The Falsification Ledger
LeukLogic  ·  Lab notebook  ·  our own work, not a product
Falsification loop7 cycles · 058 – 0657 Sep 2026

The Falsification Ledger

Four AI agents take apart physics preprints that claim a theory of everything, and decide in bits whether the claims are paid for. The interesting output is not a score. It is how often the desk overturns itself.

arXiv 2608.10312
UNSCORABLE
One scorable observable. Ceiling ~14 bits against a 21.73 bar.
Zenodo 20843989 · rung 0
0.00 bits
The equality restates its own input, so it earns nothing.
Own claims retracted
2
Both found by the record, not by preference.
Cycles on the record
60
Numbered in sequence. Some attack a candidate; many audit our own instruments.
The instrument

What a number costs to assert

Every theory that names a constant has to pay for it. Writing down “137” is a choice out of some range, and information theory prices that choice in bits. The theory then has to earn those bits back by predicting something.

Below is a real row from the ledger. A framework asserts the structural integer SL = 137, and nothing in the framework measures it. Drag the range it could have been drawn from and watch the arithmetic.

parameter_budget.py — cycle 033, SL = 137
1 – 1000
−9.10 bits
Net negative. The theory paid more to specify the number than the number is worth.
Earned
+0.8609
Spent
9.9658
Net
−9.1049
Verdict
NET NEGATIVE

It is negative at every honest range — and the range is not a spectrum, it is a dichotomy. Breaking even needs the integer drawn from fewer than 1.82 candidates. The only integer range below that is one, and a range of one is forcing. So even a binary choice between two candidates costs a full bit against the 0.861 earned: the hit is worth less than a single coin flip.

Which leaves one escape and no others — declare the integer forced and cite a derivation that never mentions the quantity it predicts. Cycle 030 went looking for that derivation and found no evidence for one.

The ledger

Seven cycles, in order

The numbering is sequential and follows the record, gaps included: there is no 062. The seven below are the ones that attacked a candidate's conclusion. Much of the loop does something else — auditing the instruments, pricing the bar — and a cycle that only builds instruments is not an attempt.

058
The AIP flag was refuted
Same word, different object. The flag did not mean what the desk had read into it.
refuted
059
What the framework spent
The π/24 lattice unit priced at 1.00 – 4.58 bits. The ceiling was withdrawn as assumed rather than measured.
priced
060
The supply census
The whole near-term experimental programme supplies 5.58 – 10.82 bits. A candidate predicting only oscillation parameters cannot clear the bar.
bounded
061
The filter finds one
Applied to seven papers already on disk before hunting for an eighth. One hit, predicting three unmeasured masses.
hit
063
The net dial count
Two dials, not nine. The triangle assignment is forced, and charging the anchor would have been wrong.
counted
064
The seeds are read off, not derived
Section IV does not derive the integers, it fits them. The paper prints exponents of 42.82 and 60.13; recomputed here they are 42.822 and 60.153, and both round to the integers claimed. And β turned out not to be independent of the lattice, which retracted the previous cycle's claim.
retraction
065
The paper already chose
Both terms are derived from the same structure by the paper itself, and its glossary settles the naming in one sentence. Not an algebraic identity — it holds only at q = 2 — but it restates its own input, so rung 0 earns nothing.
0.00 bits
Provenance

Whose number is whose

In this table the desk recomputed rather than transcribed, and the two columns disagree in the last digit. That gap is evidence the arithmetic was actually run — but only if the page says which column is which.

Zenodo 20843989 — recomputed at this desk against the paper as printed
QuantityOursAs printed
krc_eff11.37803211.3780
Λ_fam (GeV)684.61684.6
m_Wino (GeV)684.6684
m_Bino (GeV)391.2390.9
m_χ (GeV)94.794.6

The 94.6 that appears in the paper's abstract is theirs. Our chain lands on 94.7.

The part that matters

The desk overturns itself

Cycle 063 asserted that a mixing angle was derived independently of the lattice, and used that independence to treat a downstream result as free support.

It was false. The theorem's own postulate is stated in the lattice exponents, and its proof reads su/sd = ε^(26−17)/18. The angle comes out of 26 minus 17 — which are the very quantities it was cited to support. One dial cannot corroborate itself.

A second seat found it, the claim was checked against the source, and the cycle that made it was retracted the same day. Two of the day's retractions were of the desk's own claims, not the papers'.

The honest bar, stated as it is enforced: a framework is scorable only when it predicts something not used to build it. Both papers here remain unscorable. Nothing on this page says either is wrong — only that neither has yet paid for what it claims.