This branch contains the reproducible code, experiment inputs, execution logs, mathematical report, integrity manifests, and curated frozen outputs from the 2026-08-26 gate-collision experiment.
The experiment tests whether a small number of exact within-expert SwiGLU gate
collisions can expose useful full-rank GL(k) gauge orbits while keeping the
induced routed-layer output error small. It does not perform quantization,
entropy coding, Fisher/KL evaluation, perplexity measurement, or task testing.
All audited k=2, k=4, and k=8 product maps retained practical rank k,
supporting a free k^2-dimensional GL(k, R) orbit at each stored point in
exact arithmetic. The adaptive count diagnostic found:
- five width-4 blocks: 60 additional off-diagonal tangent dimensions at 0.1834% validation and 0.1703% reused-fresh routed-layer RMS;
- four width-8 blocks: 224 additional directions at 0.4165% validation and 0.2334% reused-fresh RMS;
- adding expert 29 caused both reused-fresh errors to exceed 1%, with one sequence accounting for more than 99.4% of its positive marginal error.
These operating points are hypothesis-generating, not confirmatory: both replay sets have been inspected during the evolving research program.
See the full diagnostic for the mathematics, complete statistics, confidence intervals, limitations, and next experiments.
code/— all Python and shell sources used in the experiment.inputs/— hypotheses, method proposals, earlier selections, and frozen endpoint inputs.logs/— captured execution logs from the persistent RunPod volume.docs/— the consolidated follow-up report and preceding minimization report.results/— report-critical frozen result evidence.manifests/— the complete persistent-tree, model, and cache SHA-256 manifests plus a manifest for this Git package.
The full follow-up tree contains 4,875 files and 11,249,396,003 bytes. Its
large intermediate .pt candidate/chunk trees are not suitable for ordinary
Git storage. The BF16 checkpoint and activation captures are also intentionally
excluded. Nothing is silently unaccounted for: their complete paths, sizes,
and SHA-256 values are retained in manifests/.
The authoritative full experiment tree remains on persistent RunPod network storage at:
/workspace/gauge_collision_followup_l20_20260826_v1
See ARTIFACTS.md for exact inclusion coverage and integrity details.
From the repository root:
sha256sum -c manifests/GITHUB_PACKAGE_SHA256SUMS