Skip to content

About

Scripted analog of arXiv 2610.39508 (SkillCompiler): a verified compiled skill bank matches solve-from-scratch success (0.937) while cutting synthesis to 0.025 calls per task (99.2% bank hit rate). Remove the verification gate and success collapses to 0.515. Deterministic, CPU-only.

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

2 Commits

Folders and files

Repository files navigation

skillcompiler (project #117)

Scripted analog of arXiv 2610.39508 "Compile Away Uncertainty": textual agent skills re-interpreted every execution vs executability-aware compilation into validated programs plus a program asset bank with reuse.

Run

  • python3 -m pytest tests/ -q — 10 tests
  • python3 run_pilot.py — pilot (seed 117001, n=300)
  • python3 run_headtohead.py — head-to-head (seed 117002, n=1200)

Headline (head-to-head, n=1200, paired bootstrap 2000)

arm success syntheses/task
textual 0.797 0.000
fresh compile (no bank) 0.937 3.000
compiled + validated bank 0.937 0.025
bank, no validation gate 0.515 0.025
bank, lax (convention ignored) 0.459 0.004

Textual flip rate across 3 repetitions: 0.328, CI [0.283, 0.375]. Bank − textual: +0.140, CI [0.119, 0.163]. All 8 PREREG bars pass.

Scope honesty

Scripted analog only. The paper's figures (up to 32.33% improvement, five benchmarks, six model families) are context only, never claimed as reproduced. The textual slip rate is a stated disease-model parameter; the reuse-cost, validation-gate, and compatibility effects emerge from the candidate/validation/key mechanics.

About

Scripted analog of arXiv 2610.39508 (SkillCompiler): a verified compiled skill bank matches solve-from-scratch success (0.937) while cutting synthesis to 0.025 calls per task (99.2% bank hit rate). Remove the verification gate and success collapses to 0.515. Deterministic, CPU-only.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors