Nothing here is on this page because a document says so. Every figure spoken below was recomputed from the result file that produced it, at the moment the page was built.
You Already Have the Sensors
13 SECONDS · 1080×1920 · qc_001
Your plant or farm already runs gas arrays. Process instruments. Grain sensors. You're collecting the data. You're just not listening to it.
Quantum Children reads what you already have.
✓ every number recomputed from a result file
Start With the One That Worries You
15 SECONDS · 1080×1920 · qc_020
You already know which sensor you don't quite trust. Everybody has one.
Start there. Send us its history and nothing else.
Get the free sensor audit, at quantum-children dot com slash audit.
✓ every number recomputed from a result file
Withdrawn 2026-08-23 — Six Gases, One Signature · Nothing to Retune
CLAIM WITHDRAWN · qc_021 · qc_022
These two videos were withdrawn on 2026-08-23 on a matched control that had not been run before. Both rested on "six gases separate by entropy signature at 18.5 sigma" from a public 13,910-sample gas array.
The gases do not sit evenly across the ten measurement batches, so the pooled comparison mixes gas with batch. Holding batch fixed and shuffling only the gas labels — 1000 draws, each pseudo-gas keeping the real gas's exact per-batch counts — reaches 18.44 sigma in 247 of them. p = 0.25. The control we had published, p = 0.003, permutes all 13,910 samples at once, which destroys the batch structure along with the labels, so it never tested this.
Separately, the recipe's rolling window steps by one sample, so consecutive windows share 19 of their 20 readings. On the Ammonia window series — feature 0, all ten batches pooled, 1622 windows — lag-1 autocorrelation is 0.986. The effective sample size is far below 1622: 81 counting only non-overlapping windows, 24 under an initial-positive-sequence estimate, 11 under an AR(1) one. Every sigma in this family is inflated by at least 4.5-fold and, on three of those four estimators, by eight to twelve. Corrected end to end, the pooled 18.44 falls to 4.12 with four of fifteen pairs still past 3 sigma under the most conservative correction, and below 2 with none of them under the others. This is independent of the batch question.
Correction, 2026-08-23: this paragraph previously said the effective sample size was "about 46" and the inflation "roughly sixfold". Neither figure had a result file behind it, and on recomputation neither reproduces — 46 matches no estimator, and the real inflation is larger than sixfold, not smaller. The caveat understated the problem. It is now measured and persisted: qbh/gas_overlap_inflation.json.
What survives, stated plainly, and stated as a property of this statistic rather than of the data: on entropy moments over rolling 20-sample windows, inside batch 7 alone the separation is real and large — 36.42 sigma observed, and none of 1000 shuffles came near it. On that same statistic, inside batch 10 alone there is nothing — 5.94 observed against a shuffled mean of 6.85, p = 0.72. Two batches, opposite answers. That is a genuine per-batch result, and it is not "six gases separate."
The contrast between those two batches belongs to the recipe, not to the samples. Measured 2026-08-23: a different window-level test on the same data — non-overlapping 24-row windows instead of rolling 20-sample ones — finds a gas effect inside batch 10 that is larger than the one inside batch 7, at the floor of its own permutation null. That restores nothing withdrawn above, we are not building on it, and it is recorded here for one reason: a withdrawal notice that overstates its own contrast is the same defect one level down. "Nothing in batch 10" was our result on one recipe. It is not a property of batch 10. Result file: qbh/gas_window_signal_present.json.
Control: qbh/green2_gas_within_batch_null.json. Nothing replaces them yet; we will not re-publish the number until it stands on a control that can kill it.
What "recomputed" means here. Claims are checked against two
tiers of authority. A result file — JSON, CSV, a run log written by a
script that measured something — can certify a number. Prose cannot:
not a summary document, not a README, not a constant typed into source code. A
write-up asserting a figure is a restatement of it, not evidence for it.
So the figures in the video cards above are not looked up. The checker
opens the result file and redoes the arithmetic every time the page is built. If that
arithmetic stopped landing, the page would not build.
Scope, stated exactly, because the sentence above used to say "the
figures on this page" and that was wider than the truth: the checker parses video
cards. It does not parse the withdrawal notices. The numbers inside those notices —
including every figure added on 2026-08-23 — are cited to result files by hand and are
not gated by the build. Naively widening the parser is not the fix: the checker matches
tokens, not statements, so a notice that quotes a withdrawn claim in order to retract it
reads as an assertion of it and comes back REFUTED. Measured, both notices do. Telling
"this is false" from "this is true" is a distinction the checker does not have, and
saying so is more honest than a gate that appears to cover text it never reads.
The gas separation that used to appear here still recomputes, and it
lands at 18.44σ with 11 of 15 pairs beyond 3σ. It read 18.53σ
until 2026-08-23, when the divisor was corrected: the check had been dividing by the
per-species sample count on a statistic computed over windows, of which
there are 20 fewer per species. 18.44 is also the value the matched control was built
against, so the two now agree.
That arithmetic is not what withdrew the claim, and recomputing a
number is not the same as testing it. A control did — see the withdrawal
notice above. A figure can be reproducible and still be measuring the wrong
thing.
Claims we cannot certify this way are marked unverified and
are simply absent from this page. Unverified is not a claim that they are false
— it means nobody has shown them yet, which is a different and more honest
thing to say. Those counts are kept separate and never added together. Audit it
with python3 -m src.claim_check --evidence.