Study 44 — nine billion answers, four billion ways to say them
AlphaGenome Atlas shipped on 2026-09-08: a precomputed prediction for every possible
single-letter change in the human genome, about nine billion, one petabyte, free
for non-commercial research. This study measures THE CONTAINER those numbers ship
in. It grades no model, reads no prediction, and makes no biological claim.
THE SENTENCE IT TURNS ON is theirs, from their own Apache-2.0 service definition,
field 4 of message DenseVariantScore:
// N-d array of scores, in row-major order. Values are stored as single
// precision floats.
bytes scores = 4;
The proto is shipped here at corpus/alphagenome-atlas/, sha256 037e8ca5, and the
program does not quote that sentence from memory: it verifies the digest and then
GREPS the sentence out of the file, refusing if either check fails.
THE ARITHMETIC, every figure an integer count of bit patterns:
float32 patterns, all 4,294,967,296
of which carry no number (NaN) 16,777,214
usable, widest possible reading 4,278,190,082
announced variants 9,000,000,000 REPORTED
re-derived from GRCh38 here 9,299,252,154 3,099,750,718 x 3
UNCONDITIONAL, whatever range the score takes:
variants that MUST share a value 4,721,809,918 52 of every 100
IF bounded to 0..1, a reading the program does NOT verify:
variants that MUST share a value 7,934,646,784 88 of every 100
mean variants per value 8
the same catalogue in float64: 0
The usable count is deliberately generous — it still counts both infinities and
both zeros — so every collision figure is a FLOOR.
THE FINDING IS NOT THAT ANY ANSWER IS WRONG. It is that agreement and exhaustion
look identical on the wire. Two variants returning the same number may have been
judged equally impactful, or may have run out of distinct values to be given; the
artifact returns the same bytes either way and does not say which.
WHAT IT REFUSES TO SAY, at the same volume: not that any particular pair collides,
because pigeonhole proves collisions exist and names none; not that the model is
wrong, which this study has no instrument for and never touches; and nothing
clinical — their terms say the predictions must not be used for clinical
decision-making, and this study goes further by reading no prediction at all. Not
one AVI value appears in this repository. The 88% is conditional and labelled so;
the 52% is what the study stands on.
7 arms, both directions, run before a byte of the proto is read. The one worth
arguing about hands the pigeonhole 1,000 variants and requires the answer 0 — a
pigeonhole that always reports collisions is a slogan, not an instrument. Four
refusal paths measured: PROTO_ABSENT exit 2, PROTO_DIGEST_MISMATCH exit 3,
SELFTEST_FAILED exit 4, CONTAINER_SENTENCE_ABSENT exit 6.
Seal 642d84b418825906d7bc2ec2933f2c8c50b61f3f74f11445d7c89d44ed62e631, identical
from three directories — repo root, reproduce/, and an unrelated directory with the
root as argv[1].
NO ACCOUNT, NO API KEY, NO TERMS ACCEPTED. Both inputs are public and neither is
behind a sign-in. 17 check_figure pins, each verified present in both the program's
output and the page before it was written.
White paper to 4.8, title and status to forty-four, board row 44. The Date line had
drifted at 4.6 while doc-control carried 4.7; both are corrected rather than carried.
Index surgery, ten paths, zero deletions, verified with git diff --stat.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>