app.py

import os
import time

import gradio as gr
import pandas as pd
from constants import (
    ABOUT_TEXT,
    APP_TITLE,
    BANNER,
    CITATION_TEXT,
    DEFAULT_CUSTOM_EVAL_EXAMPLE,
    INTRODUCTION_TEXT,
    LEADERBOARD_CSS,
    tab_label,
    TREBLE_TEAL,
    treble_gradio_theme,
)
import examples_renderer
import analytics
from family_registry import default_family_id
import job_queue
from metrics_config import (
    HEATMAP_SCENARIO_KEYS,
    LIVE_SCENARIO_KEYS,
    SCENARIO_METRICS,
    resolve_scenario_metric_keys,
)
from recipes.registry import RECIPE_CHOICES, resolve_recipe_id

from init import (
    LATEST_VERSION,
    is_model_on_hub,
    list_leaderboard_versions,
    load_raw_results,
    load_results,
    normalize_legacy_csv_row,
    raw_rows_for_version,
)
from utils_display import (
    AutoEvalColumn,
    SCENARIO_DISPLAY_COLS,
    SCENARIO_DISPLAY_TO_KEY,
    column_widths_for,
    fields,
    format_wer_percent,
    model_id_from_leaderboard_cell,
    styled_error,
    styled_message,
    styled_warning,
)

# ---------------------------------------------------------------------------
# Column setup
# ---------------------------------------------------------------------------

COLS = [c.name for c in fields(AutoEvalColumn)]
TYPES = [c.type for c in fields(AutoEvalColumn)]
AVG_WER_COL = AutoEvalColumn.avg_wer_core.name  # "Avg WER (%)"
MODEL_COL = AutoEvalColumn.model.name  # "Model"
ALWAYS_VISIBLE_COLS = (MODEL_COL, AVG_WER_COL)

# Empty table at import; real data loaded on demo.load (avoids blocking Space "preparing").
original_df = pd.DataFrame(columns=COLS)

# Raw CSV rows backing the currently displayed leaderboard version. ``None`` means "live"
# (filter_main_table falls back to load_raw_results()). Set whenever a version is selected
# so the Avg-WER recompute in filter_main_table matches the displayed snapshot.
_active_raw_rows: list[dict] | None = None

_NEXT_UP_PLACEHOLDER = (
    "<div class='next-up-panel' style='font-size:0.8em;opacity:0.85'>"
    "<em>Loading upcoming evaluations…</em></div>"
)

# ---------------------------------------------------------------------------
# Filter (matches open_asr_leaderboard pattern)
# ---------------------------------------------------------------------------


def filter_main_table(search_query, selected_columns):
    filtered_df = original_df.copy()
    selected_columns = list(selected_columns or [])
    if MODEL_COL not in filtered_df.columns:
        return pd.DataFrame(columns=COLS)

    # Filter by model name search (case-insensitive, supports comma-separated terms).
    if search_query:
        terms = [t.strip() for t in search_query.split(",") if t.strip()]
        if terms:
            mask = filtered_df[MODEL_COL].str.lower().apply(
                lambda cell: any(term.lower() in cell for term in terms)
            )
            filtered_df = filtered_df[mask]

    # Hide toggleable columns that the user unchecked. Model + Avg WER stay visible.
    all_toggleable = [c for c in COLS if c not in ALWAYS_VISIBLE_COLS]
    columns_to_hide = set(all_toggleable) - set(selected_columns)
    filtered_df = filtered_df[[c for c in filtered_df.columns if c not in columns_to_hide]]

    raw_rows = _active_raw_rows if _active_raw_rows is not None else load_raw_results()
    for r in raw_rows:
        normalize_legacy_csv_row(r)
    raw_by_mid = {r["model_id"]: r for r in raw_rows}

    visible_metric_keys = [
        SCENARIO_DISPLAY_TO_KEY[c]
        for c in SCENARIO_DISPLAY_COLS
        if c in selected_columns
    ]

    def _row_avg_wer(rdict: dict) -> float | None:
        if not visible_metric_keys:
            return None
        vals = []
        for k in visible_metric_keys:
            try:
                v = rdict.get(k, "")
                if v is None or str(v).strip() == "":
                    continue
                vals.append(float(v))
            except (TypeError, ValueError):
                continue
        if not vals:
            return None
        return sum(vals) / len(vals)

    def _avg_wer_key(row) -> tuple[int, float]:
        mid = model_id_from_leaderboard_cell(row[MODEL_COL])
        rdict = raw_by_mid.get(mid)
        if not rdict:
            return (1, float("inf"))
        avg = _row_avg_wer(rdict)
        if avg is None:
            return (1, float("inf"))
        return (0, avg)

    if AVG_WER_COL in filtered_df.columns:
        def _display_avg(row):
            mid = model_id_from_leaderboard_cell(row[MODEL_COL])
            rdict = raw_by_mid.get(mid)
            if not rdict:
                return "NA"
            avg = _row_avg_wer(rdict)
            if avg is None:
                return "NA"
            return format_wer_percent(avg)

        filtered_df[AVG_WER_COL] = filtered_df.apply(_display_avg, axis=1)

    filtered_df = filtered_df.assign(_sort=filtered_df.apply(_avg_wer_key, axis=1))
    filtered_df = filtered_df.sort_values(by="_sort").drop(columns="_sort")
    return filtered_df


def _leaderboard_table_update(search_query, selected_columns):
    """Return a Gradio update with table data and matching column widths."""
    df = filter_main_table(search_query, selected_columns)
    return gr.update(value=df, column_widths=column_widths_for(list(df.columns)))


def _format_version_recorded_at(ts: str) -> str:
    from datetime import datetime

    try:
        dt = datetime.fromisoformat(ts.replace("Z", "+00:00"))
        return dt.strftime("%Y-%m-%d %H:%M UTC")
    except Exception:
        return str(ts)


def _version_choices() -> list[tuple[str, str]]:
    """(label, value) pairs for the version dropdown; first entry is always 'Latest'."""
    choices: list[tuple[str, str]] = [("Latest", LATEST_VERSION)]
    try:
        for v in list_leaderboard_versions():
            label = f"{_format_version_recorded_at(v['recorded_at'])} — {v['label']}"
            choices.append((label, v["version"]))
    except Exception:
        pass
    return choices


# ---------------------------------------------------------------------------
# Submission handler
# ---------------------------------------------------------------------------


def _wer_cell(row: dict, key: str) -> str:
    from utils_display import format_wer_percent

    normalize_legacy_csv_row(row)
    v = row.get(key, "")
    if v is None or (isinstance(v, str) and str(v).strip() == ""):
        return "N/A"
    pct = format_wer_percent(v)
    return f"{pct}%" if pct != "NA" else "N/A"


def _model_hub_page_link(model_id: str) -> str:
    mid = (model_id or "").strip().replace(" ", "")
    if not mid or "/" not in mid:
        return (
            "<p style='font-size:0.8em;opacity:0.85'>"
            "Enter a model id (<code>author/name</code>) to open its Hub page.</p>"
        )
    url = f"https://huggingface.co/{mid}"
    return (
        f"<p style='font-size:0.8em'>"
        f"<a href='{url}' target='_blank' rel='noopener noreferrer'>"
        f"Open model page on the Hub</a> — request access if the repo is gated.</p>"
    )


def submit_model(
    model_id: str,
    submission_notes: str,
    contact_email: str,
    extra_requirements: str,
    setup_script: str,
    custom_script: str,
    recipe_id: str,
    is_gated: bool,
) -> str:
    """Validate and enqueue evaluation; a background worker runs approved jobs (up to 4 parallel Hub Jobs when remote mode is on).

    The backend always uses the "Auto" family (pipeline → Granite → universal → CTC),
    which covers every checkpoint we currently support, so there is no per-submission family choice.
    """
    model_id = model_id.strip()

    if not model_id:
        return styled_error("Please enter a model ID.")

    family_id = default_family_id()

    existing = load_raw_results()
    for row in existing:
        if row["model_id"] == model_id:
            return styled_message(
                f"Model '{model_id}' has already been evaluated. "
                f"Near Field Speech: {_wer_cell(row, 'wer_anechoic_speech')} | "
                f"Lab Measured: {_wer_cell(row, 'wer_lab_measured')} | "
                f"Lab Simulated: {_wer_cell(row, 'wer_lab_simulated')} | "
                f"High SNR: {_wer_cell(row, 'wer_realistic_high_snr')} | "
                f"Mid SNR: {_wer_cell(row, 'wer_realistic_mid_snr')} | "
                f"Low SNR: {_wer_cell(row, 'wer_realistic_low_snr')} | "
                f"Moving Sources (Low SNR): {_wer_cell(row, 'wer_moving_sources')}"
            )

    # API / recipe / custom-script models (e.g. zoom/scribe_v1) are not Hub repos, so the
    # Hub-existence check would wrongly reject them. Skip it whenever a recipe resolves or a
    # custom evaluator is supplied; otherwise validate the model id against the Hub.
    resolved_recipe = resolve_recipe_id(model_id, (recipe_id or "").strip() or None)
    has_custom_script = bool((custom_script or "").strip())
    if not resolved_recipe and not has_custom_script:
        on_hub, err_msg = is_model_on_hub(model_id)
        if not on_hub:
            return styled_error(f"Model '{model_id}' {err_msg}")

    if not (contact_email or "").strip():
        return styled_error("Please enter a contact email address.")

    try:
        job_queue.sanitize_contact_email(contact_email)
    except ValueError as exc:
        msg = str(exc)
        if msg == "Contact email is required.":
            return styled_error("Please enter a contact email address.")
        return styled_error("Please enter a valid email address.")

    script_hint = ""
    if (custom_script or "").strip():
        dep_warn = job_queue.custom_script_deprecated_api_warning(custom_script)
        if dep_warn:
            script_hint += f"<p style='color:orange'><strong>Note:</strong> {dep_warn}</p>"
        argparse_warn = job_queue.custom_script_argparse_warning(custom_script)
        if argparse_warn:
            script_hint += f"<p style='color:orange'><strong>Note:</strong> {argparse_warn}</p>"
        if not job_queue.custom_script_defines_evaluate(custom_script):
            script_hint += (
                "<p style='color:orange'><strong>Note:</strong> Your custom script should define "
                "<code>evaluate(file: pathlib.Path) -&gt; str</code> at the top level. "
                "It will be called once per sample during evaluation.</p>"
            )

    try:
        job_id, position, err, awaiting_mod = job_queue.enqueue(
            model_id,
            family_id,
            submission_notes=submission_notes or "",
            contact_email=contact_email or "",
            extra_requirements=extra_requirements or "",
            setup_script=setup_script or "",
            custom_script=custom_script or "",
            recipe_id=(recipe_id or "").strip(),
            is_gated=bool(is_gated),
        )
    except Exception as exc:
        return styled_warning(f"Could not enqueue submission: {exc}")
    if err == "remote_jobs_required":
        return styled_error(
            "Hub Jobs are required. Set Space secrets "
            "<code>FFASR_REMOTE_JOBS=1</code>, <code>token_for_ffasr_jobs</code>, "
            "<code>HF_TOKEN</code>, and <code>FFASR_REMOTE_EVAL_REPO_URL</code>."
        )
    if err == "moderation_misconfigured":
        return styled_error(
            "This Space has moderation enabled but <code>FFASR_MODERATOR_SECRET</code> is not set. "
            "Ask the owner to add it under Space Settings → Secrets."
        )
    if err == "pending_moderation_full":
        return styled_warning(
            "Too many models are awaiting moderator approval. Please try again later."
        )
    if err == "already_in_csv":
        existing = load_raw_results()
        for row in existing:
            if row["model_id"] == model_id:
                return styled_message(
                    f"Model '{model_id}' has already been evaluated. "
                    f"Near Field Speech: {_wer_cell(row, 'wer_anechoic_speech')} | "
                    f"Lab Measured: {_wer_cell(row, 'wer_lab_measured')} | "
                    f"Lab Simulated: {_wer_cell(row, 'wer_lab_simulated')} | "
                    f"High SNR: {_wer_cell(row, 'wer_realistic_high_snr')} | "
                    f"Mid SNR: {_wer_cell(row, 'wer_realistic_mid_snr')} | "
                    f"Low SNR: {_wer_cell(row, 'wer_realistic_low_snr')} | "
                    f"Moving Sources (Low SNR): {_wer_cell(row, 'wer_moving_sources')}"
                )
        return styled_error("Could not enqueue; please try again.")
    if err == "queue_full":
        return styled_warning(
            "The evaluation queue is full. Please try again later."
        )
    if err:
        return styled_warning(err)

    if awaiting_mod:
        return (
            script_hint
            + styled_message(
                f"Request <code>{job_id}</code> recorded for <strong>{model_id}</strong>. "
                f"<strong>A moderator must approve</strong> it before evaluation starts "
                f"(see the <strong>Moderate</strong> tab). "
                f"Approx. backlog awaiting approval: <strong>{position}</strong>."
            )
        )

    return (
        script_hint
        + styled_message(
            f"Queued job <code>{job_id}</code> for <strong>{model_id}</strong>. "
            f"Approx. position in queue: <strong>{position}</strong>. "
            f"Evaluation runs in the background; refresh the <strong>Leaderboard</strong> tab "
            f"after a few minutes to see WER when the job finishes."
        )
    )


# ---------------------------------------------------------------------------
# Analysis tab: radar + line + bar charts
# ---------------------------------------------------------------------------

_METRIC_BY_KEY = {m.key: m for m in SCENARIO_METRICS}
_METRIC_CHOICES = [
    (
        f"{_METRIC_BY_KEY[k].short}: {_METRIC_BY_KEY[k].label}"
        + (" (planned)" if _METRIC_BY_KEY[k].status == "planned" else ""),
        k,
    )
    for k in HEATMAP_SCENARIO_KEYS
    if k in _METRIC_BY_KEY
]

def _analytics_initial():
    raw = load_raw_results()
    df = analytics._raw_to_analytics_df(raw)
    ids = df["model_id"].tolist() if not df.empty else []
    avail = analytics.available_metric_keys(df)
    line_default = [k for k in HEATMAP_SCENARIO_KEYS if k in avail] or list(avail)
    tn = 10
    fig_intel = analytics.plot_avg_wer_bars(df, top_n=tn)
    fig_speed = analytics.plot_speed_bars(df, top_n=tn)
    fig_hm = analytics.plot_scenario_heatmap(df, line_default, top_n=tn)
    fig_pareto = analytics.plot_pareto_frontier(df)
    # fig_radar = analytics.plot_robustness_radar(df, sel)
    fig_b = analytics.plot_scenario_bar_summary(df, 8)
    return (
        gr.update(value=line_default),
        fig_intel,
        fig_speed,
        fig_hm,
        fig_pareto,
        fig_b,
    )


def _analytics_apply(line_keys, top_n):
    raw = load_raw_results()
    df = analytics._raw_to_analytics_df(raw)
    avail = analytics.available_metric_keys(df)
    lk = resolve_scenario_metric_keys(line_keys, _METRIC_CHOICES)
    if not lk:
        lk = [k for k in HEATMAP_SCENARIO_KEYS if k in avail] or list(avail)
    tn = int(top_n) if top_n else 10
    fig_intel = analytics.plot_avg_wer_bars(df, top_n=tn)
    fig_speed = analytics.plot_speed_bars(df, top_n=tn)
    fig_hm = analytics.plot_scenario_heatmap(df, lk, top_n=tn)
    fig_pareto = analytics.plot_pareto_frontier(df)
    # fig_radar = analytics.plot_robustness_radar(df, valid_models)
    fig_b = analytics.plot_scenario_bar_summary(df, tn)
    return fig_intel, fig_speed, fig_hm, fig_pareto, fig_b


# ---------------------------------------------------------------------------
# Gradio App
# ---------------------------------------------------------------------------

_theme = treble_gradio_theme()

with gr.Blocks(title=APP_TITLE, theme=_theme, css=LEADERBOARD_CSS) as demo:
    gr.HTML(BANNER, elem_id="banner")
    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")

    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        with gr.TabItem(tab_label("leaderboard", "Leaderboard"), elem_id="od-benchmark-tab-table", id=0):
            with gr.Row():
                search_box = gr.Textbox(
                    label="Search models",
                    placeholder="Filter by model name -- separate multiple terms with commas, e.g. whisper, nvidia",
                    scale=4,
                )
                version_dropdown = gr.Dropdown(
                    choices=_version_choices(),
                    value=LATEST_VERSION,
                    label="Version",
                    scale=2,
                    min_width=200,
                )
                leaderboard_update_btn = gr.Button(
                    "Update", variant="primary", scale=1, min_width=120
                )
            toggleable_columns = [c for c in COLS if c not in ALWAYS_VISIBLE_COLS]
            _default_visible_cols = [
                c
                for c in toggleable_columns
                if c != AutoEvalColumn.wer_lab_measured.name
                and c != AutoEvalColumn.wer_lab_simulated.name
                and c != AutoEvalColumn.wer_moving_low.name
                and c != AutoEvalColumn.wer_moving_mid.name
                and c != AutoEvalColumn.wer_moving_high.name
            ]
            column_checkboxes = gr.CheckboxGroup(
                choices=toggleable_columns,
                value=_default_visible_cols,
                label="Select columns to display",
            )
            leaderboard_table = gr.components.Dataframe(
                value=original_df,
                datatype=TYPES,
                elem_id="leaderboard-table",
                interactive=False,
                visible=True,
                wrap=False,
                column_widths=column_widths_for(COLS),
            )
            gr.Markdown(
                "\* Moving-source evaluations are a beta feature.",
                elem_classes="markdown-text",
            )

            def _set_active_version(version):
                """Load the chosen version into the module globals used by filter_main_table."""
                global original_df, _active_raw_rows
                if version:
                    _active_raw_rows = raw_rows_for_version(version)
                    original_df = load_results(version)
                else:
                    _active_raw_rows = None
                    original_df = load_results()

            def _leaderboard_refresh(search, cols, version):
                """Reload results from storage for the selected version, refresh version list."""
                from init import invalidate_results_cache

                try:
                    invalidate_results_cache()
                    choices = _version_choices()
                    valid = {v for _, v in choices}
                    if version not in valid:
                        version = LATEST_VERSION
                    _set_active_version(version)
                    return (
                        _leaderboard_table_update(search, cols),
                        gr.update(choices=choices, value=version),
                    )
                except Exception:
                    return gr.update(
                        value=original_df,
                        column_widths=column_widths_for(list(original_df.columns)),
                    ), gr.update()

            def _on_version_change(version, search, cols):
                """Swap the displayed leaderboard to the selected version, keep filters."""
                try:
                    _set_active_version(version)
                    return _leaderboard_table_update(search, cols)
                except Exception:
                    return gr.update(
                        value=original_df,
                        column_widths=column_widths_for(list(original_df.columns)),
                    )

            def _on_startup():
                """Load leaderboard + next-up list after UI is up (keeps Space prepare step fast)."""
                global original_df, _active_raw_rows
                from init import invalidate_results_cache

                version_update = gr.update()
                try:
                    invalidate_results_cache()
                    _active_raw_rows = None
                    original_df = load_results()
                    table = _leaderboard_table_update("", _default_visible_cols)
                    version_update = gr.update(choices=_version_choices(), value=LATEST_VERSION)
                except Exception as exc:
                    table = gr.update(
                        value=original_df,
                        column_widths=column_widths_for(list(original_df.columns)),
                    )
                    next_html = (
                        "<div class='next-up-panel'><p style='color:orange'>"
                        f"<strong>Startup warning:</strong> {exc}</p></div>"
                    )
                    return table, next_html, version_update
                try:
                    job_queue.ensure_worker_started()
                    next_html = job_queue.next_up_html()
                except Exception as exc:
                    next_html = (
                        "<div class='next-up-panel'><p style='color:orange'>"
                        f"<strong>Queue unavailable:</strong> {exc}</p></div>"
                    )
                return table, next_html, version_update

            filter_inputs = [search_box, column_checkboxes]
            search_box.change(_leaderboard_table_update, inputs=filter_inputs, outputs=leaderboard_table)
            column_checkboxes.change(_leaderboard_table_update, inputs=filter_inputs, outputs=leaderboard_table)
            version_dropdown.change(
                fn=_on_version_change,
                inputs=[version_dropdown, search_box, column_checkboxes],
                outputs=leaderboard_table,
            )
            leaderboard_update_btn.click(
                fn=_leaderboard_refresh,
                inputs=[search_box, column_checkboxes, version_dropdown],
                outputs=[leaderboard_table, version_dropdown],
            )

        with gr.TabItem(tab_label("submit", "Submit"), elem_id="submit-tab", id=1):
            gr.Markdown("## Submit a model for evaluation")
            gr.Markdown(
                "Paste a Hugging Face model id. The server loads every checkpoint with an automatic backend, so "
                "**Whisper, IBM Granite speech, Cohere Transcribe, efficient‑speech / custom Whisper variants, "
                "Wav2Vec2 / HuBERT CTC heads, and SpeechBrain ASR**, plus most other ASR stacks on the Hub are "
                "supported without a per‑submission choice.\n\n"
                "Submissions are **queued**; when Hub Jobs are enabled, up to **four** models evaluate in parallel. "
                "Otherwise jobs run **one at a time** on this Space. "
                "When moderation is enabled, new requests wait for approval before they run."
            )

            with gr.Row():
                model_input = gr.Textbox(
                    label="Model ID",
                    placeholder="e.g. openai/whisper-tiny",
                    scale=3,
                )
                submit_btn = gr.Button("Evaluate", variant="primary", scale=1)

            status_output = gr.HTML()

            recipe_input = gr.Dropdown(
                label="Complex model recipe (optional)",
                choices=RECIPE_CHOICES,
                value="",
                info="Pre-fills setup script, evaluate(), and deps for install-heavy stacks (e.g. Mega-ASR).",
            )
            requirements_input = gr.Textbox(
                label="Optional extra Python requirements (one per line, requirements.txt format)",
                placeholder="transformers @ git+https://github.com/huggingface/transformers.git\nsentencepiece",
                lines=4,
                max_length=8000,
            )
            setup_script_input = gr.Textbox(
                label="Optional one-time setup script (shell or Python, runs once per Hub Job)",
                placeholder="git clone …\npip install -r requirements.txt\npython scripts/download.py",
                lines=6,
                max_length=8000,
            )
            script_input = gr.Code(
                label="Optional custom evaluator (Python)",
                language="python",
                lines=16,
                value=DEFAULT_CUSTOM_EVAL_EXAMPLE,
            )
            gr.Markdown(
                "Edit the example below: define **`evaluate(file: Path) -> str`** to transcribe one WAV. "
                "Load your model **once** at module level (as in the Cohere example). "
                "Use **soundfile** (or numpy) to read each WAV (see the example). "
                "Hub Jobs disable torchcodec and fall back to librosa if you use "
                "``transformers.audio_utils.load_audio`` on file paths. "
                "The function is called **once per audio sample** inside the eval loop. "
                "**Model size:** expose your loaded model as a module-level `model` "
                "(or set `NUM_PARAMS = <int>`) so the leaderboard can report its "
                "parameter count; otherwise size is left blank. "
                "Put extra Python dependencies in the requirements box above. "
                "Use the **setup script** for git clone / weight download (runs once before evaluation). "
                "**No maintainer recipe is needed for a new model:** if your `evaluate()` imports a "
                "cloned package (e.g. `from MyASR.model import MyASR`), the worker auto-discovers it "
                "under `/tmp`. For full control, export paths from your setup script via "
                "`echo \"PYTHONPATH=/tmp/MyASR/src\" >> \"$FFASR_ENV_FILE\"` (also accepts "
                "`FFASR_IMPORT_PATHS` and any `KEY=VALUE`). "
                "See [Mega-ASR recipe](docs/recipes/mega_asr.md) for an example. "
                "Custom evaluators run on a Hub Job **only** after a moderator approves them."
            )
            with gr.Row():
                is_gated_input = gr.Checkbox(label="This is a gated repo", value=False)
            gated_link = gr.HTML(value=_model_hub_page_link(""))
            model_input.change(fn=_model_hub_page_link, inputs=[model_input], outputs=[gated_link])
            notes_input = gr.Textbox(
                label="Optional free-form notes for moderators",
                placeholder="Repo link, paper, eval caveats, hardware notes…",
                lines=2,
                max_length=4000,
            )
            email_input = gr.Textbox(
                label="Contact email",
                placeholder="you@example.com",
                max_length=254,
            )

            gr.Markdown(
                "### Contact\n\n"
                "Questions about submissions, gated repos, or evaluation issues? "
                "Email [contact@treble.tech](mailto:contact@treble.tech)."
            )
            gr.Markdown("### Next models to evaluate")
            next_up_panel = gr.HTML(value=_NEXT_UP_PLACEHOLDER, elem_classes="next-up-panel")

            def _apply_recipe_fields(model_id, recipe_id, reqs, setup, script):
                from recipes.registry import apply_recipe_to_submission

                extra, setup_out, custom, _rid = apply_recipe_to_submission(
                    (model_id or "").strip(),
                    (recipe_id or "").strip() or None,
                    reqs or "",
                    setup or "",
                    script or "",
                )
                return extra, setup_out, custom

            def _submit_and_refresh(model_id, notes, email, reqs, setup, script, recipe, gated):
                status = submit_model(model_id, notes, email, reqs, setup, script, recipe, gated)
                try:
                    nxt = job_queue.next_up_html()
                except Exception as exc:
                    nxt = f"<p style='color:orange'>Could not load queue: {exc}</p>"
                return status, nxt

            recipe_input.change(
                fn=_apply_recipe_fields,
                inputs=[
                    model_input,
                    recipe_input,
                    requirements_input,
                    setup_script_input,
                    script_input,
                ],
                outputs=[requirements_input, setup_script_input, script_input],
            )

            submit_btn.click(
                fn=_submit_and_refresh,
                inputs=[
                    model_input,
                    notes_input,
                    email_input,
                    requirements_input,
                    setup_script_input,
                    script_input,
                    recipe_input,
                    is_gated_input,
                ],
                outputs=[status_output, next_up_panel],
            )

        with gr.TabItem(tab_label("moderate", "Moderate"), elem_id="moderate-tab", id=2):
            from benchmark.dataset import CONDITION_UI_CHOICES

            _MOD_COND_CHOICES = [(lbl, key) for lbl, key in CONDITION_UI_CHOICES]
            _MOD_COND_DEFAULT = [key for _, key in CONDITION_UI_CHOICES]

            gr.Markdown("## Moderator access")
            mod_unlock_feedback = gr.HTML()
            mod_secret = gr.Textbox(
                label="Moderator secret",
                type="password",
                placeholder="FFASR_MODERATOR_SECRET",
            )
            mod_unlock_btn = gr.Button("Unlock moderator tools", variant="primary")
            mod_unlocked = gr.State(False)

            with gr.Column(visible=False) as mod_panel:
                gr.Markdown(
                    "When <code>FFASR_MODERATION=1</code> is set, new submissions **wait here** until you approve them. "
                    "Use per-row **Check**, **Approve**, **Retry**, and **Remove** buttons — popups open for "
                    "submission review, dataset selection, and approval.\n\n"
                    "When <code>FFASR_REMOTE_JOBS=1</code>, use **Open Hub Job logs** while a job runs. "
                    "The bucket <code>results/remote_artifacts/</code> folder only gets a JSON file **after** a successful run.\n\n"
                    "**Retry** re-queues failed, done, or queued jobs (not while running). "
                    "Uncheck datasets in the retry popup to run only selected packed splits; "
                    "other leaderboard WER columns are left unchanged on success. "
                    "A full re-run (all datasets selected) **replaces** the row when the model is already listed."
                )
                gr.Markdown("### Current job progress")
                mod_progress = gr.HTML(
                    value=job_queue.moderation_locked_placeholder_html(),
                    elem_classes="ffasr-card",
                )
                mod_refresh_tick = gr.Number(value=0, visible=False)
                mod_refresh = gr.Button("Refresh lists")
                mod_feedback = gr.HTML()

                mod_check_job_id = gr.State("")
                mod_retry_job_id = gr.State("")
                mod_approve_job_id = gr.State("")

                with gr.Column(visible=False) as check_panel:
                    check_header = gr.Markdown("### Check submission")
                    check_notes = gr.Textbox(
                        label="Submitter notes",
                        lines=2,
                        interactive=False,
                    )
                    check_recipe = gr.Dropdown(
                        label="Recipe",
                        choices=RECIPE_CHOICES,
                        value="",
                    )
                    check_reqs = gr.Textbox(
                        label="Extra Python requirements (one per line)",
                        lines=4,
                        max_length=8000,
                    )
                    check_setup = gr.Textbox(
                        label="One-time setup script (shell or Python)",
                        lines=6,
                        max_length=8000,
                    )
                    check_script = gr.Code(
                        label="Custom evaluator (Python)",
                        language="python",
                        lines=16,
                    )
                    with gr.Row():
                        check_save_btn = gr.Button("Save", variant="primary")
                        check_cancel_btn = gr.Button("Cancel")

                with gr.Column(visible=False) as retry_panel:
                    retry_header = gr.Markdown("### Retry job")
                    retry_eval_conditions = gr.CheckboxGroup(
                        choices=_MOD_COND_CHOICES,
                        value=_MOD_COND_DEFAULT,
                        label="Datasets to evaluate (all selected = full benchmark)",
                    )
                    with gr.Row():
                        retry_confirm_btn = gr.Button(
                            "Confirm retry", variant="primary"
                        )
                        retry_cancel_btn = gr.Button("Cancel")

                with gr.Column(visible=False) as approve_panel:
                    approve_header = gr.Markdown("### Approve job")
                    approve_run_custom = gr.Checkbox(
                        label="Use submitter's custom evaluate() function (if provided)",
                        value=False,
                    )
                    approve_eval_conditions = gr.CheckboxGroup(
                        choices=_MOD_COND_CHOICES,
                        value=_MOD_COND_DEFAULT,
                        label="Datasets to evaluate (all selected = full benchmark)",
                    )
                    with gr.Row():
                        approve_confirm_btn = gr.Button("Approve", variant="primary")
                        approve_cancel_btn = gr.Button("Cancel")

                _MOD_POPUP_HIDE = (
                    gr.update(visible=False),
                    gr.update(visible=False),
                    gr.update(visible=False),
                )

                def _mod_bump_tick(tick):
                    return int(tick or 0) + 1

                def _mod_feedback_html(ok: bool, msg: str) -> str:
                    color = "green" if ok else "red"
                    return f"<p style='color:{color}'>{msg}</p>"

                def _mod_open_check(jid: str, secret: str, unlocked: bool):
                    hidden = gr.update(visible=False)
                    if not unlocked:
                        return (
                            gr.update(visible=False),
                            hidden,
                            hidden,
                            "",
                            "### Check submission",
                            "",
                            "",
                            "",
                            "",
                            "",
                        )
                    job = job_queue.peek_job(str(jid).strip())
                    if not job:
                        return (
                            gr.update(visible=False),
                            hidden,
                            hidden,
                            "",
                            "### Check submission",
                            "",
                            "",
                            "",
                            "",
                            "",
                        )
                    return (
                        gr.update(visible=True),
                        hidden,
                        hidden,
                        job.id,
                        f"### Check submission: `{job.model_id}`",
                        job.submission_notes or "",
                        job.recipe_id or "",
                        job.extra_requirements or "",
                        job.setup_script or "",
                        job.custom_script or "",
                    )

                def _mod_open_retry(jid: str, secret: str, unlocked: bool):
                    hidden = gr.update(visible=False)
                    if not unlocked:
                        return (
                            hidden,
                            gr.update(visible=False),
                            hidden,
                            "",
                            "### Retry job",
                        )
                    job = job_queue.peek_job(str(jid).strip())
                    if not job:
                        return (
                            hidden,
                            gr.update(visible=False),
                            hidden,
                            "",
                            "### Retry job",
                        )
                    return (
                        hidden,
                        gr.update(visible=True),
                        hidden,
                        job.id,
                        f"### Retry: `{job.model_id}` (`{job.id}`)",
                    )

                def _mod_open_approve(jid: str, secret: str, unlocked: bool):
                    hidden = gr.update(visible=False)
                    if not unlocked:
                        return (
                            hidden,
                            hidden,
                            gr.update(visible=False),
                            "",
                            "### Approve job",
                            False,
                        )
                    job = job_queue.peek_job(str(jid).strip())
                    if not job:
                        return (
                            hidden,
                            hidden,
                            gr.update(visible=False),
                            "",
                            "### Approve job",
                            False,
                        )
                    has_script = bool(
                        (job.custom_script or "").strip()
                        or (job.setup_script or "").strip()
                        or (job.recipe_id or "").strip()
                    )
                    return (
                        hidden,
                        hidden,
                        gr.update(visible=True),
                        job.id,
                        f"### Approve: `{job.model_id}` (`{job.id}`)",
                        has_script,
                    )

                def _mod_reject_row(jid, secret, unlocked, tick):
                    if not unlocked:
                        return (
                            "<p style='color:red'>Unlock moderator tools first.</p>",
                            tick,
                            job_queue.progress_html(),
                        )
                    ok, msg = job_queue.reject_job(str(jid).strip(), secret)
                    return (
                        _mod_feedback_html(ok, msg),
                        _mod_bump_tick(tick),
                        job_queue.progress_html(),
                    )

                def _mod_remove_row(jid, secret, unlocked, tick):
                    if not unlocked:
                        return (
                            "<p style='color:red'>Unlock moderator tools first.</p>",
                            tick,
                            job_queue.progress_html(),
                        )
                    ok, msg = job_queue.remove_job_entry(str(jid).strip(), secret)
                    return (
                        _mod_feedback_html(ok, msg),
                        _mod_bump_tick(tick),
                        job_queue.progress_html(),
                    )

                def _mod_save_check(jid, secret, unlocked, tick, recipe, reqs, setup, script):
                    if not unlocked:
                        return (
                            "<p style='color:red'>Unlock moderator tools first.</p>",
                            tick,
                            *_MOD_POPUP_HIDE,
                        )
                    ok, msg = job_queue.update_job_script_and_requirements(
                        str(jid).strip(),
                        secret,
                        extra_requirements=reqs or "",
                        setup_script=setup or "",
                        custom_script=script or "",
                        recipe_id=(recipe or "").strip(),
                    )
                    return (
                        _mod_feedback_html(ok, msg),
                        _mod_bump_tick(tick),
                        *_MOD_POPUP_HIDE,
                    )

                def _mod_confirm_retry(jid, secret, unlocked, tick, eval_conds):
                    if not unlocked:
                        return (
                            "<p style='color:red'>Unlock moderator tools first.</p>",
                            tick,
                            *_MOD_POPUP_HIDE,
                            job_queue.progress_html(),
                        )
                    ok, msg = job_queue.retry_failed_job(
                        str(jid).strip(),
                        secret,
                        eval_conditions=list(eval_conds or []),
                    )
                    return (
                        _mod_feedback_html(ok, msg),
                        _mod_bump_tick(tick),
                        *_MOD_POPUP_HIDE,
                        job_queue.progress_html(),
                    )

                def _mod_confirm_approve(
                    jid, secret, unlocked, tick, run_custom, eval_conds
                ):
                    if not unlocked:
                        return (
                            "<p style='color:red'>Unlock moderator tools first.</p>",
                            tick,
                            *_MOD_POPUP_HIDE,
                            job_queue.progress_html(),
                        )
                    ok, msg = job_queue.approve_job(
                        str(jid).strip(),
                        secret,
                        run_custom_script=bool(run_custom),
                        eval_conditions=list(eval_conds or []),
                    )
                    return (
                        _mod_feedback_html(ok, msg),
                        _mod_bump_tick(tick),
                        *_MOD_POPUP_HIDE,
                        job_queue.progress_html(),
                    )

                def _mod_cancel_popup(tick):
                    return tick, *_MOD_POPUP_HIDE

                def _mod_retry_all(secret, unlocked, tick):
                    if not unlocked:
                        return (
                            "<p style='color:orange'>Unlock moderator tools first.</p>",
                            tick,
                            job_queue.progress_html(),
                        )
                    ok, msg = job_queue.retry_all_eligible_jobs(secret)
                    return (
                        _mod_feedback_html(ok, msg),
                        _mod_bump_tick(tick),
                        job_queue.progress_html(),
                    )

                def _mod_import_artifact(
                    secret, artifact_ref, replace_existing, import_notes, unlocked, tick
                ):
                    if not unlocked:
                        return (
                            "<p style='color:red'>Unlock moderator tools first.</p>",
                            tick,
                        )
                    ok, msg = job_queue.import_artifact_to_leaderboard(
                        artifact_ref,
                        secret,
                        replace_existing=bool(replace_existing),
                        submission_notes=import_notes or "",
                    )
                    return (_mod_feedback_html(ok, msg), _mod_bump_tick(tick))

                def _mod_refresh_lists(secret, unlocked, tick):
                    if not unlocked:
                        return tick, job_queue.moderation_locked_placeholder_html()
                    ok, msg = job_queue.verify_moderator_secret(secret)
                    if not ok:
                        return tick, job_queue.moderation_locked_placeholder_html()
                    return _mod_bump_tick(tick), job_queue.progress_html()

                gr.Markdown("### Pending moderation")

                @gr.render(inputs=[mod_refresh_tick])
                def _render_pending_jobs(_tick):
                    try:
                        jobs = job_queue.pending_jobs_for_render()
                    except Exception as exc:
                        gr.Markdown(f"*Queue unavailable: {exc}*")
                        return
                    if not jobs:
                        gr.Markdown("*No jobs awaiting approval.*")
                        return
                    for job in jobs:
                        jid = job["id"]
                        row_cls = job_queue.job_row_elem_classes(job["status"])
                        gated = " · **gated**" if job["is_gated"] else ""
                        script_bit = (
                            " · custom/setup/recipe"
                            if job["has_custom_script"]
                            else ""
                        )
                        extras = (
                            f"{job['req_count']} extra req(s)"
                            if job["req_count"]
                            else "no extra reqs"
                        )
                        email_bit = (
                            f" · 📧 `{job['contact_email']}`"
                            if job.get("contact_email")
                            else " · 📧 *(none)*"
                        )
                        with gr.Row(elem_classes=row_cls):
                            gr.Markdown(
                                f"**`{jid}`** · `{job['model_id']}` · "
                                f"`{job['family_id']}`{gated}{script_bit} · {extras}{email_bit} · "
                                f"{job['notes_preview'] or '—'} · {job['created_at']}",
                                elem_classes="ffasr-job-info",
                            )
                            check_btn = gr.Button("Check", size="sm")
                            approve_btn = gr.Button(
                                "Approve", size="sm", variant="primary"
                            )
                            reject_btn = gr.Button("Reject", size="sm", variant="stop")
                        check_btn.click(
                            fn=_mod_open_check,
                            inputs=[gr.State(jid), mod_secret, mod_unlocked],
                            outputs=[
                                check_panel,
                                retry_panel,
                                approve_panel,
                                mod_check_job_id,
                                check_header,
                                check_notes,
                                check_recipe,
                                check_reqs,
                                check_setup,
                                check_script,
                            ],
                        )
                        approve_btn.click(
                            fn=_mod_open_approve,
                            inputs=[gr.State(jid), mod_secret, mod_unlocked],
                            outputs=[
                                check_panel,
                                retry_panel,
                                approve_panel,
                                mod_approve_job_id,
                                approve_header,
                                approve_run_custom,
                            ],
                        )
                        reject_btn.click(
                            fn=_mod_reject_row,
                            inputs=[gr.State(jid), mod_secret, mod_unlocked, mod_refresh_tick],
                            outputs=[mod_feedback, mod_refresh_tick, mod_progress],
                        )

                gr.Markdown("### Recent job activity")
                retry_all_btn = gr.Button(
                    "Retry ALL eligible jobs (full benchmark)",
                    variant="secondary",
                )
                gr.Markdown(
                    "<small><em>Retry ALL</em> re-queues every <strong>failed / done / queued</strong> "
                    "job against the full dataset set (all packed conditions). Running, dispatching, and "
                    "pending-moderation jobs are skipped. Successful re-runs replace the existing "
                    "leaderboard row for each model.</small>"
                )

                @gr.render(inputs=[mod_refresh_tick])
                def _render_recent_jobs(_tick):
                    try:
                        jobs = job_queue.recent_jobs_for_render(30)
                    except Exception as exc:
                        gr.Markdown(f"*Queue unavailable: {exc}*")
                        return
                    if not jobs:
                        gr.Markdown("*No job history loaded yet.*")
                        return
                    for job in jobs:
                        jid = job["id"]
                        err = job["error"] or "N/A"
                        hub = job["hub_link_html"] or ""
                        row_cls = job_queue.job_row_elem_classes(job["status"])
                        email_bit = (
                            f" · 📧 `{job['contact_email']}`"
                            if job.get("contact_email")
                            else " · 📧 *(none)*"
                        )
                        info_md = (
                            f"**`{jid}`** · `{job['model_id']}` · "
                            f"**{job['status']}**{email_bit} · {err} · {job['updated_at']}"
                        )
                        if hub:
                            info_md += f" · {hub}"
                        with gr.Row(elem_classes=row_cls):
                            gr.Markdown(
                                info_md,
                                elem_classes="ffasr-job-info",
                            )
                            check_btn = gr.Button("Check", size="sm")
                            if job["can_retry"]:
                                retry_btn = gr.Button(
                                    "Retry", size="sm", variant="secondary"
                                )
                            if job["can_remove"]:
                                remove_btn = gr.Button("Remove", size="sm", variant="stop")
                        check_btn.click(
                            fn=_mod_open_check,
                            inputs=[gr.State(jid), mod_secret, mod_unlocked],
                            outputs=[
                                check_panel,
                                retry_panel,
                                approve_panel,
                                mod_check_job_id,
                                check_header,
                                check_notes,
                                check_recipe,
                                check_reqs,
                                check_setup,
                                check_script,
                            ],
                        )
                        if job["can_retry"]:
                            retry_btn.click(
                                fn=_mod_open_retry,
                                inputs=[gr.State(jid), mod_secret, mod_unlocked],
                                outputs=[
                                    check_panel,
                                    retry_panel,
                                    approve_panel,
                                    mod_retry_job_id,
                                    retry_header,
                                ],
                            )
                        if job["can_remove"]:
                            remove_btn.click(
                                fn=_mod_remove_row,
                                inputs=[
                                    gr.State(jid),
                                    mod_secret,
                                    mod_unlocked,
                                    mod_refresh_tick,
                                ],
                                outputs=[mod_feedback, mod_refresh_tick, mod_progress],
                            )

                gr.Markdown("### Import result from bucket artifact")
                gr.Markdown(
                    "If a Hub Job finished but the leaderboard CSV was not updated, paste the artifact "
                    "file name from <code>results/remote_artifacts/</code> (e.g. <code>a1b2c3d4.json</code>) "
                    "to merge WER/RTFx into the CSV."
                )
                mod_artifact_ref = gr.Textbox(
                    label="Artifact JSON file name or bucket path",
                    placeholder="e.g. a1b2c3d4.json or results/remote_artifacts/a1b2c3d4.json",
                )
                mod_replace_existing = gr.Checkbox(
                    label="Replace existing leaderboard row for this model",
                    value=False,
                )
                mod_import_notes = gr.Textbox(
                    label="Optional submission notes override",
                    placeholder="Leave empty to use notes from the matched queue job, if any",
                    lines=1,
                    max_length=4000,
                )
                import_artifact_btn = gr.Button("Import artifact to CSV", variant="secondary")

            def _mod_unlock(secret: str):
                ok, msg = job_queue.verify_moderator_secret(secret)
                if not ok:
                    return (
                        False,
                        gr.update(visible=False),
                        f"<p style='color:red'>{msg}</p>",
                        0,
                    )
                return (
                    True,
                    gr.update(visible=True),
                    "<p style='color:green'>Moderator tools unlocked for this session.</p>",
                    1,
                )

            def _mod_progress_gated(unlocked: bool):
                if not unlocked:
                    return job_queue.moderation_locked_placeholder_html()
                return job_queue.progress_html()

            mod_unlock_btn.click(
                fn=_mod_unlock,
                inputs=[mod_secret],
                outputs=[mod_unlocked, mod_panel, mod_unlock_feedback, mod_refresh_tick],
            )
            mod_refresh.click(
                fn=_mod_refresh_lists,
                inputs=[mod_secret, mod_unlocked, mod_refresh_tick],
                outputs=[mod_refresh_tick, mod_progress],
            )
            check_save_btn.click(
                fn=_mod_save_check,
                inputs=[
                    mod_check_job_id,
                    mod_secret,
                    mod_unlocked,
                    mod_refresh_tick,
                    check_recipe,
                    check_reqs,
                    check_setup,
                    check_script,
                ],
                outputs=[
                    mod_feedback,
                    mod_refresh_tick,
                    check_panel,
                    retry_panel,
                    approve_panel,
                ],
            )
            check_cancel_btn.click(
                fn=_mod_cancel_popup,
                inputs=[mod_refresh_tick],
                outputs=[mod_refresh_tick, check_panel, retry_panel, approve_panel],
            )
            retry_confirm_btn.click(
                fn=_mod_confirm_retry,
                inputs=[
                    mod_retry_job_id,
                    mod_secret,
                    mod_unlocked,
                    mod_refresh_tick,
                    retry_eval_conditions,
                ],
                outputs=[
                    mod_feedback,
                    mod_refresh_tick,
                    check_panel,
                    retry_panel,
                    approve_panel,
                    mod_progress,
                ],
            )
            retry_cancel_btn.click(
                fn=_mod_cancel_popup,
                inputs=[mod_refresh_tick],
                outputs=[mod_refresh_tick, check_panel, retry_panel, approve_panel],
            )
            approve_confirm_btn.click(
                fn=_mod_confirm_approve,
                inputs=[
                    mod_approve_job_id,
                    mod_secret,
                    mod_unlocked,
                    mod_refresh_tick,
                    approve_run_custom,
                    approve_eval_conditions,
                ],
                outputs=[
                    mod_feedback,
                    mod_refresh_tick,
                    check_panel,
                    retry_panel,
                    approve_panel,
                    mod_progress,
                ],
            )
            approve_cancel_btn.click(
                fn=_mod_cancel_popup,
                inputs=[mod_refresh_tick],
                outputs=[mod_refresh_tick, check_panel, retry_panel, approve_panel],
            )
            retry_all_btn.click(
                fn=_mod_retry_all,
                inputs=[mod_secret, mod_unlocked, mod_refresh_tick],
                outputs=[mod_feedback, mod_refresh_tick, mod_progress],
            )
            import_artifact_btn.click(
                fn=_mod_import_artifact,
                inputs=[
                    mod_secret,
                    mod_artifact_ref,
                    mod_replace_existing,
                    mod_import_notes,
                    mod_unlocked,
                    mod_refresh_tick,
                ],
                outputs=[mod_feedback, mod_refresh_tick],
            )

        with gr.TabItem(tab_label("analysis", "Analysis"), elem_id="analysis-tab", id=3) as analysis_tab:
            gr.Markdown("## Scenario analysis")
            gr.Markdown(
                "Charts support standard Plotly interaction (legend, zoom, pan, hover).\n\n"
                "**Leaderboard order** is by **Average WER** (mean across checked scenario columns; lower is better).\n\n"
                "**Average WER** ranks the top N models by mean WER across live scenarios (lower is better). "
                "**Speed** ranks the top N models by RTFx (audio sec / inference sec, higher is better). "
                "Bars are colored per company (e.g. all NVIDIA models share one color).\n\n"
                "**WER heatmap** shows WER (%) by model and scenario for the top N models (sorted by Average WER); "
                "lower WER corresponds to greener cells.\n\n"
                "**Pareto Front** plots Average WER on X and RTFx on Y (log scale). Models on the frontier are "
                "labeled and connected by the dashed blue line; other models render as faint dots.\n\n"
                "**WER by scenario** compares raw WER across conditions for the selected top models."
            )
            # with gr.Row():
            #     an_models = gr.Dropdown(
            #         label="Models on radar (up to 8)",
            #         choices=[],
            #         value=None,
            #         multiselect=True,
            #         max_choices=8,
            #         scale=2,
            #     )
            an_topn = gr.Slider(
                    5,
                    80,
                    value=10,
                    step=1,
                    label="Top‑N models (WER bars, heatmap & grouped WER)",
                    scale=1,
                )
            an_line_metrics = gr.CheckboxGroup(
                label="Scenarios shown in the heatmap",
                choices=_METRIC_CHOICES,
                value=[k for k in HEATMAP_SCENARIO_KEYS if k in LIVE_SCENARIO_KEYS],
            )
            an_apply = gr.Button("Apply / refresh charts", variant="primary")
            gr.Markdown(
                "### Pareto Front: Average WER vs RTFx\n\n"
                "Models on the Pareto frontier achieve the best trade-off between WER "
                "and speed (RTFx). Names are shown for frontier models; hover over "
                "other points to see their names."
            )
            an_plot_pareto = gr.Plot(
                label="Pareto Front: Average WER vs RTFx",
                elem_id="analysis-pareto-plot",
            )
            with gr.Row():
                an_plot_intelligence = gr.Plot(label="Average WER")
                an_plot_speed = gr.Plot(label="Speed")
            an_plot_compare = gr.Plot(label="WER heatmap")
            # with gr.Row():
            #     an_plot_radar = gr.Plot(label="Robustness radar")
            an_plot_bar = gr.Plot(label="WER by scenario")

            an_apply.click(
                fn=_analytics_apply,
                inputs=[
                    an_line_metrics,
                    an_topn,
                ],
                outputs=[
                    an_plot_intelligence,
                    an_plot_speed,
                    an_plot_compare,
                    an_plot_pareto,
                    an_plot_bar,
                ],
            )

            _analytics_outputs = [
                an_line_metrics,
                an_plot_intelligence,
                an_plot_speed,
                an_plot_compare,
                an_plot_pareto,
                an_plot_bar,
            ]
            _analytics_loaded = gr.State(False)

            def _analytics_on_tab_select(loaded: bool):
                """Load charts once the Analysis tab is visible so Plotly gets the real width."""
                if loaded:
                    return (loaded,) + (gr.skip(),) * 6
                (
                    metrics_upd,
                    fig_intel,
                    fig_speed,
                    fig_hm,
                    fig_pareto,
                    fig_b,
                ) = _analytics_initial()
                return (
                    True,
                    metrics_upd,
                    fig_intel,
                    fig_speed,
                    fig_hm,
                    fig_pareto,
                    fig_b,
                )

            analysis_tab.select(
                fn=_analytics_on_tab_select,
                inputs=[_analytics_loaded],
                outputs=[_analytics_loaded, *_analytics_outputs],
            )

        with gr.TabItem(tab_label("examples", "Examples"), elem_id="examples-tab", id=4):
            examples_renderer.ensure_assets()
            _dry = examples_renderer.sample_paths("dry")
            _hi = examples_renderer.sample_paths("high_snr")
            _mid = examples_renderer.sample_paths("mid_snr")
            _lo = examples_renderer.sample_paths("low_snr")
            _scene = examples_renderer.scene_image_path()

            def _example_audio(path: str | None, label: str):
                return gr.Audio(
                    value=path,
                    type="filepath",
                    label=label,
                    interactive=False,
                    editable=False,
                    show_download_button=False,
                    show_share_button=False,
                    waveform_options=gr.WaveformOptions(
                        waveform_color="#5c6670",
                        waveform_progress_color=TREBLE_TEAL,
                        show_recording_waveform=True,
                    ),
                    elem_classes="examples-audio",
                )

            with gr.Column(elem_classes="examples-scene-wrap"):
                gr.Image(
                    value=_scene,
                    type="filepath",
                    label="Treble scene",
                    interactive=False,
                    show_label=False,
                    show_download_button=False,
                    show_fullscreen_button=False,
                    height=300,
                    container=False,
                )

            gr.Markdown("### Near Field Speech")
            with gr.Row():
                _example_audio(_dry[0], "Example")

            gr.Markdown("### High SNR")
            with gr.Row():
                _example_audio(_hi[0], "Example")

            gr.Markdown("### Mid SNR")
            with gr.Row():
                _example_audio(_mid[0], "Example")

            gr.Markdown("### Low SNR")
            with gr.Row():
                _example_audio(_lo[0], "Example")

        with gr.TabItem(tab_label("about", "About"), elem_id="about-tab", id=5):
            gr.Markdown(ABOUT_TEXT, elem_classes="markdown-text")

    _footer_html = (
        "<div class='ffasr-footnote'>"
        "Leaderboard data loads from storage after the page opens · "
        "Evaluation runs on Hub Jobs against a held-out set · "
        "<a href='https://huggingface.co/' target='_blank' rel='noopener' "
        "style='color: inherit; text-decoration: underline; text-decoration-style: dotted;'>"
        "Models loaded from the Hugging Face Hub</a>"
        "</div>"
    )
    gr.HTML(_footer_html)

    with gr.Row():
        with gr.Accordion("Citation", open=False):
            gr.Textbox(
                value=CITATION_TEXT, lines=7,
                label="Copy the BibTeX snippet to cite this source",
                elem_id="citation-button",
                show_label=True,
            )

    # Periodic queue refresh (Timer added in Gradio 4.44). Avoid demo.load(..., every=); unsupported on some builds.
    if hasattr(gr, "Timer"):
        with gr.Row(visible=False):
            _next_up_timer = gr.Timer(15)
            _progress_timer = gr.Timer(2)
        _next_up_timer.tick(fn=job_queue.next_up_html, outputs=[next_up_panel])
        _progress_timer.tick(fn=_mod_progress_gated, inputs=[mod_unlocked], outputs=[mod_progress])

    demo.load(fn=_on_startup, outputs=[leaderboard_table, next_up_panel, version_dropdown])

def _launch_gradio_demo() -> None:
    """
    Avoid OSError: Cannot find empty port in range: 7860-7860.

    Hugging Face / tooling sometimes sets ``GRADIO_NUM_PORTS=1``, so only 7860 is tried.
    After a Space restart (e.g. post pip install), that port can still be in use briefly.
    We drop an overly small ``GRADIO_NUM_PORTS`` so Gradio's default port sweep applies, then
    retry with short backoff.
    """
    raw_np = os.environ.get("GRADIO_NUM_PORTS", "").strip()
    if raw_np.isdigit() and int(raw_np) < 10:
        os.environ.pop("GRADIO_NUM_PORTS", None)
    server_name = os.environ.get("GRADIO_SERVER_NAME", "0.0.0.0")
    last_err: OSError | None = None
    for attempt in range(24):
        try:
            demo.launch(server_name=server_name, server_port=None)
            return
        except OSError as e:
            last_err = e
            msg = str(e).lower()
            if "cannot find empty port" not in msg and "address already in use" not in msg:
                raise
            time.sleep(0.45 + attempt * 0.12)
    if last_err is not None:
        raise last_err


_launch_gradio_demo()