USR-2026-0034 · Quantum
No lazy-PRNG structure found in the delivered quantum bits examined: 32,768 post-processed ANU vacuum-fluctuation bits (bias 0.0003, compression 1.003, next-bit prediction 0.501) and 56,832 SHA-512-whitened NIST beacon bits (bias 0.0006, compression 1.002, prediction 0.506), with weak-LCG and LFSR-32 controls correctly caught.
PROVISIONAL
null on delivered quantum bits — bounds the delivered stream, not the raw device
re-run & confirmed 2026-07-09 (R0 — founder CI; independence pending)
CORRECTED 2026-07-26: the numbers previously registered (1.001 / 0.497 / 199-of-400) were the os.urandom CONTROL row, not quantum data. And the ceiling must be stated — both feeds are bias-corrected or hash-whitened, so this bounds laziness in the DELIVERED stream, not in raw device output.
Falsify-box — how to kill this claim
SURVIVED
Run the battery in randomness_seam/analyze_nature_bits.py on quantum-RNG bits and declare a DETECTION, killing the null, if on a feed of >= 10^4 bits any of the script's own thresholds trips - monobit |mean - 0.5| > 0.02; zlib compression ratio < 0.98; order-10 Markov next-bit accuracy > 0.53; or Berlekamp-Massey linear complexity below 0.35x the profile length on a 3000-bit profile, i.e. < 1050 - AND the same branch trips on a second, independently sourced quantum feed. The 10^4-bit floor and the replication requirement are load-bearing, not decoration: at analyze_nature_bits.py's own minimum admissible length (1000 bits, lines 96 and 98) the monobit branch sits at z = 0.02/(0.5/sqrt(1000)) = 1.27, a 20.6% chance false-positive rate on truly random bits, so an unpinned 'any single one is fatal' would kill a true null on noise; at 10^4 bits the same branch is z = 4.0, P < 1e-4. Calibration proof it can still fire despite the guards: weak LCG gives compression 0.019 against the 0.98 threshold and order-12 prediction 0.980 against 0.53; LFSR-32 gives LC(3000) = 32 against the 1050 threshold - margins of 50x and 30x. Sensitivity ceiling that must travel with the bound: at a 3000-bit BM window nothing above LC ~1050 is visible, and Mersenne Twister (GF(2) complexity 19937) measures LC(3000) = 1498, statistically indistinguishable from the ANU feed's 1499 - so this bounds only lazy generators, and no efficient test can ever exclude a cryptographic PRNG.
Verification record — every quoted number, re-run 2026-07-09
| quantity | measured | verdict |
|---|---|---|
| ANU vacuum-fluctuation bits | 32,768 bits: bias 0.0003, compression 1.003, prediction 0.501 | MATCH |
| NIST beacon bits | 56,832 SHA-512-whitened: bias 0.0006, compression 1.002, prediction 0.506 | MATCH |
| controls | weak LCG and LFSR-32 correctly caught | MATCH |
| ceiling | both feeds are whitened — this cannot test the raw source | MATCH |
| registered numbers, superseded | 1.001 / 0.497 / 199-of-400 were the os.urandom control row | MATCH |
Provenance
script: randomness_seam/run_randomness_seam.py
script status: PRESENT
datasets: Quantum-RNG bitstreams (qdata) · NIST-style statistical battery · claim-scripts (downloadable from /data — run it yourself)
re-run: 2026-07-09
script status: PRESENT
datasets: Quantum-RNG bitstreams (qdata) · NIST-style statistical battery · claim-scripts (downloadable from /data — run it yourself)
re-run: 2026-07-09
Status, honestly
CORRECTED 2026-07-26 by the kill-test sweep; the audit flag it carried is resolved. Status here is computed from evidence — the author cannot set it, and neither can we. Independent reproduction would move it; nothing else will.
§ Figures
2 from randomness_seam/run_randomness_seam.py
What this claim's own script draws. These are the founder's committed outputs — the same plots the script regenerates on any machine that runs it. You can redraw them yourself below.
§ Reproduction
run_randomness_seam.py
Two runs of the same script — mine, and one you can start right now. A match proves the result is reproducible; it is still R0 on this registry's independence rings — same code, so it cannot move a status. Only an outside run does that.
Line 5 is the true/crypto control row, drawn fresh from os.urandom on every run. It cannot repeat — on any machine, including the same one twice. That is the control working, not a reproducibility failure: the one script about randomness is the one that is not reproducible.
A · Recorded
founder's machine · 2026-08-02
Building linear-complexity profiles ...
wrote rs_linear_complexity.png
Running the full battery (100k bits each) ...
true / crypto compress=1.001 predict=0.499 lin.complexity(400b)=200
Mersenne Twister compress=1.001 predict=0.498 lin.complexity(400b)=201
LFSR-32 compress=1.001 predict=0.500 lin.complexity(400b)=32
weak LCG compress=0.012 predict=0.980 lin.complexity(400b)=199
wrote rs_battery.png
====================================================================
THE RANDOMNESS SEAM — summary
====================================================================
A deterministic simulator can't make true randomness; it runs a PRNG.
- weak LCG / short LFSR : CAUGHT by compression + prediction fast
- Mersenne Twister : passes short-memory tests, but is LINEAR
(GF(2) complexity 19937) -> catchable by Berlekamp-Massey with ~40k
bits. 'Good enough for science' is NOT good enough to hide.
- cryptographic PRNG : passes every efficient test (provably, if
one-way functions exist) -> the ceiling. Only a LAZY sim is caught.
Physics hook: real quantum RNGs (vacuum noise, photon timing, decay)
produce the bits. Running THOSE through this battery is a real, cheap,
do-able search for a computed universe. Structure there = a yes.
Bell twist: loophole-free Bell violations already rule out a LOCAL
deterministic source. A simulator can still fake them — but only by
computing entangled outcomes JOINTLY (nonlocal bookkeeping), which is
exactly the expensive entanglement it was trying to avoid. The two
seams are the same wall seen from two sides.
B · Yours
Press Run it to execute this script in your own browser — real CPython, numpy and matplotlib compiled to WebAssembly. Nothing is sent to us; it runs on your CPU. First run downloads the runtime (~5 MB), then takes a few seconds. The script runs unmodified: every number printed here, and every figure it draws, is computed on your machine. The figures appear below.
Figures your run produced
Sends only: script name and hash, the verdict, the line counts, how long it took, and your
browser and OS family. No IP is stored (it is hashed), no fingerprint, nothing you typed.
Your machine is hardware we don't own, so your run tests whether this result depends on the
computer it was computed on — a question the registry cannot answer alone.