Borbax

Chemical evolution — abiogenesis up to protocells — in a universe whose chemistry is invented from a 64-bit seed, not modelled on the real world.

Molecules are shapes. Two of them react when their shapes fit together and their surface characters oppose. Catalysis, membranes and heredity are not implemented; they are downstream consequences of that one test.

The premise

Real chemistry is, at the level that matters for life, about shape. Borbax takes that literally and refuses to model anything else:

For each of the 60 rotations R in the icosahedral rotation group:
    for each direction i:
        j = ANTI[ PERM[R][i] ]
        shape_term  = −( r_A[i] + r_B[j] − IDEAL_GAP )²      # bumps meet hollows
        charge_term = −( a_A[i] + a_B[j] )²                  # + faces meet −
    score(R) = Σᵢ ( w_shape · shape_term + w_charge · charge_term )

affinity(A,B) = max over R of score(R)

That function is the entire mechanism. Two molecules sticking together, an enzyme recognising a substrate, a membrane self-assembling, and a pore gating what crosses it are all the same call at a different scale. If a feature seems to need a second mechanism, the feature is wrong.

Three things that make it unusual

The chemistry is invented, and that is enforced. There is no periodic table in this repository. Elements, valences, bond energies and folding rules are all generated from universe_seed, so a Borbax universe has its own chemistry that nobody — including us — knew in advance. A CI gate rejects real element names, real chemical file formats, and any data file in a chemistry crate. Temperature is measured in thermals, energy in quanta, distance in spans: distinct types, and mixing them does not compile.

Nothing about life is hardcoded. No struct Cell, no replication rule, no "if conditions are right, spawn a protocell". Catalysis is not implemented — a folded chain that happens to have two well-placed cavities holds two reactants adjacent, and the rate enhancement falls out of the geometry. Protocells and death are detected, via autocatalytic-set closure, never declared. A special case for life would mean the physics is wrong.

The same seed gives the same universe, everywhere. Bit-identical on macOS/aarch64, Linux/x86-64 and Windows/x86-64, enforced by a CI matrix that diffs golden state hashes across all three. That means counter-based RNG, no HashMap iteration in any result path, pinned float accumulation order, fixed-point mass so conservation is exact by construction, and every transcendental routed through one portable implementation — because exp and cos genuinely differ between platform libms, and a world you cannot reproduce is a world you cannot share.

Status

V0, in progress — Tasks 1–2 of 21 complete. The workspace, toolchain and fiction-guarantee gate are in; so are the invented-unit types and the portable transcendental chokepoint. Universe generation is next.

V0	Chemistry in a test tube: universe generation, molecules, shape and binding, folding, decay, a beaker harness, SVG rendering. No world, no live viewer.	← here
V1	One focused region, multi-scale engine, keyframes, neutral shadow, the Chronicle. Target: protocells.	The hard part
V2	Template-copying polymers, mutation, true heredity.	Life gets a genome
V3	Cells, ecology, a planetary field.	The world gets big

Genomes are deliberately deferred: V1 succeeds or fails on its own criteria, however tempting V2 becomes.

Getting started

proto install          # toolchain, from .prototools
cargo xtask setup      # prek + git hooks, once per clone
cargo test --workspace
cargo xtask            # fiction-guarantee checks

CLAUDE.md has the full quality gate and explains why each command is on it.

Layout

crates/            the simulation, in dependency order — ✅ exists, ○ planned
  ✅ borbax-units      invented units + the portable-transcendental chokepoint
  ✅ borbax-rng        counter-based deterministic streams
  ✅ borbax-universe   generated elements, bond energies, constants
  ○  borbax-molecule   graph → canonical form → 3D embedding → signature → binding → fold → cavity
  ○  borbax-reaction   reaction classes and rates
  ○  borbax-beaker     the well-mixed reactor
  ○  borbax-geometry   → borbax-render (headless SVG, golden-tested)
  ○  borbax-cli        beaker, sweep, band, battery, render, goldens
experiments/       measurement harnesses that answer a question and then stay as evidence
xtask/             the fiction-guarantee gate
docs/
  superpowers/specs/    the PRD — numbered sections (§n) are cited everywhere
  superpowers/plans/    the V0 implementation plan, Tasks 1–21
  experiments/          measured results, with their methods and near-misses

Crates later in that list depend on earlier ones and never the reverse. The planned ones arrive in plan order; nothing is stubbed ahead of time.

How the maths works, per crate

Each crate has a README written for a reader who wants to understand the ideas rather than the API — what problem the maths solves, why that approach and not the obvious one, and where the constants came from. They assume no background and cite sources.

README	What it explains
crates/borbax-rng	Why random numbers are computed from a counter rather than shuffled, and how Philox scrambles them
crates/borbax-units	Units the compiler enforces, why mass is an integer, and why exp is banned
crates/borbax-universe	Elements as packed spheres — the shell law, the binding peak, and where valence and bond energies come from
experiments	The measurement harnesses, locality, and the neutral shadow
xtask	The fiction-guarantee gate and why it is a program rather than a checklist

Reading order

1. docs/superpowers/specs/2026-07-26-borbax-prd.md — the design, and why each decision is the way it is. §3 is the principles; §8 is the chemistry; §13 is the determinism contract.
2. CLAUDE.md — the working agreement: hard invariants, review precedence, and how work reaches main.
3. docs/superpowers/plans/ — the task-by-task plan, with the "as built" notes recording where execution disagreed with the plan and who won.

Two claims this design refuses to make

Both come from the prior-art survey, and both are places where the field's consensus is weaker than it looks:

• Never assume self-maintaining units compose into higher-order ones. It is the most-cited claim in artificial chemistry and its weakest result. V2 onward treats composition as an open question rather than something that will fall out.
• Never declare a plateau by eyeballing a flattening curve. Fit competing models and compare, and read every novelty metric against a neutral shadow run — otherwise "look what evolved" is a person looking at a graph and feeling optimistic.

Licence

MIT — see LICENSE. [workspace.package] license already said so; this is the file that makes it true.