Lensemble MVP — Historical SO-100 Gauge Audit¶
Historical Evidence Status¶
These runs predate the correction of the outer-update direction and the public-probe-v2 target-binding contract. Their metrics are retained for audit history only and do not validate the corrected runtime; GitHub issue #335 tracks the replacement run.
This card preserves the immutable Phase 3 MVP measurements for provenance and failure analysis. It is not current correctness evidence and must not be used to claim that the corrected distributed training loop converges.
Recorded Run Shape¶
- Four participant silos and one disjoint held-out SO-100 split
- From-scratch latent model (
latent_dim=256,depth=8,image_size=224) - Simulated secure aggregation
- Relaxed-DP (DP-off) gauge-probe regime
Historical Metrics¶
| Control | Final effective rank | Final prediction loss | Final frame drift |
|---|---|---|---|
| Anchored federation | 14.815850257873535 | 22.19140625 | 83.59595115984035 |
| Naive FedAvg | 2.6838886737823486 | 203776.0 | 180.0 |
| Local only | 105.507 (mean) | 0.025432 (mean) | 180.0 |
Pinned immutable revisions: anchored-federation a6f5a96174f76af90ef3f4975bdda101e5ce6d45 · naive-fedavg cd8481c4c041d610008491f13e64bf092d3bc94d · local-only 9345bc3cf831238508e540981a15e8884acedec1
These numbers describe the superseded implementation. Under that implementation, the anchored configuration exhibited less gauge collapse than naive FedAvg, but the sign error in the outer update prevents treating the comparison as evidence for the corrected runtime.
Claim Boundaries¶
- Held-out magnitude collapse was approximately
~7.5e-6latent variance; the checked-in downstream and inference reports preserve that audit trail. The associated central ceiling diagnostic did not establish downstream usefulness. - In plain terms: skill_vs_identity is gameable; effective_rank is scale-invariant. Therefore, neither is a binding usefulness metric.
- Latent-MPC
success_rate=0.0is a negative result. - The SO-100 checkpoint is not a downstream-useful world model.
- No closed-loop physical SO-100 task success, paper-scale LeWorldModel performance, effective differential privacy, secure-sum optimizer integration, or cryptographic honest-computation proof is claimed.
- Issue #335 tracks a clean rerun using the corrected outer update and target-bound public probe.
Machine-readable source: phase3_mvp_benchmark_report.json.