HERMES BENCHMARKS // FLEET TELEMETRY

← NEWS
BOOTING awaiting uplink

30-DAY TOKEN FLOW

INOUTKANBAN DONE

MODEL GRADES

    ALL FLEET DETAILSusage leaderboard · kanban · report cards · categories · skills · crew · cron

    MODEL LEADERBOARD

      KANBAN THROUGHPUT

        TOP ASSIGNEES

          REPORT CARDS

          PROFILE REPORT CARDS

          CRON RELIABILITY

            TASK CATEGORIES · TIME TO COMPLETE

              Work time = minutes agents actually ran on a card (sum of its runs), median per category and per model. Categories come from a fixed keyword rule on each card; synthetic and canary cards are excluded. Faster is not better on its own: compare models within a category, and treat PROVISIONAL (fewer than 5 cards) as a lead.

              TOP 10 SKILLS USED

                Counted from skill loads (skill_view calls) across every profile; linked-file reads and failed loads excluded, each call counted once.

                HOW GRADING WORKS

                rubric v1 · refreshed every 5 min
                40%KANBAN OUTCOMESfirst-pass, reopens and blocked cards on completed work
                20%CRON RELIABILITYsuccess rate of scheduled job runs
                20%BLIND REVIEWan LLM judge from the default profile scores output against the brief — the doer's model is hidden
                20%PROVING GROUNDfixed bench tasks with deterministic pass/fail checks

                Composite = weighted mean over the tiers where a model has enough samples, weights renormalised; small-sample tiers never count. A letter needs at least two such tiers, otherwise the row shows — and PROVISIONAL. Cron grades count only jobs that call a model. Grades: A ≥ 90 · B ≥ 80 · C ≥ 70 · D ≥ 60 · F below.

                CREW ROSTER

                CRON HEALTH

                —

                LAST 24 HOURS