Phipps Eval — PE Energy AI Benchmark

Cloud benchmark

Hosted model quality and economics

The 33-task PE benchmark compares cloud/API models on one artifact-backed board. Local PE runs use the older harness and stay available as a separate legacy view—not mixed with production qualification.

33 tasks · 3× Finals highlighted

Overall /100 50% Phipps behavior + 35% broad capability + 15% BFCL tool policy. Missing components are re-normalized and labeled with coverage; hard failures remain disqualifying outside the weighted score.

Speed vs Overall /100

Production qualification

Fully tested Daily Driver comparisons

The primary board shows only current-manifest 3× Finals and keeps one best-evidenced Overall score per model family. Screens, reasoning alternatives, superseded router builds, and every valid historical lane remain available below.

Same v4 harness · Finals only

Validated families

—

Final runs

—

Run history

—

Overall /100 50% Daily Driver behavior + 35% broad capability + 15% BFCL tool policy. Coverage stays visible until every external receipt is complete.

No signed-off model runs yet

The Daily Driver board will populate only from valid database receipts. Invalid or incomplete results never enter the ranking.

36-task core/smoke manifest — database pending

Methodology

Daily Driver is a separate 36-task suite covering interaction quality, tool judgment, instruction following, memory/context, knowledge calibration, and coding/problem solving. A complete 1× lane is a Screen; only a complete 3× lane is a Final. Screens never rank on the primary board.

Status

Waiting for the signed-off suite export.

Manifest: —