Status: PROGRAM RETIRED — July 2026.

1. Summary

Null Worldtube Theory was a research program (2026) claiming that the Standard Model’s dimensionless constants, particle masses, and Newton’s constant follow from torus-knot topology and the K₇/Spin(7) structural algebra with zero free dimensionless parameters. Twenty-two papers were published on Zenodo; an open-source library (nwt-substrate) computed the claims; reproduction code was published (nwt-analysis).

In July 2026 the program’s own audit machinery — precision scoring at experimental uncertainty, value-provenance linting, a quantified look-elsewhere test, and an independent memory-blind audit protocol — was turned on the full claim set. The claims did not survive. The program is retired, the repositories are archived unmodified as the record, and this document plus the case-study publication (in preparation) are the authoritative statement of what was found.

The decision to retire was made by the author on the basis of the results below, not external criticism. Nothing has been deleted: every paper, commit, and dataset remains available, because the audit trail — a five-months-long, AI-amplified theory-building effort that constructed the instruments of its own falsification and then used them — is the program’s actual contribution.

2. What the audit found

All results below are reproducible from the pinned code (nwt-substrate, branch audit-dag-prereg-surface, commit a8f3335; modules benchmarks/surface.py, benchmarks/neighbouring_value.py, docs/BENCHMARK_TRIAGE_2026-07-12.md).

2.1 Precision accounting (S-NOW). Scored against each measurement’s own 1σ uncertainty rather than a percent-level tolerance, 31 of 36 registered dimensionless claims are excluded as exact statements:

claim residual experimental precision exclusion
1/α = 25π√3 + 1 7.6 ppm 1.5 × 10⁻¹⁰ ~50,000σ
m_p/m_e (walk formula) 0.11% 6 × 10⁻¹¹ ~3.6 × 10⁶σ
v_EW/m_e 27.7 ppm 2.6 × 10⁻⁷ ~107σ
m_e/M_Pl (substrate-pure) ~75 ppm 11 ppm ~7σ

The five claims that survive at 2σ do so either because their targets are loosely measured (Cabibbo λ, sin²θ₁₃, a broad resonance) or because they are the very claims whose provenance is under dispute as fitted or post-selected (η_B, Ω_b/Ω_c — the latter compatible at 0.006σ, the signature of a fit to the central value).

2.2 Look-elsewhere volume, measured. The constants were assembled from a menu: a structural-integer prefactor × a half-integer power of α × up to two rational-coefficient correction stages. Running the same procedure against random targets (neighbouring_value.py) fits them to a median ~2 ppm, with ~83% landing inside CODATA G’s ±22 ppm error bar (full menu; ~24% even on a minimal 6-integer menu). A ppm-scale match to any constant is therefore the expected outcome of the procedure and carries no information about nature.

2.3 Input-hygiene failure in the flagship claim. The published statement that the derived G matches CODATA to −11 ppm, “inside the ±22 ppm experimental error bar” (Paper 18; repeated in this repository’s README and website) was produced with the measured CODATA α as input (gravity/coupling.py defaults to ALPHA_QED). Computed with the theory’s own α = 1/(25π√3+1), the chain misses G by ~150 ppm — approximately 7σ outside the bar. This correction is issued regardless of any other finding.

2.4 The benchmark suite, classified. A failure-mode triage of all 38 public benchmarks (docs/BENCHMARK_TRIAGE_2026-07-12.md): 18 menu-fitted, 13 standard textbook results with the substrate structure inert (e.g. C₆₀’s 174 modes = 3N−6 + icosahedral group theory; magic numbers; Hückel counting), 3 self-comparisons that cannot fail (including the electron anomalous moment scored against its own Schwinger expression), 3 driven by measured inputs (including a hardcoded PDG α_s graded against itself and attributed to a “K₇ Wilson loop”), 1 untested forward claim. The published statement that the master table is computed “with no row-specific tuning” (Paper 19) is contradicted by the library’s own source (per-particle walk integers, corrected post-hoc; per-state rational coefficients) and is likewise corrected.

2.5 Independent audit verdict. The memory-blind Auditor (audit 2026-07-12-constants-provenance-disputes, verdict 4032d04; evidence: pinned clones only, SHA-256 manifest 13/13, no program memory or chat read) adjudicated the four disputed rows and refuted the DERIVED self-tag on every one:

The Auditor also fixed the claimed order at NLO: the NNLO α² bracket is documented target-selection — its coefficient computed from CODATA, then rounded to “the nearest structural integer” — and is retired from claim status (it may remain as code, but may never be cited as the claim). Both dispute exhibits (§2.2 look-elsewhere, §2.3 the measured-α G chain) were reproduced independently and were seed- and menu-robust. The single vulnerability the Auditor disclosed on the record: the four replacement tags lean on the program’s own retrospective admissions, so η_B’s POST_SELECTED in particular is closed by pinning the primary 17-formula sweep artifact — supplied as a dated appendix (audit deposit CLAIM.md Appendix A, 2026-07-12), which pins the memo and adds a reproducible enumeration (eta_B_menu_density.py) confirming the same menu fits ~55–63% of random targets at Planck precision.

2.6 Related retirements. The black-hole-cosmogenesis branch (Paper 22 and the AoE/Cold-Spot motivation) was retired separately on 2026-07-12 after a pre-registered, blind-reviewed joint null found the anomaly constellation not established (the one provenance-clean external leg: p = 0.16; the apparent disk alignment was a 2-of-12 selection artifact). An earlier program (vortex-vision) was retired 2026-07-07. All three nulls are documented in the case study (in preparation).

3. What is corrected, specifically

  1. G “inside the error bar” — withdrawn (§2.3). The honest figure is ~150 ppm / ~7σ outside, substrate-pure.
  2. “No row-specific tuning” — withdrawn (§2.4).
  3. “80 observables at 0.1% median residual” and all similar aggregate accuracy claims — reframed: the residuals are as reported, but they are postdictions whose agreement is (a) excluded at experimental precision where targets are sharp, and (b) expected under the measured look-elsewhere volume where they are not. No claim in the series should be cited as a confirmed prediction.
  4. “Zero free parameters” — withdrawn as an aggregate claim: the per-particle integer assignments, per-state rationals, and staged correction coefficients are free discrete parameters drawn from a menu whose fitting power is quantified in §2.2.

4. What survives — and, explicitly, what does not

A full salvage inventory was compiled at retirement under one rule of inclusion: a result survives if and only if it would be true had NWT never been proposed (SALVAGE_INVENTORY.md, to be published with the case study, in preparation). In summary:

Salvaged:

NOT salvaged (explicit, so nothing is resurrected by ambiguity — mirroring SALVAGE_INVENTORY.md §F):

5. Reopen conditions (frozen)

The program retires with its remaining forward claims frozen, not erased. To be precise about the relation to §4: these claims are not salvaged — per the inventory’s rule they are not established truths about anything — they are the pre-registered conditions under which the retired claims would earn re-examination. NWT would warrant reopening if and only if:

  1. A frozen K₈ dark-matter rung is observed (the per-rung integers N_e are pinned at retirement; any post-hoc integer choice voids this clause) — e.g. a ~2 keV sterile-neutrino line (XRISM) or a ~98 GeV direct-detection state (LZ-G3);
  2. A forcing chain is exhibited that derives a claimed constant exclusively — with no menu of alternative integers/coefficients — and demonstrably beats the measured noise-fit baseline of §2.2; or
  3. The S-FORWARD register (post-2018 measurement updates, append-only) shows sustained convergence of new measurements toward a frozen prediction. Its first entry (CODATA-2022 α) moved away.

6. Record

7. Method note and acknowledgment

This program was built in intensive collaboration with AI systems (Claude), which accelerated both the construction of the theory and — the case study’s central subject — the construction of the machinery that killed it. The failure modes documented here (confirmation-amplification, fit-then- interpret, yardstick equivocation, selection-as-signal) were catalogued from the inside. Readers evaluating similar AI-amplified theory-building programs may find the countermeasures more transferable than the physics.

The D12RG community’s on-record cautions — in particular regarding the rigor of the Chern–Simons functional-integral step — were correct, and are acknowledged with thanks.

8. Lineage note

NWT descended from the toroidal-electron family of models (Williamson & van der Mark’s confined-photon electron and its successors). Two facts about our own conduct toward that lineage belong in the record:

  1. Our early survey treated the family’s independent convergence on α-as-geometric-ratio as support. Under this retrospective’s accounting, that convergence is what shared apparatus freedom predicts — several researchers drawing small-integer/π/√ combinations near 137 from the same menu is multiplicity, not independence (§2.2). Convergence-as-evidence was one of our own catalogued failure modes, and it operated on us through this lineage.
  2. The one family mechanism we tested quantitatively — the Robinson charge-overlap model of nuclear binding — is falsified by that test: the mechanism yields net repulsion for the deuteron (+0.64 MeV predicted vs −2.22 MeV measured; no parameter in the scanned range binds it). The reproduction script (nwt_robinson_nuclear.py) is in the published analysis code (nwt-repro run nwt_robinson_nuclear). We report this as a result about a model we imported, not as commentary on any person or community.

We offer no verdicts on claims we have not tested. The instruments published with this retrospective — precision scoring at experimental uncertainty, input-hygiene checks, and the neighbouring-value look-elsewhere test — apply to any closed-form-fits-constants program, ours first of all. Anyone who wishes to audit a claims table, including their own, can.

— Jim Galasyn, July 2026

This page mirrors RETROSPECTIVE.md in the repository, which is canonical.