Three tools for one question: can you trust this system tomorrow?
All three grade with fixed predicates — no LLM-as-judge — so a score that moves means the system moved, not the test. Bring your own key; it never leaves your tab.
Same discipline pointed at different failure modes.
A frozen 35-task suite runs against 16 live models every day. Every run is kept, and a drop against the previous run is flagged automatically — including the false alarms, which get published too.
Point your own model at adversarial batteries — injection, tool-abuse, hallucination-bait, refusal calibration, retrieval grounding, agent trajectories — and get a severity-weighted vulnerability report with a fail card for every miss.
A harness config is data: roles, models, scaffolding, topology. Draw the shape or paste the JSON, declare the axes to vary, and it runs the whole matrix — against the built-in coding suite or your own tasks. Then it does the part a leaderboard skips: a per-task paired comparison that says whether one shape actually beat another, or whether the suite is too small to tell.
A verdict is a function of the output, so it reproduces exactly. Rerun it and argue with it.
server.pyThe drift board is the one that hasn't been folded in yet.
It was written before gradecore was extracted and still grades with its own copy
of the predicates. That is a duplication to pay down, not a design — and it is named here
because a page claiming "one engine" while one of its three tools quietly runs a second copy
would be exactly the kind of unchecked claim these tools exist to catch.
python check executes locally and labels itself as such in every result.
eval_run.json, so a crash-test result and a
drift run land in the same history and can be compared over time.