How it works

Two hands, and a firewall between them.

Most engineering-AI systems let the same model that proposes a design also score it. That is how mirages ship. PIE splits the two roles at the architecture level and forbids either side from crossing.

The generative half

Evolutionary search proposes geometry: CPPN networks, field-lattices, L-systems, hybrid genomes, MAP-Elites archives. Neural surrogates and LLMs may suggest and route — a 968,761-theorem semantic math shelf and cited engineering references feed inspiration — but nothing on this side can promote a candidate.

Authority policy, enforced in code, not in a style guide:

The authority policy enforced in code: each row pairs a policy flag with the value it is set to, showing which parts of the system may promote a candidate and which may not.
`physics_gates_promote`	true
`request_text_promotes`	false
`scouts_promote`	false
`llm_narration_promotes`	false
`surrogate_screening`	observe_only

The verification half

Real solvers certify: CalculiX finite-element analysis (static and eigenfrequency) on Gmsh tetrahedral meshes, D3Q19 lattice-Boltzmann pore flow, and in-house cores that are themselves benchmark-validated —

the 2D structural core matches an external topology-optimization reference within 1.66% recompute-verified
the in-house hydraulic solver matches real EPANET within 0.0143 m across the published Hanoi network, and PIE's evolved Hanoi design lands within +0.269% of the best-known published cost recompute-verified

The evidence-tier ladder

Every number PIE produces carries a tier, and the tier caps the claim:

The three evidence tiers. For each tier the table gives its name, what kind of evaluation it is, and the strongest claim a number carried at that tier is allowed to make.
Tier	What it is	What it may claim
analytical_proxy	Fast in-loop screening models that drive the search	Advisory only. Never published as a result.
regime_specific_solver	Validated in-house physics cores	Claim-grade within their benchmarked regime
validated_sidecar	CalculiX / Gmsh on the real geometry	Claim-grade.

A proxy may drive the search. Only the real solver's numbers may be claimed. When the two disagree, the proxy is wrong by definition — and the disagreement itself is measured and recorded. See what that looks like in practice.

The pipeline

The five stages of the pipeline in order. The left column names each stage; the right column describes what happens there.
GENERATOR	CPPN / field-lattice / hybrid genomes propose candidates
FIREWALL	suggestion may pass down; judgment may never pass up
SOLVERS	CalculiX + Gmsh + LBM + validated cores evaluate every declared load case
HARD GATES	worst case governs; every gate proven able to fail before the run counts
LEDGER	append-only, content-addressed record of every certification — passes and failures

See what this produced The full honesty apparatus