kugguk/recaman-independent-check-bundle

Recaman independent-check bundle

This bundle contains inspectable numerical artifacts from commit ccc0b620a57c630550f96cbcf247e1e42bdd1edc. It is evidence for finite computational runs, not a proof that any integer is absent from the infinite Recaman sequence.

Contents

• embeddings/: numeric delay, spatiotemporal, and arc-lift coordinates as compressed NumPy files; parameters are in embeddings/metadata.json.
• models/: 5 load-verified joblib artifact(s) are in models/.
• evidence/: the rerun JSON/CSV results and exact obstruction catalogue.
• source/: the scripts and dependency declarations needed to inspect or regenerate the artifacts.
• environment.json: commit, catalogue checksum, Python, OS, and package versions.
• SHA256SUMS: integrity hashes for every other file in this bundle.

The saved models are fitted on all examples solely for deterministic artifact inspection. Do not quote their in-sample performance. Evaluation evidence is the forward-held-out or cross-validation AUC stored in evidence/.

2026-09-01 deep-obstruction frequency addendum

analysis/deep_obstruction_frequency/ is a self-contained, source-backed test of whether deeper catalogued obstruction runs become more frequent across equal multiplicative value scales after the first known hole at 852,655.

Across 24 equal-width log10(value) bins, event counts increase at every predeclared contiguous-run threshold (>=1, >=2, >=10, >=100, and >=1000). Spearman correlations range from 0.685 to 0.905; all five one-sided permutation tests remain significant after Holm correction (adjusted p <= 0.00060). The late half contains 6.48x as many total events and 33.67x as many events of length at least 1,000 as the early half.

This supports a catalogue-level pattern per equal multiplicative value interval. It does not establish the proposed visited-set-saturation mechanism: the catalogue contains neither Recaman landing-opportunity counts nor survivor times. The addendum includes the corrected 3,103-event Chaffin catalogue, machine-readable results, bin counts, a readable report, the exact script, and commit provenance.

The original bundle snapshot and its model artifacts remain unchanged. They predate restoration of the 852,655 singleton and should not be treated as models fitted to the corrected addendum catalogue.

Minimal verification

import joblib
import numpy as np

delay = np.load("embeddings/delay_tau2.npz")
print(delay["points"].shape)

artifact = joblib.load("models/version_c_D.joblib")
print(artifact["kind"], len(artifact["feature_names"]))

Verify file integrity with sha256sum -c SHA256SUMS on Linux/macOS or compare the listed hashes with Get-FileHash -Algorithm SHA256 in PowerShell.