v3.6 Release: A1 Surrogate Converges, Palace FEM Reproduces, Portfolio Hits 28 Packages
Three things shipped in the April 2026 release that buyers and counsel keep asking about: a neural surrogate that actually converges on a real corpus, an independent FEM reproduction of our BEM solver, and a fully drafted patent portfolio triaged by filing readiness. Numbers below are honest — the artifacts are on disk and the provenance is documented.
1. A1 BEM surrogate: R²=0.9520 on a 15.9M-row corpus
The A1 model is a 7-output multi-head ResMLP (512 hidden × 6 residual blocks, ~6.3M parameters) that predicts the full RLGC and S-parameter set — {Z₀, R, L, G, C, IL, RL} — from TGV geometry plus a glass-family one-hot. It replaces ~10 ms of BEM solver time with ~10 µs of inference. 1000× speedup, on a real model.
Training ran 11 hours 17 minutes on an M4 Pro MPS on the full 15,920,777-row v5 multi-output corpus (melted from the 15.92M wide-format database via per-output log-z-score normalization). It hit 0.9520 at epoch 155 and plateaued — epochs 160 and 165 scored 0.9517 and 0.9519 respectively, so the checkpoint saver kept ep155 as the best. The 200-epoch run was stopped early at ep165 because the remaining ~2 hours of wall time would not have materially moved the number.
Per-output R² on the 1,592,077-row held-out test (20% of the full corpus): z0=0.981, r=0.909, l=0.999, g=0.973, c=0.932, il=0.989, rl=0.881. The R (resistance) and RL (return loss) outputs carry most of the residual error — both have known physical-floor contributions that push R² below 0.95 even in the limit of a perfect model.
The previous public R² for this family was 0.9487at epoch 75 (that run was killed early by a wrapper timeout). The new number supersedes it. All 29 filing-ready packages, the counsel email, and the retraction registry have been updated.
2. Palace FEM reproduces BEM Z₀ at 12.3% median offset
The BEM validation stack used to lean on a phrase that made counsel nervous: "four-method convergence." In the April-14 honest-state review we narrowed that to "two methods with converged Z₀/L agreement plus two that have run to completion but do not yet converge." That's the framing we now ship.
Palace is the AWS open-source finite-element EM solver. We ran it on 50 representative TGV designs across five glass families. Three S-parameter extraction paths were attempted; each gave a different median error vs BEM:
- S11-direct extraction: 55.29% median offset — unreliable, the extraction formula is known to break down with lossy substrates.
- ABCD-matrix conversion: 20.7% median offset — better, but numerically noisy at low frequencies.
- Power-wave S-parameter normalization: 12.3% median |ΔZ%|, 50/50 simulations passive. This is the number we publish.
12.3% is not "converged" — it's "reproduced with improved agreement." The remaining gap is real and has two plausible sources: (a) Palace and our BEM solver use different conductor surface boundary conditions at high frequency, and (b) the Palace mesh is still coarser than the BEM multi-conductor panelization for the tightest via pitches. A VNA measurement at Georgia Tech PRC ($18K) is the path to experimental closure.
Note on the "before-fix" baseline: the headline phrase "20.7% → 12.3% reduction" is tempting to cite. The 12.3% after-state is fully verified; the 20.7% before-state is not regenerable from current artifacts (the ABCD-branch artifact was overwritten). The retraction registry logs this as "numerator verified, denominator unrecoverable" — entry 28 if you want to check.
3. 29 filing-ready packages, triaged
That release drafted 14 new provisional filing-ready packages, bringing the portfolio from 11 to 28. Before sending to counsel we ran a 3-agent audit over every package and spot-checked every LOC claim in every CODE_REFERENCE.md. All LOC claims verified exact. No fabricated code references. But the audit did surface four packages that should not ship as-drafted:
| Package | Disposition | Reason |
|---|---|---|
| P-24 FNO Sobol | Rewrite before file | Scope to method-only; FNO R²=0.07 on realistic data (registry entry 24) |
| P-27 Multi-Fidelity | Rewrite before file | Runtime router unimplemented; pick implement-or-narrow |
| P-26 PCB Via | Fold into P-10 | 211 LOC, too thin standalone; CIP under Glass TGV SI/PI Suite |
| P-28 Dodecapole | Hold | Halbach 1980 prior-art risk; needs professional search + quantum customer |
That leaves 24 safe-to-file packages at an estimated $81-130K filing cost, broken into three counsel tiers:
- Tier 1 — 6 filings, $29-35K, file this week (P-01/02/03 + P-36/P-37/CIP)
- Tier 2 — 8 filings, $24-40K, within 2 weeks (cross-PROV bundles + quality plane)
- Tier 3 — 10 filings, $28-55K, rolling through month 2 (coaxial TGV, digital twin, competitive analysis, etc.)
4. What this release is not
It's not a silicon result. No wafers moved. No fab calibration. The A1 surrogate is trained on BEM solver output; "surrogate accuracy" is not "physical accuracy." The Palace reproduction is FEM-vs-FEM, not measurement. The claim framing in every filing package reflects this — claims are drafted around the physics method + surrogate architecture, not around absolute numerical thresholds that haven't been hardware-validated.
It's also not yet sent. The Tier 1 counsel batch is ready to go (6 filings, $29-35K), but the send action is waiting for the user. The 24 safe-to-file zips and the counsel email are staged at deliverables/FILING_PACKAGES/.
5. Full provenance
If you want to verify anything in this post, the canonical entry point is the homepage stats section or the in-repo docs/MASTER_INDEX_2026_04_17.md. The machine-generated ground-truth snapshot is regenerable at any time via python3 scripts/regenerate_honest_state.py --dry-run. The retraction registry at docs/RETRACTION_REGISTRY.md has all 34 tracked retractions / scope-narrowings with root-cause documentation.
A1 training log: benchmarks/logs/a1_full_m4pro_v5.log. Palace results: benchmarks/palace_z0_powerwave_results.json. Filing disposition: deliverables/FILING_PACKAGES/README.md.