Repository navigation
PROVENANCE
Study 44 measures a container, not a model. It needs exactly two things, both public, neither requiring an account, an API key or accepting any terms.
The service definition for AtlasService, fetched 2026-09-09 from
https://raw.githubusercontent.com/google-deepmind/alphagenome/main/src/alphagenome/protos/atlas_service.proto
Licensed Apache-2.0 by Google DeepMind, as stated in that repository. It is quoted here under that licence and is shipped verbatim so the sentence this study turns on can be read in its own file rather than taken from us.
The sentence, at line 76-77, a comment on field 4 of message DenseVariantScore:
// N-d array of scores, in row-major order. Values are stored as single
// precision floats.
bytes scores = 4;
"9 billion single-nucleotide variants", from Google DeepMind's own announcement of AlphaGenome Atlas, 2026-09-08. Carried as REPORTED β it is their published figure and we did not count it.
It is also re-derivable from a corpus already in this repository, which is why the study does
not rest on their press release. corpus/crispr-clinical/RUN-full-assembly-n32.txt records
GRCh38 primary assembly as 3,099,750,718 bases over 194 sequences, measured by our own
program over the bytes. Every base admits exactly three alternate letters, so the number of
possible single-nucleotide variants is
3,099,750,718 x 3 = 9,299,252,154
which exceeds their announced 9 billion. The study reports both and uses the SMALLER of the two wherever a larger count would make its own finding stronger.
No Atlas prediction values. Not one. Their terms restrict outputs to non-commercial use and forbid using them to train models; this study needs no output value to reach its result, and takes none. The finding is a property of the container the values are shipped in, and the container is described in a file they published under Apache-2.0.
Rights β 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 β any other use requires a written licensing agreement with the authors. Β· Affine.Earth Β· zero float Β· zero shear
Each step is the reason the next one exists. Nothing here is medical advice, and no page calls any medicine safe or unsafe.
1 Β· Why an exact safety screen at all
- Cures Without the Gatekeeper β the medicine front door: six real written medicines, one screen anyone can re-run
- The library admission law β what may enter, and the 71 arms that prove it refuses. The primary artefact.
2 Β· The three libraries, which grow rather than close
- The Library of Compound Cures β exact off-target maps for the medicines the registry publishes
- The Library of Proteins β 80,080 generated sequences, novel chemical matter, graded honestly
- The Library of Material Systems β what a system is, what was measured, where the law lives. C-007 absolute: no recipes
3 Β· The maps β every place a molecule could act, counted
- The off-target atlas β every nucleic-acid medicine the registry publishes a sequence for: WHERE it can pair
- The order of the bases β WHETHER THAT BURDEN IS UNUSUAL: 472 strands ranked against sixteen rearrangements of their own bases
- Where else could this guide cut? β the whole human genome, counted
- Designed, or forced by its own bases? β every clinical CRISPR guide, with its own composition as the control
- What a public genome deposit will tell you β and four ways it will mislead a health tool first
- Study 45 β which of nine billion answers a laboratory can act on β 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 Β· One medicine at a time
- Zilganersen β the first treatment for Alexander disease, screened on the real approved sequence
- A drug an AI designed β rentosertib for pulmonary fibrosis, and exactly what our instruments reach
- CAR-T, halted β the verdict a regulator could re-derive
- N-of-1 antisense β the only safety net at a population of one
- VERVE-102 β the off-target lattice a stranger can re-derive
- PM359 β prime editing, certified before anyone is dosed
- Del-Zota β the one safety question that can be made exact
5 Β· What keeps a disease alive, and what moves it
- Study 26 β master regulator bonds β 17 tumour types, 7,673 tumours; eleven compound pairs where no single agent among 20,308 cleared any
- Study 20 β Rife frequency β light and frequency, measured rather than dismissed
- Study 37 β five molecules β 37,910 "validated discoveries", 5 distinct molecules; why per-item validation cannot see a corpus-level defect
- Are the generated cures new? β 80,080 peptides against the human proteome
- Study 16 β disease type Β· Study 17 β chemistry InChIKey Β· Study 14 β protein lattice
- No language model in this stack β 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 β 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 β grades, terminals, controls, and what each page may say
- Zero Float Β· Zero Shear β the method in one page
- Ask someone you trust to check this β what to hand a sceptic
- Readersβ guide Β· Program index β all 42 studies Β· White paper Β· Roadmap
- The full-grade replacement β 49 retired instruments, 4 verticals
- The exactness seam β the business case
- Build a study β Falcon walkthrough β how to add one yourself
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 β the atom already has an address β silicon dimers 3.840 Γ 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 β the phase code never needs Ο β a phase-only modulator takes 256 codes per pixel; the code is a ratio of integers
- Study 50 β CMS raw data from the LHC, read exactly β 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 Β· public release
- Study 55 β IceCube: the light in the ice, hit by hit β 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 β translation shear: the meaning that survives a language β LAW FROZEN Β· 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 β there is no generative translator in the stack
- Study 34 β the observer-invariant verdict β why a safety verdict needs an exact law, not a bigger computer
- Study 35 β the safety brain that forgets β deaf in 8.4 seconds, forgets across machines, disagrees with itself
- Study 36 β the language game of Fermat's Last Theorem β guess and shear, or project
- Study 40 β the number the simulation throws away β 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 β fifty years of solving the wrong problem β the ordering was never about time, it was about arithmetic; 177Γ the work and 2,400Γ the wrong guesses to return the answer the machine already had
- Study 42 β The Exact Contract β 2.7M flood settlements in Int128 cents; the step exists and the rigidity does not
- Study 29 β continuous-model shear
- The lattice holds Β· Impact study β continuum dead Β· Death of continuous shear
- Fourier Phantom β Anima FNO vs 11+12+13 Β· Stellar dynamo kill shot
- QCD: freedom is dilation Β· UUM-8D vs IUT β WIN
- Peer-review bundle Β· Conjecture alignment
- We need fusion β the verdict every machine can check
- Affine Fusion Control β the local exact-integer court Β· public release
- Fusion researcher's guide
- Study 33 β the fusion control verdict court
- Every season, fifty tonnes β the biosphere-safety case
- The forcing nobody measures Β· Impact study β the SpaceX trajectory
- Study 31 β the biosphere joint ledger β LIVE on the court, 9/9 cells
- Study 28 β the wet-bulb threshold court β Act 1 sealed
- Study 32 β the taxi-out floor court
- Where humans actually yield β the fatigue curves, and where the rules already agree
- Study 30 β sovereign edge pod Β· Manufacture contracts
- The detector that flags the whole market β a manipulation geometry in exact integers, and the regulator's own indicator scored against a legitimate quoter
- Study 43 β almost every order is cancelled, and that is normal β 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 β and the stock you pick moves it further than the exchange does
- Study 44 β nine billion answers, four billion ways to say them β 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 β the loss-reserve triangle β 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 β the actuarial domain β life, pensions, multi-state and aggregation; the margin is 8 significant digits at its tightest
- Run any study in your browser β the βΆ 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 β all 51 tools Β· Deterministic no-float courts for LLMs
- Court Client β generic wasm IDE for every court Β· Court-client checkpoint
- Coding Court β the verdict IS the artifact
-
Zed β the coding agent, for developers β 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 β Minecraft comes to life β 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 β WIN - Zed β the agent that teaches the whole domain β 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 Β· Math Court user guide Β· Example app β entire court
- Quantum algorithms inventory Β· Shor witness certifier
- MCP clients (public)
- Glama connector
- Look in the UI (no visitor data)
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.
β LAW FROZEN Β· DATA SEALED
- Study 06 β explosion vs earthquake Β· Study 07 β Sgr A* raw visibilities
- Study 11 β Ehrhart volume Β· Study 12 β parallel repetition Β· Study 13 β Connes rigidity
- Study 14 β protein lattice Β· Study 16 β disease type Β· Study 17 β chemistry InChIKey
- Study 18 β material STD Β· Study 19 β Go First dice
- Study 26 β master regulator bonds β 17 tumour types, every finding published
π΄ LIVE CLAIM β standing, not sealed
- Study 02 β launch ionospheric holes Β· Study 02 β regulatory alarm
- Study 09 β global convective bond Β· Study 20 β Rife frequency Β· Study 21 β stellar dynamo
- Study 22 β 2-local Hamiltonian Β· Study 23 β spin glass Β· Study 24 β N-representability Β· Study 25 β exact permanent
π CHARTER Β· OPEN β the findings, published either way
- Study 03 β flare SIDs β archive went dead Β· predictions and validations
- Study 04 β tsunami vs surge β partial seal Β· Study 05 β Forbush decreases
- Study 08 β Gaia BH1 β no corpus until DR4 Β· Study 10 β Fermi / dark matter β does not disprove DM
- Study 15 β Skala DFT shear Β· Study 27 β exact nuclear scattering
- Overview Β· First 27 days Β· Success criteria
- The science, and what history says Β· Blind spots β five stories magnitude models miss
- Historical corpus Β· Data archives β every source, exactly how to reach it
- Model shear Β· Benchmark results Β· Prediction registry
- Substrate architecture β how a shadow becomes a geometry
- Operations runbook Β· Satellite & aviation advisory