{% extends "base.html" %}
{% block title %}Agreement — FreeWeight{% endblock %}
{% block content %}
<h2>{{ goal.pack.name }}</h2>
{% include "partials/_wizard_steps.html" %}
<h3>6. {{ step.heading }}</h3>

{% if error %}
<div class="error-state" role="alert"><p>{{ error }}</p></div>
{% endif %}

{% if outcome %}
{% set verdict = outcome.verdict %}
<div class="{{ 'notice' if verdict.state.value == 'calibrated' else 'warning-state' }}" role="status">
  <h3>
    {% if verdict.state.value == 'calibrated' %}
    The jury agrees with you well enough to be worth trusting
    {% elif verdict.state.value == 'uncalibrated' %}
    Not measurable yet — and that is a useful answer
    {% else %}
    {{ verdict.state.value | replace('_', ' ') }}
    {% endif %}
  </h3>
  <p class="band">
    {% if verdict.weighted_kappa_w is none %}
    There is no agreement figure yet.
    {% else %}
    Across your judged criteria, the jury and you agree at
    <strong>{{ '%.2f' % verdict.weighted_kappa_w }}</strong>
    over <strong>{{ verdict.n_holdout }}</strong> held-out samples.
    {% endif %}
  </p>
  <p>
    {% if verdict.state.value == 'calibrated' %}
    Results from this goal emit capability evidence, discounted by a judge validity factor of
    {{ '%.2f' % verdict.judge_validity_factor }}.
    {% elif verdict.state.value == 'uncalibrated' %}
    Your goal still runs, and every sample stays inspectable. What it will not do is emit
    capability evidence — not discounted evidence, none. The criteria below are where the jury
    and you diverged most, and that is where the rubric needs work.
    {% else %}
    Grade more samples: the minimum is {{ verdict.min_samples }} and you have
    {{ verdict.graded_samples }}. Too few grades is a different thing from poor agreement, and
    this is the first one.
    {% endif %}
  </p>
</div>

<h3>What the number means</h3>
<div class="table-scroll">
  <table>
    <caption>The band, and the consequence, in words. The coefficient is beside it, never instead
      of it.</caption>
    <thead>
      <tr><th scope="col">Band</th><th scope="col">&kappa;<sub>w</sub></th><th scope="col">What it means for you</th></tr>
    </thead>
    <tbody>
      {% for label, range_text, meaning in bands %}
      <tr>
        <td><strong>{{ label }}</strong></td>
        <td class="mono">{{ range_text }}</td>
        <td>{{ meaning }}</td>
      </tr>
      {% endfor %}
    </tbody>
  </table>
</div>

<h3>Per criterion</h3>
<div class="table-scroll">
  <table data-table="agreement">
    <caption>
      Every coefficient carries its <code>n_holdout</code>. A &kappa;<sub>w</sub> without its n is
      a number pretending to be a fact.
    </caption>
    <thead>
      <tr>
        <th scope="col">Criterion</th>
        <th scope="col" class="numeric">Weight</th>
        <th scope="col" class="numeric">&kappa;<sub>w</sub></th>
        <th scope="col" class="numeric">&rho;</th>
        <th scope="col" class="numeric">MAE</th>
        <th scope="col" class="numeric">Bias</th>
        <th scope="col" class="numeric">n<sub>holdout</sub></th>
        <th scope="col">Band</th>
        <th scope="col">The lint's read</th>
      </tr>
    </thead>
    <tbody>
      {% for agreement in outcome.criteria %}
      <tr>
        <td>{{ agreement.criterion_key }}</td>
        <td class="numeric">{{ '%.0f' % (agreement.weight * 100) }}%</td>
        <td class="numeric">{{ '%.3f' % agreement.result.kappa_w }}</td>
        <td class="numeric">{{ '%.3f' % agreement.result.rho }}</td>
        <td class="numeric">{{ '%.2f' % agreement.result.mae }}</td>
        <td class="numeric">
          {{ '%+.2f' % agreement.result.bias }}
          <br><span class="unit">{{ 'generous' if agreement.result.bias > 0 else 'harsh' }}</span>
        </td>
        <td class="numeric">{{ agreement.result.n_holdout }}</td>
        <td>{{ agreement.band | replace('_', ' ') }}</td>
        <td class="muted">{{ agreement.lint or '—' }}</td>
      </tr>
      {% endfor %}
    </tbody>
  </table>
</div>

{% for agreement in outcome.criteria %}
{% if agreement.disagreements %}
<h3>Where {{ agreement.criterion_key }} diverged most</h3>
<p class="muted">
  Your grade, your note, and the jury's. This is the part that teaches: the number is attached to
  a sample you remember rather than to an abstraction.
</p>
{% for disagreement in agreement.disagreements %}
<div class="table-scroll">
  <table>
    <tbody>
      <tr>
        <th scope="row">You graded</th>
        <td class="numeric">{{ disagreement.author_grade }}</td>
        <td>{{ disagreement.author_note or '—' }}</td>
      </tr>
      <tr>
        <th scope="row">The jury graded</th>
        <td class="numeric">{{ disagreement.jury_grade }}</td>
        <td>{{ disagreement.jury_rationale or '—' }}</td>
      </tr>
    </tbody>
  </table>
</div>
<details>
  <summary>The sample</summary>
  <pre>{{ disagreement.excerpt }}</pre>
</details>
{% endfor %}
{% endif %}
{% endfor %}

{% if outcome.warnings %}
<div class="warning-state">
  <h3>Before you read those numbers</h3>
  <ul>
    {% for warning in outcome.warnings %}<li>{{ warning }}</li>{% endfor %}
  </ul>
</div>
{% endif %}

<p class="muted">
  FreeWeight proposes no rewritten criterion text. It names the problem and shows the evidence.
  The rubric is yours, and a model that edited it until it could measure it would be measuring its
  own edit.
</p>
{% else %}
<div class="empty-state">
  <h3>No calibration yet</h3>
  <p>Grade this goal's samples, then run the calibration:</p>
  <pre>freeweight goals calibrate {{ goal.slug }}</pre>
  <p><a href="/goals/{{ goal.slug }}/grade">Go to grading</a></p>
</div>
{% endif %}
{% endblock %}