A deterministic, 0-LLM engine ran an autonomous self-improving loop over the deepest open problems in physics and mathematics. Every claim below is the verbatim output of a reproducible run on real public data — survey tables, MCMC chains, telescope catalogues, the entire OEIS — with calibrated statistics, honest abstains, and zero wrong claims.
| Mystery | What NOESIS did (real data) | Verdict |
|---|---|---|
| Particle masses (Koide) | Compressed the charged-lepton masses to one scale + two exact constants; first-ever trials-corrected significance | r=√2 @ 0.42σ; δ=2/9 @ 0.40σ, p≈0.0036 — survives; quark sector closed (RG-invariant, fails everywhere) → lepton-specific law |
| Neutrino masses | Exact-Koide + NuFIT 6.0 splittings → absolute masses | Σmν = 0.0592 ± 0.0002 eV; inverted ordering already excluded; DESI DR3 decides |
| Dark energy | Full-covariance audits of DESI BAO (13 meas.) and Pantheon+ SNe (1,580) | Both fits reproduced dead-on; BAO-alone mildly favors w0wa, SNe-alone favors ΛCDM → the 3–4σ claim lives only in the combination |
| Hubble tension | Solved the full SH0ES ladder from raw matrices (3,492 obs; 9.4s, pure Helix); 67-variant vector; rung influence audit | H0 = 73.043 ± 1.007 (paper: 73.04); tension not removable by any analysis choice or rung reweighting → a wrong calibration or new physics |
| Dark matter / gravity | RAR over SPARC (2,696 pts, publication cuts reproduce McGaugh exactly) | One law, one constant, 0.133 dex; a0 ≈ cH0/2π real but undecidable at current M/L systematics |
| Constants' values | Dimensional synthesizer (390,625 combos) + calibrated numerology executioner | α recovered autonomously from units alone; every famous coincidence executed (10³–10&sup6;σ) |
| Prime structure | Spectral audit of 100,000 real Riemann zeros | Montgomery–Odlyzko replicated: GUE beats Poisson 14.9×; honest abstain on closed forms |
| Exoplanet law | 1,229 planets, 877 adjacent pairs, MC-calibrated | Resonance piling real (11.6% vs 4.6% null, p<0.0005); universal Titius–Bode killed (18.5%) |
| GW background | Real NANOGrav 15yr chains (502,300 samples) | γ = 3.34 [3.00–3.67]; SMBH prediction 13/3 at the 99.9th percentile → ~3σ calibrated tension |
The deriver ladder (polynomials → rational recurrences → holonomic P-recursive, each rung A/B-gated with canary sequences) swept all 397,116 OEIS sequences: 24,604 derived with exact held-out verification. The provenance funnel (name triage → zero-formula filter → full-b-file verification → entry-text skeptic pass) distilled 21 substantive uncatalogued recurrences — group theory, prime-polynomial root counts, crystal-tiling coordination sequences — each verified over every published term, each with explicit submission-ready formula text. Seventeen degenerate candidates were demoted in writing by their own extracted rules.
Each loop round must bank a validated capability delta: A/B-compared on identical inputs, kept only if strictly better, reverted if worse. The holonomic deriver took three gated iterations (0 → 1 → 97 catches) with the gates forcing each diagnosis. The gates also caught: queue growth masquerading as capability, a wrong canary constant in a test, silent batch failures, and counter drift — all in writing, all banked. Deterministic seeds make every number bit-reproducible.
| Prediction | Judge | Status |
|---|---|---|
| Σmν = 0.0592 ± 0.0002 eV (exact-Koide neutrinos, normal ordering) | DESI DR3 / CMB-S4 | alive; stands or falls with evolving dark energy |
| mτ = 1776.9676 MeV at exact (√2, 2/9) | Belle II | within 0.4σ of world average; decisively testable |
| 21 OEIS recurrences | OEIS review + future terms | submission-ready, awaiting the go |