Benchmark run gpt-6-luna-all-r10-c5-prescriptive-20260927
This page contains exactly the publishable benchmark observations recorded in run gpt-6-luna-all-r10-c5-prescriptive-20260927; it does not select or aggregate them. Return to the run index or view aggregate metrics.
Interactive browser
Rows are pre-aggregated by task, target, prompt variant, MCP-instruction mode, context, arm, and restriction. Numeric values show median (p50), IQR, p90, maximum, and sample count. Boolean checks show pass counts and success rates. Filters apply to these summary groups; per-run observations remain available on their run pages.
Loading benchmark results…
Baseline vs semedit
Comparison evaluates baseline and semedit values within each restriction policy, target, and context group. Negative changes mark improvements for these lower-is-better metrics; positive changes mark degradations.
Detailed observations
Empirical telemetry measuring wall-clock latency, token usage, tool turns, and compiler correctness when AI coding agents perform code modifications.
Every benchmark pairs Vanilla LLM (standard file editing tools) against Semedit MCP (deterministic AST compiler operations).
Benchmark Methodology & Transparency
- Reproducibility & Provenance: Self-contained txtar archives define test scenarios. Links below resolve to commit-anchored GitHub source files for published commits, or cryptographic SHA-256 fingerprints for uncommitted local fixtures.
- Model Cost: Tables report model-specific credit consumption where a target model has a declared schedule: (uncached input × input credits + cached input × cached credits + (reasoning + visible output) × output credits) / 1,000,000. Thinking tokens use the output rate. ADR-0043 defines the per-million-token credit rates.
- Multi-Level Correctness Oracle: Four validation levels evaluate each trial:
- Level 1 (Mutation Policy): Restricts file modifications strictly to authorized paths.
- Level 2 (AST Invariants): Compiler AST verification of required symbols, imports, and relative declaration ordering.
- Level 3 (Clean Build): Strict compilation validation with zero compiler or typecheck errors.
- Level 4 (Test Suite): Unit and integration test suite execution.
- Isolated Sandboxing: Each trial executes in an ephemeral directory with pristine git and module state.
Note
No benchmark evaluations exist yet in data/benchmarks/results/. Run the benchmark harness using go run ./tools/benchmark-harness --target ... to record empirical results.