reversible-circuit-8b-tool — tool-driven reversible-circuit synthesis (Qwen3-8B)

• Base: Qwen/Qwen3-8B (Apache-2.0) · License: Apache-2.0 · Method: LoRA SFT (Unsloth/TRL on Modal)
• Full writeup (read this): docs/PROCESS_LOG.md · docs/WRITEUP.md

What it does

Honest evaluation (held-out, 40 tasks/band, best-of-5)

What we learned (and what did NOT work — stated plainly)

• The tool removes the real bottleneck. Without it, a 1.5B and a 7B model one-shot-synthesize identically (~4.8%) — the limiter is symbolic execution, not capacity. With the tool, scale then matters (a trained 1.5B caps at n=4; this 8B reaches n=5).
• A self-harvest "flywheel" (expert iteration on the model's own verified solutions) did NOT improve held-out capability — a clean negative result. base ≈ iter-1 ≈ iter-2 (~58% best-of-5). An earlier apparent "n=6 cracked 0→7.5%" was a best-of-2 sampling artifact (this base already solves n=6 at ~5% with enough attempts). SFT on a model's own correct outputs re-teaches what it already does; it cannot push the frontier.
• Measurement discipline was the real lesson: under-sampled evals manufactured two phantom "wins" that an adequately-sampled, fixed held-out set erased.

Intended use & limitations

Reproduce