Add three loop extension repos (autoexperiment, autoadversary, metaresearch)

[shyamsivakumar]

Summary

Three novel loop extensions for Karpathy's autoresearch pattern:

autoexperiment

Hypothesis-driven loop requiring the agent to maintain a causal model (theory.md) and track change-type effectiveness (credit.json). Produces transferable knowledge artifacts.

autoadversary

Two-agent adversarial loop with an adversary that systematically finds failure modes, preventing Goodhart's Law. Produces an auto-growing adversarial test suite.

metaresearch

Nested outer loop that optimizes program.md itself - discovering empirically validated research strategies automatically.

All three extend the standard autoresearch loop while using the same nanoGPT base for direct comparability.

[ZhengyaoJiang]

Hey, those ideas all sound cool. I’m wondering if you have actual run trajectories. Are they better than the vanilla autoresearch loop (for example, in terms of sample efficiency)?

[shyamsivakumar]

---

Hey Zhengyao, thanks for asking! We ran a controlled comparison — 4 variants × 10 experiments each, same nanoGPT/TinyShakespeare base (100 training iters/experiment), MiniMax M2.7 as the research agent making per-step decisions.

What we see so far:

• metaresearch found the steepest improvement curve — the "start with regularization" meta-strategy led it to systematically ramp attn_dropout (0.1→0.15→0.2), finding 3 consecutive wins by exp 5
• autoexperiment used credit scores to double down on depth (n_layer 8→10→12→14), correctly identifying architecture as the productive axis
• autoadversary was slower to start but kept improving through exp 10 via cautious regularization (weight_decay + targeted mlp_dropout)
• vanilla tried things semi-randomly and plateaued early