An autonomous design loop for positional molecular assembly, built in public. Looping coding agents as harness, physics engines as arbiter, domain experts with veto. Aimed at one sub-capability of the unclaimed Feynman Grand Prize — not at winning it. Failures published alongside results.
cheiron proposes candidate positional-assembly reaction steps — a tooltip
that, under positional control, makes or breaks a specific bond at a specific
site — and submits each to physics engines that judge whether it is:
- favorable — thermodynamically downhill,
- feasible — reachable over a low-enough barrier under mechanical approach,
- selective — hitting the intended site only, with the tool surviving.
The one sub-capability we own is designing and computationally validating those
steps. We are not building hardware and not trying to win the prize. See
docs/design/00-goal-and-scope.md.
PROPOSE ─▶ BUILD ─▶ ARBITER ─▶ SCORE ─▶ SELECT/EVOLVE ─┐
▲ (physics engine) │
└──────────────── VETO (human) ◀─────────────────────┘
every candidate ─▶ append-only ledger (successes and failures alike)
Architecture: docs/design/01-loop-architecture.md ·
Arbiter: docs/design/02-arbiter.md ·
Milestones: docs/design/03-milestones.md
M0–M2 complete; M3 (search, not enumeration) in progress. Highlights, all produced unattended by the loop and pushed as they happened:
- Favorability solved for the family: 40/40 tool×workpiece candidates measured (5 tools × 8 workpieces, UKS/PBE/def2-SVP), zero unresolved failures; reaction energies decompose exactly (Hess's law), so the additive model now predicts and the arbiter only verifies.
- Feasibility is per-tool-family: relaxed approach scans give a hybrid-grade barrier map that tracks literature activation energies (hydroxyl 1.8 vs exp ≈1.7; methyl 10.6; amino 8.4; vinyl 6.7; ethynyl 0) — and ΔE alone cannot rank tools kinetically.
- Selectivity is positional, not chemical (the project premise, measured): on adamantane — the diamondoid surface model — the thermodynamic site margin is 1.14 kcal/mol and the kinetic margin under clamped positional control is zero. Whichever C–H the tool is held over reacts.
- First step datasheet (criterion S2, VETO-pending):
docs/datasheets/habs-adamantane.md. - Handle-mounted tools work: an ethynyl tip on an adamantyl frame costs only 0.8 kcal/mol of driving force vs the free radical — cheap surrogate screening is predictive of realistic tooltips.
- A second operation (M4): radical addition. The loop is not hardwired to hydrogen abstraction — it now also does a bond-forming step (a radical adding across a C=C), characterized to the same depth: favorability (validated vs known ΔE), approximate additivity, a certified PBE0 barrier (methyl+ethylene 3.84 kcal/mol), and anti-Markovnikov regioselectivity (3.8 kcal/mol). Two findings only two operations could give: tool ranking is operation-dependent (abstraction strength doesn't predict addition strength), and abstraction needs the machine to pick the site while addition has real intrinsic regiochemistry to lean on.
The narrative, including every failure and correction, lives in
JOURNAL.md; raw append-only records in
experiments/m0_hydrogen_abstraction/results/ (abstraction) and
experiments/m1_radical_addition/results/ (addition).
# one-time environment (no system packages required)
python3 -m venv .venv --without-pip
curl -sS https://bootstrap.pypa.io/get-pip.py | .venv/bin/python
.venv/bin/pip install -e .
# fast pipeline smoke test (rough, small basis)
.venv/bin/python experiments/m0_hydrogen_abstraction/run.py --workpieces methane --fast
# the M0 result run
.venv/bin/python experiments/m0_hydrogen_abstraction/run.py --workpieces isobutane
# a second-operation run: radical addition (M4)
.venv/bin/python experiments/m1_radical_addition/run_addition.py --tool ethynyl --substrate C2H4
# tests (no quantum chemistry needed)
.venv/bin/pytestThis is built in public: the narrative is in JOURNAL.md and the
prompts steering the project are archived under history/prompts/.
GNU AGPL-3.0-or-later — see LICENSE.