Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Skill loop experiment

What this is

This repository contains a pre-registered skill-improvement experiment run on 2026-08-11 and 2026-08-12. It transfers the controller pattern from Andrej Karpathy's autoresearch—bounded mutation, a fixed budget, a frozen evaluator, and keep-or-revert selection—from LLM training to a single agent skill file.

Companion article (live at publication): The Cost of Unmeasured Change.

The data are synthetic: all 16 tenancy contracts were generated, and every entity in them is invented.

Headline result

Finalist Held-out score
v0 (starting skill) 0.902778
v1 (loop-final) 1.000000
bo12-04 (best blind rewrite) 0.875000

H1 was supported: v1 beat v0 by +9.72pp, above the pre-registered 1.94pp held-out threshold.

H2 had no directional prediction: v1 beat the best blind rewrite, bo12-04, by +12.5pp.

H3 was not run under SPEC amendment A3.

Repo map

  • 00-control/ — protocol, freeze record, reviews, and holdout ledger.
  • 01-fixtures/ — generator, synthetic corpus, and gold records.
  • 02-evaluator/ — mechanical grader and scorer.
  • 03-skill/ — every skill version from v0 through v5 plus all bo12 attempts.
  • 04-runner/ — runner for the pinned worker invocation, evidence-pack builder, and tests.
  • 05-runs/ — every raw worker output and manifest, preserved byte-exact.
  • 06-log/ — chronological experiment log.
  • 07-analysis/ — locked RESULTS.md, trajectory, diff evolution, and gaming audit.

Verify the freeze

From the repository root, re-hash each frozen file and compare the output with 00-control/FREEZE.md §1:

sha256sum 01-fixtures/generator/generate.py
sha256sum 02-evaluator/grader.py
sha256sum 02-evaluator/score.py
sha256sum 04-runner/run.py
sha256sum 03-skill/versions/v0/SKILL.md
sha256sum 00-control/worker-system-prompt.txt
sha256sum 00-control/worker.json
sha256sum 04-runner/evidence.py

Regenerate the corpus in a temporary directory and byte-compare all committed documents and gold records:

python3 01-fixtures/generator/generate.py --check

The batch manifests in 05-runs/ bind the skill, document, and runner hashes for each batch.

Rerun it

The published results were produced through subscription CLIs with the worker pinned to the alias model string claude-sonnet-5 and Claude Code CLI 2.1.227. That is an alias pin, not a dated model snapshot; see FREEZE.md §5. A rerun is therefore a new experiment, not a replay.

Use the worker invocation in 00-control/worker.json together with 00-control/worker-system-prompt.txt. Follow 04-runner/README.md to execute batches and use 02-evaluator/ to score them. 00-control/SPEC.md contains the full protocol and append-only amendment trail.

Registered caveats

  • The model pin is an alias; no dated snapshot was exposed.
  • Measured headroom is concentrated in two derivation rules: contract_end_date and security_deposit_aed.
  • The decoys contribute zero measured difficulty because each is neutralized by an explicit disclaimer clause.
  • The deposit is a constant ratio corpus-wide, so clause-reading and constant-memorization strategies are observationally equivalent on this corpus; see 07-analysis/evaluator-missed-audit.md.
  • All experiment data are synthetic.
  • H3 was waived under SPEC amendment A3: status reporting contaminated the founder's independence, so the manual baseline was confounded and not run.

Attribution

The controller pattern is Karpathy's: autoresearch supplies the bounded mutable surface, fixed budget, frozen evaluation, and keep-or-discard loop, while Verifiability provides the broader framing for tasks that can be efficiently attempted and mechanically rewarded. This experiment transfers that pattern from training code to an agent skill file.

The evidentiary layer—noise-floor calibration and a keep threshold, held-out single-shot evaluation, independent baselines, pre-registration with amendments, hash-freeze manifests, and adversarial cross-review—is this experiment's addition. It is not attributed to the autoresearch repository.

License

MIT.

About

Pre-registered skill-improvement experiment: the autoresearch controller pattern (bounded mutation, frozen evaluator, keep-or-revert) transferred from LLM training to an agent skill file. Full protocol, runs, and locked results.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages