIKey
Study 18 — material STD · Study 19 — Go First dice
Study 26 — master regulator bonds — 17 tumour types, every finding published

🔴 LIVE CLAIM — standing, not sealed

🌊 CHARTER · OPEN — the findings, published either way

Study 03 — flare SIDs — archive went dead · predictions and validations
Study 04 — tsunami vs surge — partial seal · Study 05 — Forbush decreases
Study 08 — Gaia BH1 — no corpus until DR4 · Study 10 — Fermi / dark matter — does not disprove DM
Study 15 — Skala DFT shear · Study 27 — exact nuclear scattering

Study 34 Observer Invariant Verdict

Study 34 — The Observer-Invariant Verdict: why an exact law, not a bigger computer, makes the safety verdict the same for everyone

We need fusion

Affine Fusion computes the verdict in exact integers

What the two programs measure

Why an observer-invariant verdict matters

The limits — a court states its own

Rights — source-available, not open-source

Reproduce it yourself

Related

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!