Nine campaigns. The search got better each time.
solver run · check these numbers against their artifacts
A different declared contract than case study A: an engine-support bracket, not the 200 kg payload bracket. Over nine real-solver- certified campaigns run between 2026-08-05 and 2026-08-09, PIE's own best safety-passing design for this contract went from 5.49 kg to 4.54 kg to 2.40 kg — every one of those numbers CalculiX/Gmsh certified, not proxy.
Read this before the numbers below: there is no baseline beat here. The declared naive design for this contract — a conventional plate-and-gusset support — is itself CERTIFIED FAILING at 0.953 kg, safety factor 0.748 against a required 1.5. There is no passing baseline to beat on mass. What follows is evidence that PIE's search is improving release over release at finding lighter designs that still pass — measured against itself, not against a working reference design.
The trajectory, every point real-solver certified
| Candidate | Mass | Governing SF | Certified | |
|---|---|---|---|---|
| Declared naive baseline | plate-and-gusset | 0.953 kg | 0.748 — fails | validated_sidecar |
| v9's own first certified pass (12h run, gen 24) | bracket_g00000 | 5.485 kg | 1.833 — passes | validated_sidecar |
| Prior campaign's certified winner (v8, 8h run) | bracket_g00062 | 4.539 kg | 1.730 — passes | validated_sidecar |
| v9's final winner (12h run, gen 203) | bracket_g00087 | 2.397 kg | 1.759 — passes | validated_sidecar |
Every row: CalculiX static + eigenfrequency on a Gmsh tetrahedral mesh, same declared load cases, same AlSi10Mg material-library entry. 47.2% lighter than the prior campaign's winner (v8 → v9); 56.3% lighter than this same run's own first certified pass (gen 24 → gen 203) — and still 2.52× heavier than the baseline above, because that baseline fails and cannot be beaten, only avoided.
The winner's receipt
{
"candidate_id": "bracket_g00087_ae6c49af05444",
"evidence_tier": "validated_sidecar_calculix_gmsh",
"material": "AlSi10Mg",
"governing_case": "engine_torque_216Nm_declared_1.5x",
"governing_safety_factor": 1.7592,
"mass_kg": 2.3971,
"first_mode_margin_ratio": 0.5767,
"mesh": "gmsh · 50,752 nodes / 227,765 tets",
"passes": true, "failures": []
} No render yet — like case study A's winner, this candidate's STL wasn't retained after certification, only its solver outputs. A render will appear here when the geometry export is regenerated.
What this is, and isn't
This campaign has not yet been through the audit process that promotes a result into the discovery ledger — it exists today as raw, real-solver-certified campaign artifacts, not yet a ledgered discovery record. We're publishing it anyway, at this stage, because the honesty apparatus applies the same way to an in-progress trend as to a finished claim: every number above traces to a hashed CalculiX artifact, and the caveat that matters — no baseline was beaten — is stated first, not buried.
Back to evidence Bring us a contract like this one