{% extends "base.html" %} {% block title %}Trials{% endblock %} {% block body %}

Model trials

The same task, several models, one comparison. Scores accumulate here so the tier-to-model map is a decision, not a guess.

Filter by contender
{% if filter_harness %}Clear{% endif %}
Leaderboard
{% if rows %} {% for r in rows %}{% endfor %}
contendertrialswinswin rateavg scoreavg cost$ per pointavg tokens in / outfailedenv failed
{{ r.label }}{{ r.trials }}{{ r.wins }}{% if r.win_rate is not none %}{{ (r.win_rate * 100)|round|int }}%{% endif %}{% if r.avg_score is not none %}{{ r.avg_score }}{% endif %}{% if r.avg_cost is not none %}${{ '%.2f' % r.avg_cost }}{% endif %}{% if r.cost_per_point is not none %}${{ '%.3f' % r.cost_per_point }}{% endif %}{{ '{:,}'.format(r.avg_input_tokens) }} / {{ '{:,}'.format(r.avg_output_tokens) }}{{ r.failed }}{{ r.env_failed }}

env failed: the contender's own harness or environment stopped it (quota, login, a sandbox denial, a missing setup) — excluded from the comparison, not scored as a loss.

$ per point is cost divided by the comparison score: the cheapest way to a good PR is the lowest number, not the lowest cost.

{% else %}
No trials yet. On a ready task's page choose Model trial… and name two or more contenders, e.g. claude:sonnet, claude:opus.
{% endif %}
{% for t, md in trials %}
{{ t.task }} {{ t.at[:16] }}{{ md|md|safe }}
{% endfor %} {% endblock %}