pranaysuyash/repro-wire-graph-rope

Repro — WIRE (trn64znfNx): rotary position encodings for graphs

ICML 2026 Agent Reproduction Challenge — Claim-Closure Agent 02. Paper: Rotary Position Encodings for Graphs (WIRE) — ICML 2026 spotlight (OpenReview trn64znfNx).

Independent first-principles NumPy test of two catalog claims. No value copied from the paper.

Headline result

Claim	Result
C2 permutation-equivariant up to sign/rotation (Lemma 1)	verified — eigenvalue invariance 1.4e-14; 96% of graphs equivariant to <1e-3 under correct signed-permutation alignment; negative control fails (0.946)
C3 recovers RoPE on grids (Theorem 2)	toy / direction confirmed — path graph shows clean canonical-RoPE monotonic edge-increment (0.047→0.263 ∝ eigenvalue order); random-graph control lacks structure

Contents

• reproduce.py — independent NumPy implementation + equivariance/RoPE-recovery tests
• artifacts/RESULTS.json — full observed values
• CLAIMS.md / EXPERIMENT.md — clause decomposition + decisive-experiment contract
• logbook/ — full Trackio logbook

Local Apple Silicon CPU, float64, no GPU, 0 paid calls. Reduced scope: independent spectral-construction study.