{% extends "base.html" %}
{% block title %}{{ data.review_name }}{% endblock %}
{% block content %}
  <section>
    <p class="muted"><a href="{{ url_for('review_launcher') }}">← Back to review</a></p>
    <h1>{{ data.review_name }}</h1>
    <p class="muted" style="margin: 0.2rem 0 0.8rem 0; font-size: 1rem;">{{ data.paper.title }}</p>
    <p class="muted" style="border-left: 3px solid var(--warn); padding: 0.5rem 0.75rem; background: rgba(0,0,0,0.02);">
      <strong>Reminder:</strong> this output is a <strong>draft-polishing aid</strong> — it is not a peer-review
      generator. Most venues have strict policies against using LLMs in assigned reviews. Please use it at your
      own discretion, and indicate when you have used it. Every comment is a
      <strong>suggestion to evaluate</strong>, not a finding to accept. AI reviewers hallucinate, miss context,
      and over-confidently flag non-issues. Expect to reject roughly half of what you see.
    </p>
    <div class="download-row">
      <a class="button" href="{{ url_for('review_download', job_id=job_id, kind='md') }}">Download report (.md)</a>
      <a class="button" href="{{ url_for('review_download', job_id=job_id, kind='data') }}">Download per-reviewer data (.md)</a>
      {% if data.clarity_review and data.clarity_review.comments %}
        <a class="button" href="{{ url_for('review_download', job_id=job_id, kind='clarity') }}">Download writing clarity review (.md)</a>
      {% endif %}
      <form method="post" action="{{ url_for('review_delete', job_id=job_id) }}"
            style="display:inline;"
            onsubmit="return confirm('Delete this review? Files will be removed from disk and cannot be recovered.');">
        <button type="submit" class="button button-danger" title="Delete this review and its files">Delete review</button>
      </form>
    </div>
    {% if data.llm_provider and data.llm_model %}
      <p class="muted" style="margin-top: 0.8rem; font-size: 0.9rem;">
        <strong>LLM:</strong> <code>{{ data.llm_provider }}</code> / <code>{{ data.llm_model }}</code>
        &nbsp;·&nbsp; <strong>Base URL:</strong> <code>{{ data.llm_base_url or '(default)' }}</code>
        {% if data.launched_at %}
          &nbsp;·&nbsp; <strong>Launched:</strong> {{ data.launched_at }}
        {% endif %}
        {% if data.ended_at %}
          &nbsp;·&nbsp; <strong>Ended:</strong> {{ data.ended_at }}
        {% endif %}
      </p>
    {% endif %}
  </section>

  {% if run_files %}
    {# One macro for the three run-files tables so column changes land in one place. #}
    {% macro run_files_table(files) -%}
      <table class="rows run-files">
        <colgroup>
          <col class="col-file">
          <col class="col-size">
          <col class="col-description">
        </colgroup>
        <thead><tr><th>File</th><th>Size</th><th>Description</th></tr></thead>
        <tbody>
          {% for f in files %}
            <tr>
              <td><code>{{ f.abs_path }}</code></td>
              <td class="muted" style="white-space: nowrap;">{{ f.size }}</td>
              <td>{{ f.description }}</td>
            </tr>
          {% endfor %}
        </tbody>
      </table>
    {%- endmacro %}

    <section class="card">
      <h2>Source files on disk</h2>
      <p class="muted" style="font-size: 0.9rem;">
        Everything produced by this review lives in a single directory. Paths
        below are absolute so you can copy-paste straight into a shell.
      </p>
      <p><strong>Run directory:</strong> <code>{{ run_files.run_dir }}</code></p>

      {% if run_files.inputs %}
        <h3 style="margin-top: 1rem;">Inputs</h3>
        {{ run_files_table(run_files.inputs) }}
      {% endif %}

      {% if run_files.outputs %}
        <h3 style="margin-top: 1rem;">Outputs</h3>
        {{ run_files_table(run_files.outputs) }}
      {% endif %}

      {% if run_files.internal %}
        <details style="margin-top: 1rem;">
          <summary class="muted">Internal artifacts ({{ run_files.internal|length }})</summary>
          <div style="margin-top: 0.5rem;">
            {{ run_files_table(run_files.internal) }}
          </div>
        </details>
      {% endif %}
    </section>
  {% endif %}

  <section class="card">
    <h2>Selected reviewers</h2>
    <p class="muted">{{ data.selected|length }} reviewers selected by topic similarity, then diversified across personas.</p>
    <table class="rows">
      <thead><tr><th>ID</th><th>Domain</th><th>Persona</th><th>Selection relevance</th></tr></thead>
      <tbody>
        {% for r in data.selected | sort(attribute='id') %}
          <tr>
            <td><a href="{{ url_for('reviewer_detail', rid=r.id) }}"><code>{{ r.id }}</code></a></td>
            <td>{{ r.domain }}</td>
            <td>{{ r.persona }}</td>
            <td>{{ "%.3f"|format(r.score) }}</td>
          </tr>
        {% endfor %}
      </tbody>
    </table>
    {% if data.n_reviewers_total is defined %}
      <p class="muted" style="margin-top: 0.8rem; font-size: 0.9rem;">
        <strong>Format-fix retries:</strong>
        {{ data.n_format_repairs or 0 }} of {{ data.n_reviewers_total }}
        reviewer(s) (incl. the clarity reviewer)
        needed a markdown-repair pass to produce usable output.
        {% if (data.n_format_repairs or 0) > 0 %}
          A high count suggests the model is deviating from the expected
          comment format — consider a different provider/model or
          tightening the persona prompt.
        {% endif %}
      </p>
    {% endif %}
  </section>

  {% if data.clarity_review and data.clarity_review.comments %}
    <section class="card">
      <h2>Writing clarity review</h2>
      <p class="muted">
        Always-on reviewer focused on writing quality (flow, terminology,
        grammar, figures, structure). Not part of the ranked issues and
        not compared against human reviewers during Validation —
        <code>{{ data.clarity_review._reviewer_id }}</code> /
        {{ data.clarity_review._persona }}.
      </p>
      {% for c in data.clarity_review.comments %}
        {% set sev = c.severity or 'minor' %}
        <article class="issue issue-{{ sev }}">
          <header>
            <span class="sev sev-{{ sev }}">{{ sev|upper }}</span>
            <span class="muted" style="font-size: 0.9rem;">
              category: <code>{{ c.category or 'clarity' }}</code>,
              section: <code>{{ c.section_reference or 'general' }}</code>
            </span>
          </header>
          {% if c.summary %}<p><strong>{{ c.summary }}</strong></p>{% endif %}
          {% if c.description %}<p>{{ c.description }}</p>{% endif %}
        </article>
      {% endfor %}
    </section>
  {% endif %}

  <section>
    <h2>Ranked issues</h2>
    <p class="muted">
      Ordered by <strong>commonality × importance</strong>: issues multiple
      reviewers raise, weighted by how severe they are, rise to the top.
    </p>
    <details style="margin: 0.5rem 0 1rem 0;">
      <summary><strong>How the score is computed</strong></summary>
      <div style="margin-top: 0.6rem; font-size: 0.95rem;">
        <p>
          Each issue cluster groups comments that look like the same concern
          (cosine similarity on summary + keywords). Clusters are then scored:
        </p>
        <pre style="padding: 0.6rem 0.9rem;"><code>score = num_distinct_reviewers × (0.5·avg_severity + 0.5·max_severity)</code></pre>
        <p>
          where severity weights are:
        </p>
        <ul>
          <li><span class="sev sev-major">MAJOR</span> = 3 (would block acceptance)</li>
          <li><span class="sev sev-moderate">MODERATE</span> = 2 (significant revision)</li>
          <li><span class="sev sev-minor">MINOR</span> = 1 (improvement but not blocking)</li>
        </ul>
        <p>
          Higher-severity comments AND broader reviewer agreement both push
          a cluster up. Intuition: one reviewer's minor issue scores 1·(0.5·1 + 0.5·1) = 1;
          three reviewers agreeing on a moderate issue scores 3·(0.5·2 + 0.5·2) = 6;
          five reviewers on a mix of major and moderate scores
          5·(0.5·2.4 + 0.5·3) = 13.5.
        </p>
        <p style="margin-top: 0.6rem;">
          <strong>The severity badge next to each cluster shows the worst
          individual comment in the cluster</strong>, not an average. A
          cluster can legitimately have score 6 with a <span class="sev sev-minor">MINOR</span>
          label — that means six reviewers all flagged the same thing as
          minor, and none escalated it to moderate or major. The count of
          agreeing reviewers still pushes its score up. The small coloured
          chips beside the label ("<span class="sev-chip sev-minor">6 minor</span>"
          etc.) show the severity mix so you can tell a broadly-agreed nit
          from a genuinely severe issue at a glance.
        </p>
      </div>
    </details>
    <details style="margin: 0.5rem 0 1rem 0;">
      <summary><strong>What the <code>category</code> and <code>section</code> fields show</strong></summary>
      <div style="margin-top: 0.6rem; font-size: 0.95rem;">
        <p>
          Each cluster row shows two tags pulled from the representative
          comment (the highest-severity member):
        </p>
        <ul>
          <li>
            <code>category</code> — a classification of <em>what kind of
            concern</em> this is, snapped to a fixed vocabulary at parse
            time. Examples: <code>novelty</code>, <code>evaluation</code>,
            <code>methodology</code>, <code>reproducibility</code>,
            <code>presentation</code>, <code>deployment</code>,
            <code>related_work</code>, <code>correctness</code>. This is
            what the validator uses to route each miss to an expected
            persona (Methodology Critic catches <code>methodology</code>,
            Novelty Hunter catches <code>novelty</code>, etc.). Shown as
            <code>general</code> if the LLM didn't pick a specific one.
          </li>
          <li>
            <code>section</code> — a normalized anchor into <em>where in
            the paper</em> the comment applies, extracted from the
            reviewer's text by pattern matching: <code>Section 3.2</code>,
            <code>Table 2</code>, <code>Figure 4</code>,
            <code>Algorithm 1</code>, <code>Eq. 7</code>, or named parts
            like <code>Abstract</code>, <code>Introduction</code>,
            <code>Related Work</code>, <code>Conclusion</code>,
            <code>References</code>. Shown as <code>general</code> when
            the comment doesn't point at any specific part of the paper.
          </li>
        </ul>
        <p>
          Clusters group comments by semantic similarity, so the
          representative's category and section are shared by most of
          the cluster — but individual members may phrase section
          references slightly differently. Expand the "other reviewers
          raised the same issue" details below each cluster to see the
          full text each reviewer wrote.
        </p>
      </div>
    </details>
    {% for c in data.ranked_clusters %}
      {% set rep = c.representative %}
      {% set sev = rep.severity or 'minor' %}
      <article class="issue issue-{{ sev }}">
        <header>
          <span class="rank">#{{ c.rank }}</span>
          <span class="sev sev-{{ sev }}">{{ sev|upper }}</span>
          {# Severity mix — makes "score 6, MINOR" readable as "6 reviewers all flagged this as minor". Only shown for multi-member clusters. #}
          {% if c.members|length > 1 %}
            {% set n_major = c.members|selectattr('severity', 'equalto', 'major')|list|length %}
            {% set n_moderate = c.members|selectattr('severity', 'equalto', 'moderate')|list|length %}
            {% set n_minor = c.members|selectattr('severity', 'equalto', 'minor')|list|length %}
            <span class="sev-mix" title="How the {{ c.members|length }} comments in this cluster break down by severity. The label above reflects the worst individual comment.">
              {% if n_major %}<span class="sev-chip sev-major">{{ n_major }} major</span>{% endif %}
              {% if n_moderate %}<span class="sev-chip sev-moderate">{{ n_moderate }} moderate</span>{% endif %}
              {% if n_minor %}<span class="sev-chip sev-minor">{{ n_minor }} minor</span>{% endif %}
            </span>
          {% endif %}
        </header>
        <p class="issue-meta">
          <span>Score <strong>{{ c.score }}</strong></span>
          <span>{{ c.size }} comment{{ 's' if c.size != 1 }}</span>
          <span>{{ c.num_distinct_reviewers }} reviewer{{ 's' if c.num_distinct_reviewers != 1 }}</span>
          <span>category: <code>{{ rep.category or 'general' }}</code></span>
          <span>section: <code>{{ rep.section_reference or 'general' }}</code></span>
        </p>
        <p><strong>Description.</strong> {{ rep.description }}</p>
        {% if c.members|length > 1 %}
          <details>
            <summary>{{ c.members|length - 1 }} other reviewer{{ 's' if c.members|length - 1 != 1 }} raised the same issue</summary>
            <ul class="alt-phrasings">
              {% for m in c.members %}
                {% if m is not sameas rep %}
                  <li>
                    <div class="alt-phrasing-head">
                      <a href="{{ url_for('reviewer_detail', rid=m._reviewer_id) }}"
                         title="View this reviewer's persona / system prompt"><code>{{ m._reviewer_id }}</code></a>
                      <span>({{ m._persona }})</span>
                      <span class="sev sev-{{ m.severity or 'minor' }}">{{ (m.severity or 'minor')|upper }}</span>
                      <span class="muted" style="font-size: 0.9rem;">
                        category: <code>{{ m.category or 'general' }}</code>,
                        section: <code>{{ m.section_reference or 'general' }}</code>
                      </span>
                    </div>
                    {% if m.summary %}
                      <p class="alt-phrasing-summary"><strong>{{ m.summary }}</strong></p>
                    {% endif %}
                    {% if m.description %}
                      <p class="alt-phrasing-description">{{ m.description }}</p>
                    {% endif %}
                  </li>
                {% endif %}
              {% endfor %}
            </ul>
          </details>
        {% endif %}
      </article>
    {% else %}
      <p>No issues were raised — surprising; the reviewers produced no comments for this paper.</p>
    {% endfor %}
  </section>
{% endblock %}