{% extends "base.html" %}
{% block title %}Voxint — synthdetect report {{ run.id.hex[:8] }}{% endblock %}
{% block body %}
<div class="container">
<h1>Synthetic speech detection report</h1>
<p class="muted"><a href="/runs/{{ run.id }}">← Back to run {{ run.id.hex[:8] }}</a></p>

{% if synthdetect_job is none %}
  <p>No synthdetect job found for this run.</p>
{% elif synthdetect_job.status != "succeeded" %}
  <p>Scoring {{ synthdetect_job.status }}{% if synthdetect_job.error %}: {{ synthdetect_job.error }}{% endif %}</p>
{% else %}

<section>
<h2>Summary</h2>
<table class="key-value">
  <tr><th>Scored turns</th><td>{{ synthdetect_job.scored_turns }} of {{ synthdetect_job.total_turns }}</td></tr>
  {% if synthdetect_job.skipped_turns %}
  <tr><th>Skipped turns</th><td>{{ synthdetect_job.skipped_turns }}</td></tr>
  {% endif %}
  {% if synthdetect_job.mean_risk is not none %}
  <tr>
    <th>Mean risk</th>
    <td><span class="risk-chip {% if synthdetect_job.mean_risk < 0.1 %}risk-low{% elif synthdetect_job.mean_risk < 0.5 %}risk-medium{% else %}risk-high{% endif %}">{{ "%.1f"|format(synthdetect_job.mean_risk * 100) }}%</span></td>
  </tr>
  <tr>
    <th>Max risk</th>
    <td><span class="risk-chip {% if synthdetect_job.max_risk < 0.1 %}risk-low{% elif synthdetect_job.max_risk < 0.5 %}risk-medium{% else %}risk-high{% endif %}">{{ "%.1f"|format(synthdetect_job.max_risk * 100) }}%</span></td>
  </tr>
  {% endif %}
  <tr><th>Calibration policy</th><td><code>{{ synthdetect_job.calibration_policy_id }}</code></td></tr>
  <tr><th>Inference space</th><td><code>{{ synthdetect_job.inference_space }}</code></td></tr>
</table>
</section>

{% if skipped_count > 0 %}
<section>
<h2>Skipped turns</h2>
<p class="muted">Skipped turns were not scored. An unscored turn is not the same
   as a clean turn; the model had insufficient audio to make a determination.</p>
</section>
{% endif %}

{% if high_risk_scores %}
<section>
<h2>Highest-risk turns</h2>
<p class="muted">Turns with the highest calibrated risk scores, descending. Scores
   are calibrated risk values specific to this model and calibration policy, not
   portable probabilities.</p>
<div class="table-wrapper">
<table>
  <thead>
    <tr>
      <th>Turn</th>
      <th>Speaker</th>
      <th>Start</th>
      <th>End</th>
      <th>Risk</th>
      <th>Windows</th>
    </tr>
  </thead>
  <tbody>
    {% for score in high_risk_scores %}
    <tr>
      <td>{{ loop.index }}</td>
      <td>{{ score.speaker_label or "—" }}</td>
      <td>{% if score.start_seconds is not none %}{{ "%.1f"|format(score.start_seconds) }}s{% else %}—{% endif %}</td>
      <td>{% if score.end_seconds is not none %}{{ "%.1f"|format(score.end_seconds) }}s{% else %}—{% endif %}</td>
      <td><span class="risk-chip {% if score.calibrated_score < 0.1 %}risk-low{% elif score.calibrated_score < 0.5 %}risk-medium{% else %}risk-high{% endif %}">{{ "%.1f"|format(score.calibrated_score * 100) }}%</span></td>
      <td>{{ score.window_count }}</td>
    </tr>
    {% endfor %}
  </tbody>
</table>
</div>
</section>
{% endif %}

<section>
<h2>All scored turns</h2>
<div class="table-wrapper">
<table>
  <thead>
    <tr>
      <th>Speaker</th>
      <th>Start</th>
      <th>End</th>
      <th>Risk</th>
      <th>Raw logit</th>
      <th>Windows</th>
      <th>Skip reason</th>
    </tr>
  </thead>
  <tbody>
    {% for score in all_scores %}
    <tr>
      <td>{{ score.speaker_label or "—" }}</td>
      <td>{% if score.start_seconds is not none %}{{ "%.1f"|format(score.start_seconds) }}s{% else %}—{% endif %}</td>
      <td>{% if score.end_seconds is not none %}{{ "%.1f"|format(score.end_seconds) }}s{% else %}—{% endif %}</td>
      <td>{% if score.calibrated_score is not none %}<span class="risk-chip {% if score.calibrated_score < 0.1 %}risk-low{% elif score.calibrated_score < 0.5 %}risk-medium{% else %}risk-high{% endif %}">{{ "%.1f"|format(score.calibrated_score * 100) }}%</span>{% else %}—{% endif %}</td>
      <td>{% if score.raw_logit is not none %}{{ "%.3f"|format(score.raw_logit) }}{% else %}—{% endif %}</td>
      <td>{{ score.window_count }}</td>
      <td>{{ score.skip_reason or "—" }}</td>
    </tr>
    {% endfor %}
  </tbody>
</table>
</div>
</section>

<section>
<h2>Known limitations</h2>
<ul>
  <li><strong>Flow-matching TTS evasion</strong> (<a href="https://github.com/bengizmo/voxint/issues/252">#252</a>):
    Flow-matching text-to-speech systems (Chatterbox, F5-TTS) produce audio that
    this model frequently scores as bona fide. A low risk score does not rule out
    synthetic speech from these generators.</li>
  <li><strong>Channel confounds</strong> (<a href="https://github.com/bengizmo/voxint/issues/253">#253</a>):
    Bona fide speech recorded over low-quality channels (VoIP, telephony,
    heavy compression) can produce elevated false-positive rates. Scores are most
    reliable on clean, studio-quality or near-field recordings.</li>
  <li><strong>Calibration scope</strong>: Scores are calibrated risk values
    produced by the <code>{{ synthdetect_job.calibration_policy_id }}</code>
    policy on the <code>{{ synthdetect_job.inference_space }}</code> model.
    They are not portable probabilities and should not be compared across
    different models or calibration policies.</li>
</ul>
</section>

{% endif %}
</div>
{% endblock %}