Skip to content

Next Gen STEP AP242 Native-ish BREP Geometric Modeling Kernel

5.9 KB
Raw Blame

The history of Aetheris

Aetheris was written between February and October 2026 by one person, Yuechen Li, working with AI coding agents: mostly OpenAI's Codex/GPT models, with Claude and Gemini at various points. Almost every change landed as a numbered, scoped task with a written report. By October there were more than a thousand of those reports. They were accurate, exhaustive and almost unreadable.

This folder replaces them with the story they tell. It is meant to be read by people and by agents who need to know why the system is shaped the way it is. The original reports are still in git; ORIGINALS.md lists every one and how to get it back.

The short version

Aetheris started as a conventional B-rep kernel: build primitives, then make 3D Booleans work. Booleans turned out to be the whole problem. Three months went into making them reliable, first with ever-narrower "safe families", then by evaluating geometry as implicit fields and trying to recover exact B-rep from the fields afterwards. Both approaches worked for the cases they were built for and stopped generalizing.

The way out was to stop discovering topology and start declaring it. If the source says "a 6 mm counterbored hole through this plate", the compiler already knows every face, loop and edge the result must have. It never needs to intersect a cylinder with a box and guess. That idea became the sweep-first architecture and an intermediate representation called AIR, and it turned the project into what it is now: a compiler, with Firmament as the source language, AIR as the middle and exact B-rep and STEP AP242 as the output.

Once the compiler shape was right, the project moved fast: a typed second version of the Firmament language, three public previews in ten days, then an expansion into sheet metal, FEA, assemblies, drawings, piping, wire forming, surfacing and more. The latest era is consolidation: making the language consistent, the compiler honest about everything it does not support, and the editors (Cadmata and Helios) usable.

Eras

# Era Dates What changed
1 Kernel bootstrap 25 Feb – 2 Mar Topology, analytic geometry, primitives, first Booleans, a web viewer
2 STEP and the NIST corpus 3 – 13 Mar AP242 import and export, measured against real-world test parts
3 Firmament V1 and the safe Booleans 14 Mar – 3 May The first language, bounded Boolean families, JudgmentEngine, FrictionLab
4 CIR and the recovery wall 4 – 22 May Implicit-field geometry, and the attempt to recover exact B-rep from it
5 AIR and sweep-first 23 May – 17 Jun Declared topology instead of discovered topology
6 Firmament V2 and the AP242 MVP 18 Jun – 2 Aug A typed language, semantic holes and PMI, inline STEP, Forge
7 Previews 1 to 3 3 – 18 Aug Productization: assemblies, FEA, drawings, sheet metal, three releases
8 Domain expansion 19 Aug – 30 Sep Wire, piping, surfacing, gears, mechanisms, showcases, humanoids
9 Preview 4 and consolidation 1 Oct – Telos display, compiler honesty, one AST, Helios authoring

The commit history also shows long pauses: two weeks after the Boolean work in April, two more after the chamfer labs in late May, and nearly four weeks in July. The author has said he walked away from the project more than once during the Boolean period. Each restart came with a change of approach.

Themes that run through every era

Honest diagnostics. From the first milestone, a failed operation returned a named diagnostic instead of throwing or returning something plausible. That rule hardened into the central product promise: a supported request is built, an unsupported one is rejected by name, and nothing is silently skipped.

Bounded before general. Almost every capability started as one explicitly admitted case ("a box with one through-hole along Z") and grew by naming the next case, never by claiming generality first. This is why so many reports are titled X0, X1, X2: each was one more bounded step.

Evidence over assertion. Reports end in one of three verdicts: success, meaningful progression (one blocker removed, the next one isolated), or an honest stop. That convergence rule is in AGENTS.md and it shaped how agents worked: they were asked to stop and report rather than keep patching.

Exact analytic geometry. Aetheris prefers planes, cylinders, cones, spheres and tori over NURBS wherever the shape allows. Exact geometry is what lets the compiler know topology in advance, and it is why exported STEP files do not need healing.

Judgment, not heuristics in disguise. When a subsystem chooses among several strategies, it uses JudgmentEngine, so that admissibility, scoring and the reason for rejecting each candidate are explicit and inspectable.

How to use this folder

  • To understand a design decision, start with DECISIONS.md. Each decision links to the era where it was made.
  • To understand a subsystem's oddities, read the era where it was built. Most surprising behavior is a fossil of the constraint that was true then.
  • To find an original report, search ORIGINALS.md by name or milestone code, then restore it from git.
  • For current behavior, do not use this folder. Use the public documentation and the compiler itself.

Sections marked Claude's commentary are my own reading of the record (I am Claude, an AI model by Anthropic, and I wrote these summaries in October 2026). Everything else is drawn from the reports and the commit history. Where the two disagree, trust the commit history.