Skip to content

Latest commit

 

History

History

README.md

LiveCodeBench Hard: RA2R Logic API on Competitive Programming

Claude Opus 4.6 with maximum-effort extended thinking solves 85.7% of 28 hard AtCoder problems. With one Ejentum Logic API call per task, it solves 100%.

Results

Condition Passed Rate
Baseline (Opus 4.6 max effort) 24/28 85.7%
+ Logic API injection 28/28 100.0%
Delta +4 +14.3pp

Zero regressions. Every task that passed baseline also passed augmented.

Blind Evaluation

A blind evaluator scored both solutions per task without knowing which used the injection:

  • 3.5x magnitude asymmetry. Average injection win: +5.7 points. Average baseline win: -1.6 points.
  • Never loses on correctness (2-0) or robustness (4-0).
  • Independent bug discovery. The evaluator traced a fatal sentinel-collision bug in the baseline, scored it 2/10, without knowing which solution used the injection.

Repository Structure

README.md           This file
REPORT.md           Full benchmark report (296 lines, 13 sections)
skill.md            The Logic API skill file used in the benchmark
run_lcb.py          Benchmark runner (reproducible)
results/
  baseline.json     28 baseline results (metadata, no code)
  augmented.json    28 augmented results (metadata, no code)
  blind_eval.json   27 blind evaluation results with full commentary
  three_way_evals.json   2 three-way blind evaluations

Reproduction

Requires an Ejentum API key (ejentum.com) and Claude CLI.

# Install dependencies
pip install httpx

# Set your API key
export EJENTUM_API_KEY="your-key-here"

# Run baseline
python run_lcb.py --condition baseline --run-id my_run

# Run augmented
python run_lcb.py --condition augmented --run-id my_run

Task data is not included (LiveCodeBench IP). Download from the LiveCodeBench HuggingFace dataset and place batch files in the same directory.

Links