{% set ar = ablation_report %}
<section class="ablation">
  <h2>Component ablation
    <span class="muted small">bundle <span class="mono">{{ ar.bundle_hash[:12] }}</span> on {{ ar.base_variant }} · every component tested against its own absence</span>
  </h2>
  {% for note in ar.notes %}<p class="muted small">{{ note }}</p>{% endfor %}
  <h3 class="small muted">Whole bundle: full vs minimal (no components)</h3>
  {% set p = ar.full_vs_minimal %}
  {% include "partials/paired.html" %}
  <table class="data compact">
    <thead><tr><th>component</th><th>verdict</th><th class="num">Δ pass rate</th><th class="num">interval</th><th class="num">Δ {{ ar.full_vs_minimal.cost_kind }}</th><th class="num">Δ llm_calls</th><th class="num">wins / losses / ties</th><th class="num">sign test p</th></tr></thead>
    <tbody>
    {% for e in ar.components %}
      {% set c = e.comparison %}
      <tr>
        <td class="mono">{{ e.component }}</td>
        <td><span class="badge {{ 'badge-pass' if e.verdict == 'helps' else ('badge-fail' if e.verdict == 'hurts' else 'badge-neutral') }}">{{ e.verdict }}</span></td>
        <td class="num">{% if c.pass_rate_diff %}{{ (c.pass_rate_diff.estimate * 100)|delta(1) }} pts{% else %}—{% endif %}</td>
        <td class="num">{% if c.pass_rate_diff %}[{{ (c.pass_rate_diff.low * 100)|delta(1) }}, {{ (c.pass_rate_diff.high * 100)|delta(1) }}]{% else %}—{% endif %}</td>
        <td class="num">{% if c.cost_diff %}{{ c.cost_diff.estimate|delta(4) }}{% else %}—{% endif %}</td>
        <td class="num">{% if c.llm_calls_diff %}{{ c.llm_calls_diff.estimate|delta(2) }}{% else %}—{% endif %}</td>
        <td class="num">{{ c.wins }} / {{ c.losses }} / {{ c.ties }}</td>
        <td class="num">{% if c.sign_test_p is not none %}{{ '%.3f'|format(c.sign_test_p) }}{% else %}—{% endif %}</td>
      </tr>
    {% endfor %}
    </tbody>
  </table>
  <p class="muted small">Effect = full bundle minus the bundle without that component, per task. A component "helps" only when the interval excludes zero over enough tasks; "no evidence" is a reason to prefer the simpler harness.</p>
</section>