Research Findings
Validated findings from Phase 3 experiments (F1–F8). Each finding includes the claim, supporting evidence, Tarka verification status, and artifact paths. Results are gated by independent recomputation and sign-off.
Filter by category:
MMD converges linearly to its magnetic fixed point where GDA cycles
Claim
On symmetric zero-sum matrix games, fixed-magnet MMD (uniform anchor, lr=0.1, τ=0.1) exhibits linear (geometric) last-iterate convergence to its fixed point, while simultaneous GDA at the same stepsize cycles, bounded away from equilibrium.
Evidence
10 md5-distinct seeds, 2000 steps (exp01 iteration 2): log-linear fit on distance-to-fixed-point, last 50%: R²=0.9948 (matching pennies), 0.9015 (RPS). GDA final NashConv: 1.93 [1.90,1.96] (MP), 1.76 [1.62,1.88] (RPS), R²<0.09 (no decay).
Key Numbers
Verification & Status
Artifacts
Uniform-anchor MMD fixed points ≠ logit-QRE for asymmetric games
Regularized Nash Dynamics reach Nash universally with periodic reference resets
DEQ peak activation memory is O(1) in effective depth
Anderson acceleration beats Picard exactly where theory predicts: stiff fixed points
Second-price token auction is exactly truthful; weighted aggregation is manipulable
Warm-started homotopy accelerates QRE path tracing
Undamped logit-QRE fixed-point iteration requires damping beyond small λ
About Tarka Verification
Tarka is an independent verification system that recomputes findings and audits experimental config hashes, seeds, and git commits. All findings in this explorer have undergone Tarka review; VALIDATED status means the claim has been independently reproduced. SIGN-OFF PENDING means the verification is complete but awaits final operator signature.
Every result includes:
- Config hash (SHA-256) for parameter reproducibility
- Random seed pinning for deterministic runs
- Git commit reference for code versioning
- Artifact paths for accessing raw data and logs