Frozen Predictions

A pre-registration is a claim written down and time-stamped before the measurement that will test it — so that no one, including its author, can quietly revise it afterwards. Here the commit hash is the timestamp: these files were committed to the public repository and tagged prereg-2026-08-02. Every number below is regenerated from CODATA 2022 inputs by the scripts in this directory; nothing is quoted from memory.

This is a scoreboard, not a showcase. A registry that lists only the surviving predictions would misrepresent the corpus and destroy the credibility the freezes were written to earn. The neutron–proton result is recorded here at the top, marked FIRED — 7.24σ, alongside the live ones.

The scoreboard

Commit = the commit tagged prereg-2026-08-02 in the repository; that hash is the timestamp of record. Last checked = the date the standing was last re-computed against measurement. Two further documents are not predictions and carry no row: RIGIDITY.md (why the integer 21 cannot move — the answer to "this is numerology") and PROTOCOL-lattice-qcd-mapping.md, a sealed-envelope procedure for the one blind test still available — explicitly not a freeze.

The registration — proton–electron mass ratio at order α³

Freeze date: 2026-08-02 · Author: G · Studio G · the420code.org · Scope: AP30 (The Resistance) and Ø Predictions Part I
Reproduction: verify_prereg.py recomputes every number below from CODATA 2022 inputs.

1. What this document registers, and what it does not

It registers one number: the value the AP30 formula produces when the series is truncated where AP30 truncates it, at order α².

It does not register a value for the order-α³ coefficient. As of this freeze date the corpus contains no derivation of that coefficient, and none is claimed here. §5 explains why any value stated today would be a fit rather than a prediction.

This is a narrower claim than earlier drafts of this document made. The narrowing is deliberate.

2. Inputs (CODATA 2022, NIST, retrieved 2026-08-02)

Earlier drafts used α⁻¹ = 137.035999084. That is the CODATA 2018 value. The correction moves the frozen value by 0.013 ppt — numerically irrelevant against a 10 ppt effect, but the label was wrong and is corrected here.

3. The formula, with its epistemic ledger

What each term rests on, stated as the corpus itself states it:

The integers 21, 3, 4 are fixed in AP10/AP28, upstream of any mass calculation. That ordering is the load-bearing fact and it is checkable against the corpus's own dependency graph.

4. The frozen value

Evaluated at CODATA 2022 α:

Against measurement:

D lies above the measured value. Any order-α³ term that closes the gap must therefore be negative, while the α and α² terms are both positive.

D is the registered prediction. It was fixed by structure, not by the measurement, and it is the only value in this document not selected with the residual in view.

4a. The α² coefficient is empirically pinned (registered 2026-08-02)

Subtracting the exactly known 1836 and O(α) terms from the measurement and dividing by α²:

Written as an integer numerator over 1836, the measurement requires 335.4 ± 1.1. The integers inside the 3σ window are 332 … 339. Exactly one is a product of powers of {21, 3, 4}:

336 = 21 × 16 = 84 × 4 = 4² × 21

Consequence for the recursion. The naive rank-3 continuation 4³ = 64 gives c₃ = 21 × 64/1836 = 0.732 — excluded at 9.5σ (same sign) or 8.3σ (reversed). Any derivation producing 16 at second order must not produce 64 at third. See RESEARCH-NOTE-deriving-336.md.

5. Why no α³ coefficient is registered

Write the third-order term as c₃α³. The measurement constrains c₃ to:

c₃ = −0.047 ± 0.082 (1σ), i.e. the interval [−0.129, +0.035]

The uncertainty is 1.7× the magnitude of the central value; the interval contains zero. A bounded enumeration over the corpus integers returns 910 distinct admissible values inside 1σ. The sign is undetermined, the coefficient is unconstrained, and the generating rule is owed.

The honest position: the framework predicts D. It does not yet predict anything at order α³.

6. Correction to the uniqueness claim

Ø Predictions Part I states three conditions on the decomposition of 1836 and reports that exactly one sum satisfies all three. Independent enumeration (verify_prereg.py) reproduces the eleven under the first two conditions and returns two under all three:

Condition 4 (formalised here). The number of structural factors per term, counted with multiplicity, must be non-increasing across the ordered triple. Under conditions 1–4 the enumeration returns exactly one decomposition: 1764 + 63 + 9.

Declared: condition 4 was formalised on 2026-08-02, after the alternative surfaced. AP30 §3 already states the criterion in prose ("each successive term using fewer body-of-work numbers"), but as strictly decreasing; the adopted decomposition is 3, 2, 2 — non-increasing. AP30's prose is corrected to "no more than." KS-30.4's within-family limb is recorded as reopened 2026-08-02 and closed the same day by condition 4, with the ordering disclosed.

7. Binding conditions

1. KILL D. D dies at 3σ once measurement uncertainty reaches 3.3 ppt, and at 5σ at 2.0 ppt. Current best is 17.4 ppt (CODATA 2022) / 20.2 ppt (HD⁺) / 25.6 ppt (H₂⁺). This is a real kill condition and it arrives within 3–5 years.
2. CONFIRM D. The measurement converges on 1836.152673444 as uncertainty falls below 5 ppt.
3. NO POST-HOC c₃. If D dies, the framework may not repair itself with a coefficient chosen against the residual.
4. KS-30.4 stays open until §6 is resolved.
5. No retro-editing. This file's hash and date are the record.

9. The muon: a domain boundary, registered

A three-term sum with strictly decreasing 21-exponents requires a leading term of at least 21² = 441, a second of at least 21, and a third of at least 1. The construction has a floor of 463. m_μ/m_e = 206.77 lies below that floor and is unreachable at any exponents. The floor is itself a kill condition, and it is registered as one.

10. What this establishes and what it does not

Establishes: a single frozen number, derived from structure fixed in advance, currently 0.57σ from the best measurement, with a kill condition that binds within roughly five years and cannot be evaded by adjusting a free coefficient.

Does not establish: that the decomposition is unique (§6 — it is not); that the α² coefficient is derived (KS-30.3); that the layers add (KS-30.1); or that a match at 17 ppt selects this construction over the many others that survive at the same precision. A confirmation can only remove rivals. It can never establish uniqueness.

Errata — 2026-08-02

What changed on 2026-08-02 and why, in a form a referee can check. Full text: ERRATA-2026-08-02.md.

E1 — α mislabelled

1/137.035999084 is CODATA 2018. CODATA 2022 is 1/137.035999177(21). Effect on the proton prediction: 0.013 ppt — numerically irrelevant, the label was wrong.

E2 — uniqueness of the 1836 decomposition

Two decompositions survive the three stated conditions, not one. A fourth condition (factor count non-increasing) leaves exactly one. AP30's wording is corrected from "fewer" to "no more than." Condition 4 formalised 2026-08-02, after the alternative surfaced. Declared.

E3 — KS-30.4 status event

Within-family limb reopened 2026-08-02, closed same day by condition 4. Between-family limb unchanged and open. Total switch count unchanged at 561.

E4 — the α² coefficient 16

AP30 §6 states its decomposition is owed under KS-30.3. Any text presenting 16 as derived is corrected to match AP30's own ledger.

E5 — the α³ residual

Restated as: 9.98 ppt above measurement, 0.57σ, with c₃ = −0.047 ± 0.082 unconstrained. There is no α³ prediction.

E6a — the neutron–proton difference: KS-NPP.1 has fired

Predicted 2.53099393, measured 2.530988574 ± 0.00000074. The offset is 2.12 ppm against a bar of 0.288 ppm — 7.24 σ, verified two independent ways. Recorded FIRED on its empirical limb; the structural limbs (KS-NPP.2, KS-NPP.3) are unaffected. No repair is offered: the required correction is negative, every existing term is positive.

E6b — the visible fraction

Planck 2018 gives 4.86–4.91% depending on convention; the corpus quoted 4.86% (the bottom). Against the midpoint the deviation is 2.5 %, i.e. 2.46 σ, not 2.0 %.

E6c — the α² coefficient is now pinned (a strengthening)

The measurement requires 0.1826623 ± 0.000601 (0.33%). 21×16/1836 agrees at 0.57σ; 336 is the only product of powers of {21, 3, 4} inside 3σ. The value is settled; the decomposition is what is owed.

E7 — citation

Alighanbari et al. is Nature 644, 69–75 (2025). Corrects any occurrence of "Nature 625."

The frozen documents

• 2026-08-02-mp-me-alpha3.mdThe proton freeze — the standard the others follow. Registers D = 1836.152673444331.
• 2026-08-02-H0-KS45.1.mdThe Hubble constant, 74.3 ± 1.2 km/s/Mpc, from the acceleration floor.
• 2026-08-02-dark-clock-KS42.6.mdThe dark-sector clock, τ/t_H = 6/21, and the direction Ω_DM/Ω_b rises with time.
• 2026-08-02-visible-fraction-KS41.1.mdThe visible fraction, 1/21, with the corpus's own comparison corrected.
• ERRATA-2026-08-02.mdThe seven corrections and strengthenings of 2026-08-02, E1–E7.
• RESEARCH-NOTE-deriving-336.mdBrief for the owed KS-30.3 derivation — deriving 336, and what would falsify it.
• RIGIDITY.mdThe answer to "this is numerology": why the integer 21 cannot move. Not a freeze.
• PROTOCOL-lattice-qcd-mapping.mdThe sealed-envelope protocol for the one blind test still available. Not a freeze — the procedure for making one.

Reproduce it yourself

The one external dependency is mpmath. Every number on this page comes out of these three scripts:

Scripts: verify_prereg.py · verify_cosmology.py · verify_rigidity.py. The head of the proton script:

from mpmath import mp, mpf, pi, sqrt
mp.dps = 50

# Inputs are CODATA 2022 (NIST, physics.nist.gov/cuu/Constants)
AINV, AINV_U = mpf('137.035999177'), mpf('0.000000021')   # CODATA 2022
MP_ME, MP_ME_U = mpf('1836.152673426'), mpf('0.000000032')
ALPHA = 1 / AINV
ALPHA_U = ALPHA * (AINV_U / AINV)

# The frozen O(alpha^2) value
T0 = mpf(21**2 * 4 + 21 * 3 + 3**2)            # = 1836
T1 = ALPHA * 21 * (1 - 1 / (84 * pi))
T2 = ALPHA**2 * 21 * mpf(16) / 1836
D = T0 + T1 + T2                               # = 1836.152673444331

resid = D - MP_ME                              # = +9.98 ppt
# discrepancy = resid / u = 0.57 sigma; D lies above measurement,
# so any alpha^3 term that closes it must be NEGATIVE.

Or verify them right here — offline. Each button re-derives the numbers independently in pure JavaScript at 60-digit precision, entirely in your browser: no install, no Python, no network, nothing sent anywhere. The reference scripts (verify_prereg.py, verify_cosmology.py, verify_rigidity.py) agree to every digit — a second, independent implementation that matches is stronger than re-running one file twice.