AGE turns an engineering goal into a plan, executes it with sandboxed skills, and refuses to declare success until an independent verifier agrees. In the software domain that means running the test suite; in the physics domain it means a closed-form answer cross-checked by a numeric simulator written independently of the training signal, and a PhysFormer that predicts answers from the parameters alone. This repo is the project root: the agent, the physics core, both the NMI and IEEE manuscripts, and the committed data behind every number. The component studies — physics transformers, PhysBench, the verification gate, the loss channel, few-shot law acquisition, field consistency — live in their own repos, linked below.
The system
Give AGE a goal and it runs the loop: plan, act, verify, reflect, journal. Two brains — a deterministic mechanical brain that works with zero API keys, and an LLM brain behind any OpenAI-compatible endpoint. Writes are confined to the working directory; destructive commands are refused. Every failed verification writes a lesson, so the next run starts smarter.
the loop is the heart: every step can carry expect: 'ok', and a failed verification fails the mission and writes a lesson into the journal.
every physics design is computed twice, independently: closed-form solution cross-checked by Euler/RK4 or finite-difference simulators written from a different formulation.
49 physics tests (closed forms, verifiers, the 10-law set) + 16 node agent tests. The IEEE paper and NMI paper both compile clean.
The physics core
physx/ is a physics-informed engineering core: projectile, pendulum, spring, beam, cantilever, RC, damped oscillator, Kepler, LC circuit, linear drag, Burgers, and 2D heat — each with an exact closed form, an independent numeric verifier, and a PhysFormer head trained on exact trajectories with the governing-equation residual in the loss. The multi-law protocol is what the component papers dissect: 10 laws, one shared body, 3 seeds, pre-registered.
The series at a glance
| Law | real | dummy | benefit |
|---|---|---|---|
| beam | 0.047 | 0.289 | +0.84 |
| cantilever | 0.129 | 0.252 | +0.49 |
| projectile | 0.102 | 0.129 | +0.21 |
| pendulum | 0.050 | 0.056 | +0.11 |
| spring | 0.091 | 0.088 | -0.03 |
| rc | 0.099 | 0.115 | +0.14 |
Reproduce
git clone https://github.com/sehajr-singhs/AGE-artificial-general-engineer cd AGE-artificial-general-engineer npm test # 16 node tests python3 -m unittest physx.test_physx # 49 physics tests node age.js --demo # two-act demo: software scaffold + physics design
Simulation-only, CPU-scale, deterministic seeds. No GPU required.
Sister papers in the series
Seven manuscripts, one codebase, one guarantee: every number traces to a committed JSON and regenerates from a committed script.
the PhysFormer architecture; the falsified regime theory
the 12-domain verifiable benchmark
the gate as the missing control in agent evaluation
when physics in the loss helps — and when it only enforces consistency
transfer across laws: what carries the knowledge
the cost of consistency on 2D fields