Skip to content

LEADERBOARD.md

Latest commit

 

History

History
37 lines (25 loc) · 3.55 KB

LEADERBOARD.md

File metadata and controls

37 lines (25 loc) · 3.55 KB

TwinBench Leaderboard

TwinBench currently has two public result classes:

  • first-party measured harness artifacts under the earlier TwinBench v0.2 schema
  • modeled baseline comparisons under the current simplified TwinBench v1 scaffold

No public third-party submissions are listed yet.

First-Party Measured Artifacts

These are recorded benchmark runs checked into the repository. They are first-party evaluations and should be read as repository evidence, not neutral external audits.

System Version Benchmark Version Date Evaluated MR IC CCC TC PG Total Score Evidence Caveats / Notes
Nullalis local openended race local openended race 0.2 2026-03-25 - - - - - 75.9 coverage-adjusted verified legacy/results/nullalis-live-2026-03-25-openended.json Real first-party harness run. Uses legacy 10-dimension schema, not v1 five-metric scoring. Measured coverage 83.5%.
Nullalis local live baseline local live 0.2 2026-03-24 - - - - - 68.1 coverage-adjusted verified legacy/results/nullalis-local-2026-03-24.json Real first-party harness run. Earlier stable baseline before later failures.
Nullalis auth-mismatch degraded run local live 0.2 2026-03-25 - - - - - 5.8 coverage-adjusted verified legacy/results/nullalis-live-2026-03-25.json Real first-party degraded run. Chat turns failed with unauthorized errors; not suitable as headline evidence.
Nullalis targeted degraded run local live 0.2 2026-03-24 - - - - - 0.0 coverage-adjusted verified legacy/results/nullalis-targeted-2026-03-24.json Real first-party degraded run. Runtime unavailability dominated the result.

Modeled Baseline Comparisons

These entries are comparison baselines designed to show what TwinBench distinguishes. They are not live measured runs.

Displayed score scale: 0.0-1.0 normalized

System Version Benchmark Version Date Evaluated MR IC CCC TC PG Total Score Evidence Caveats / Notes
ChatGPT-like stateless conversational AI baseline-simulated-v1 TwinBench v1 2026-04-03 0.10 0.42 0.22 0.18 0.08 0.20 results/chatgpt_baseline_v1.json Modeled baseline, not a live product run.
OpenClaw-like agent system baseline-simulated-v1 TwinBench v1 2026-04-03 0.40 0.35 0.33 0.56 0.20 0.37 results/openclaw_baseline_v1.json Modeled baseline, not a live product run.
ZAKI / Nullalis-like persistent system baseline-simulated-v1 TwinBench v1 2026-04-03 0.78 0.74 0.71 0.76 0.80 0.76 results/zaki_baseline_v1.json Modeled persistence-oriented profile. Should not be confused with the measured Nullalis harness runs above.

Interpretation Notes

  • The strongest measured result in the repository is the Nullalis 2026-03-25 open-ended harness run.
  • The v1 baseline rows are easier to compare quickly, but they carry lower evidentiary weight because they are modeled.
  • TwinBench should prefer measured artifacts over modeled baselines whenever both exist.