AI makes the abandoned road searchable.

The opportunity is not that an LLM can write a persuasive anti-quantum essay. The opportunity is that candidate dynamics can now be generated, executed, scored and killed at machine speed.

The computational bottleneck changed

Symbolic-regression systems such as AI Feynman have recovered compact analytic expressions from numerical data. Modern machine-learning research explicitly targets discovery of governing PDEs and reduced dynamical models. Evaluator-driven systems such as AlphaEvolve show the broader pattern: generate executable candidates, score them objectively, preserve successful mutations, repeat.

That architecture is unusually well matched to Classical Reconstruction because the project has a huge candidate space and brutally specific empirical constraints.

The loop is now public and runnable

Open the executable reconstruction protocol. It implements encode → generate → simulate → score → falsify → mutate in the browser and ships a downloadable Python loop. The current numerical benchmark is deliberately synthetic; the next adapters will replace it with real field equations and experimental datasets.

Why adversarial agents matter

What AI cannot be allowed to do

No candidate earns credit by curve fitting one graph. No LLM-generated equation is accepted without execution. No mutation may silently change locality, measurement independence, detector physics or state space. No failed candidate disappears from history. Description length, conservation laws, cross-domain transfer and Q1–Q5 status belong in the fitness function.

Research basis

Udrescu & Tegmark (2020), AI Feynman
Brunton & Kutz (2024), machine learning for PDE discovery
DeepMind (2025), AlphaEvolve evaluator-driven evolution

If AI merely agrees with the thesis, it adds nothing. If it can generate a model and another agent can kill it, the project has become physics.