PIE·INDUSTRIAL
PIE — Personal Invention Engine

Invention you can check.

PIE is an evolutionary invention engine where physics is the only judge. It evolves engineering geometry from scratch — and it is architected so it cannot lie about the results. Every solver result on this site traces to a hashed artifact, and the manifest that proves it is published.

See the evidence What it can discover The run that failed honestly
Claim-grade result
SF 2.055vs0.427

CalculiX-certified declared envelope

An evolved AM payload bracket carries a declared 200 kg-payload rotorcraft load envelope where the conventional plate-and-gusset design fails torsion outright. Same solver, same material, same load cases. It is heavier than the conventional part — 1.165 kg vs 0.946 kg. We say that plainly, because the alternative is you finding out later.

The receipt on the right is not a mock-up. It is the certification record for that bracket — CalculiX static + eigenfrequency on a Gmsh mesh of 31,487 nodes / 126,665 tets. Full case study →

certification record certification.json ✓ PASS
{
  "candidate_id": "bracket_g00095_21d1fa03afc24",
  "evidence_tier": "validated_sidecar_calculix_gmsh",
  "material": "AlSi10Mg",
  "governing_case": "torsion",
  "governing_safety_factor": 2.0554,
  "mass_kg": 1.1647,
  "first_mode_hz": 171.58,
  "mesh": "gmsh · 31,487 nodes / 126,665 tets",
  "passes": true,
  "failures": []
}
ledger record e9ff00d8… · append-only · content-addressed
Architecture

Two hands, and a firewall between them.

01 · Generator proposes

Evolutionary search over CPPN networks, field-lattices, hybrid genomes. LLMs and surrogates may suggest and route. They are firewalled from ever judging fitness.

02 · Physics disposes

CalculiX FEA on Gmsh meshes, lattice-Boltzmann pore flow, validated in-house cores. Only real-solver results are claim-grade.

03 · Receipts survive

Preregistered claims, an append-only ground-truth ledger, and a compiler that machine-traces every published number to a hashed artifact.

The same engine and the same firewall run at every scale — crystal microstructures, architected lattices, printable components, water and power networks, coupled machines, dynamical systems. The full range, with each family's evidence tier →

How the architecture works
Curved Inconel 625 annular-sector test coupon, rendered from the print-ready STL
Inconel 625 annular-sector coupon — rendered from the print-ready STL.
Also on the bench

A combustion-chamber sector coupon, print-ready.

Built to a turbojet OEM's publicly documented envelope: CalculiX FEA (axial E 136.72 GPa, first yield 165.01 MPa), converged pore-flow LBM, watertight print files. Its report's 10 headline claims machine-compiled ALL_TRACED.

Case study
Getting better, measurably

Nine campaigns. The search got better each time.

5.49 kg4.54 kg2.40 kg

validated_sidecar not a baseline beat

Across nine real-solver-certified campaigns on the same declared engine-support bracket contract, PIE's own best safety-passing design got 47.2% lighter campaign to campaign, and 56.3% lighter within a single 12-hour run — every point CalculiX/Gmsh certified. The declared naive baseline for this contract still fails outright (SF 0.748 against 1.5 required), so there is nothing passing to beat. This is the search improving against itself, stated as exactly that. Full case study →

Honest failure

The proxy promised twelve winners. The solver failed all twelve.

The proxy's champion bracket — the real solver failed it at torsion SF 1.08
The proxy's champion, rendered from its certification STL. Predicted SF 1.586 proxy — advisory — the real solver failed it at SF 1.082 0 / 12 passed

A 1,063-generation run, driven by a fast in-loop proxy, sent its twelve best candidates to CalculiX certification. Real torsion safety factors came back 0.97–1.28, all below the 1.5 bar. Nothing shipped.

Then the system turned on its own proxy:

  • measured it 40.1% optimistic — 12 of 12 candidates over-promised
  • found its reported accuracy understated ~81× against reality
  • found its confidence band ~48σ too narrow
  • filed all 12 failures in the failure museum as evidence for the next run

Most engineering-AI pipelines would have published the mirage. PIE's architecture exists so it can't.

The full story, with numbers
The honesty apparatus

Held to its own rules.

PREREGISTRATION

Envelope, load cases, safety bar, baseline and the exact claim ceiling — frozen before iteration 1, so the claim cannot be adapted to the result.

GROUND-TRUTH LEDGER

Append-only, content-addressed record of every real-solver certification — passes and failures. Proxy results cannot enter by construction.

CLAIM COMPILER

Every outbound number maps to a JSON path in a hashed artifact and is machine-checked: ALL_TRACED, or named failures. No manifest, no publication.

BLIND TWO-LANE REVIEW

An independent verification lane re-runs the rulers itself rather than accepting reported numbers. No result grades itself.

The full apparatus — nine mechanisms

Bring a component you care about.

Your load envelope, your boundary conditions. PIE evolves and solver-certifies candidates; you get print-ready geometry and the full evidence chain.

How collaboration works Email directly