app.py

import gradio as gr
import pandas as pd
import os
from datetime import date
from huggingface_hub import HfApi

CSV_PATH = "observations.csv"
REPO_ID = "Flame-Forged/coherence-gap-explorer"
HF_TOKEN = os.environ.get("HF_TOKEN")

custom_css = """
body, .gradio-container {
    background-color: #0d0d1a !important;
    color: #e0e0f0 !important;
    font-family: 'Georgia', serif !important;
}
.gradio-container h1 {
    font-size: 2.2em !important;
    font-weight: 900 !important;
    background: linear-gradient(90deg, #a855f7, #f59e0b) !important;
    -webkit-background-clip: text !important;
    -webkit-text-fill-color: transparent !important;
    padding-bottom: 6px !important;
}
.gradio-container p, .gradio-container label {
    color: #c4b5fd !important;
}
.tab-nav {
    background: #1a1a2e !important;
    border-bottom: 2px solid #7c3aed !important;
}
.tab-nav button {
    color: #a0aec0 !important;
    font-weight: 600 !important;
    font-size: 1em !important;
    border-radius: 6px 6px 0 0 !important;
    padding: 10px 24px !important;
}
.tab-nav button.selected {
    background: #7c3aed !important;
    color: #ffffff !important;
    border-bottom: none !important;
}
input[type="text"], textarea, select, .gr-box {
    background-color: #1a1a2e !important;
    color: #e0e0f0 !important;
    border: 1px solid #4c1d95 !important;
    border-radius: 6px !important;
}
input[type="text"]:focus, textarea:focus {
    border-color: #a855f7 !important;
    outline: none !important;
    box-shadow: 0 0 0 2px rgba(168, 85, 247, 0.3) !important;
}
button.primary {
    background: linear-gradient(90deg, #7c3aed, #a855f7) !important;
    color: white !important;
    border: none !important;
    font-weight: 700 !important;
    font-size: 1em !important;
    padding: 10px 28px !important;
    border-radius: 8px !important;
    cursor: pointer !important;
    transition: opacity 0.2s !important;
}
button.primary:hover { opacity: 0.85 !important; }
table {
    background-color: #12122a !important;
    border-collapse: collapse !important;
    width: 100% !important;
}
th {
    background-color: #4c1d95 !important;
    color: #f59e0b !important;
    font-weight: 700 !important;
    text-transform: uppercase !important;
    font-size: 0.78em !important;
    letter-spacing: 0.08em !important;
    padding: 10px 14px !important;
    border-bottom: 2px solid #7c3aed !important;
}
td {
    background-color: #0d0d1a !important;
    color: #e0e0f0 !important;
    padding: 6px 14px !important;
    border-bottom: 1px solid #1e1e3a !important;
    font-size: 0.9em !important;
    overflow: hidden !important;
    text-overflow: ellipsis !important;
    white-space: nowrap !important;
    vertical-align: middle !important;
}
tr:hover td { background-color: #1a1a2e !important; }
.result-count p {
    color: #f59e0b !important;
    font-style: italic !important;
    font-size: 0.9em !important;
}
.table-wrap {
    max-height: 500px !important;
    overflow-y: auto !important;
}
::-webkit-scrollbar { width: 6px; height: 6px; }
::-webkit-scrollbar-track { background: #0d0d1a; }
::-webkit-scrollbar-thumb { background: #7c3aed; border-radius: 3px; }
"""

def load_data():
    return pd.read_csv(CSV_PATH, on_bad_lines='skip', engine='python')

def get_choices(df, column):
    if column not in df.columns:
        return ["All"]
    vals = sorted(df[column].dropna().unique().tolist())
    return ["All"] + [str(v) for v in vals]

def filter_data(platform, behavior_category, confidence, search_term):
    df = load_data()
    if platform != "All":
        df = df[df["platform"].astype(str) == platform]
    if behavior_category != "All":
        df = df[df["behavior_category"].astype(str) == behavior_category]
    if confidence != "All":
        df = df[df["interpretive_confidence"].astype(str) == confidence]
    if search_term.strip():
        mask = df.apply(
            lambda row: row.astype(str).str.contains(
                search_term.strip(), case=False, na=False
            ).any(),
            axis=1
        )
        df = df[mask]
    return df

def update(platform, behavior_category, confidence, search_term):
    df = filter_data(platform, behavior_category, confidence, search_term)
    count = f"*Showing {len(df)} of {len(load_data())} observations*"
    return df, count

def build_stats():
    df = load_data()
    total = len(df)

    def make_bar(count, total, width=20):
        filled = int(round(count / total * width)) if total > 0 else 0
        return "█" * filled + "░" * (width - filled)

    def section(title, series):
        lines = [f"### {title}\n"]
        for val, count in series.sort_values(ascending=False).items():
            bar = make_bar(count, total)
            pct = round(count / total * 100)
            lines.append(f"`{bar}` **{val}** — {count} ({pct}%)")
        return "\n\n".join(lines)

    platforms  = section("By Platform",                df["platform"].value_counts())
    categories = section("By Behavior Category",       df["behavior_category"].value_counts())
    confidence = section("By Interpretive Confidence", df["interpretive_confidence"].value_counts())
    repro      = section("By Reproducibility Status",  df["reproducibility_status"].value_counts())

    evidence_pct = round(df["raw_prompt_available"].astype(str).str.lower().eq("yes").mean() * 100)
    response_pct = round(df["raw_response_available"].astype(str).str.lower().eq("yes").mean() * 100)

    return f"""## Dataset Overview

| Metric | Value |
|---|---|
| Total observations | **{total}** |
| Platforms covered | **{df['platform'].nunique()}** |
| Behavior categories | **{df['behavior_category'].nunique()}** |
| Raw prompt available | **{evidence_pct}%** of observations |
| Raw response available | **{response_pct}%** of observations |

---
{platforms}

---
{categories}

---
{confidence}

---
{repro}
"""

def submit_observation(
    f_date, platform, model_version, memory_enabled,
    window_type, prompt_class, behavior_category,
    description, evidence, screenshot_ids,
    raw_prompt, raw_response, confidence,
    alternative_explanations, reproducibility_status, coder_notes
):
    new_row = {
        "date": f_date, "platform": platform,
        "model_version": model_version, "memory_enabled": memory_enabled,
        "window_type": window_type, "prompt_class": prompt_class,
        "behavior_category": behavior_category, "description": description,
        "evidence": evidence, "screenshot_ids": screenshot_ids,
        "raw_prompt_available": raw_prompt, "raw_response_available": raw_response,
        "interpretive_confidence": confidence,
        "alternative_explanations": alternative_explanations,
        "reproducibility_status": reproducibility_status,
        "coder_notes": coder_notes
    }
    df = load_data()
    new_df = pd.concat([df, pd.DataFrame([new_row])], ignore_index=True)
    new_df.to_csv(CSV_PATH, index=False)
    if HF_TOKEN:
        api = HfApi(token=HF_TOKEN)
        api.upload_file(
            path_or_fileobj=CSV_PATH,
            path_in_repo="observations.csv",
            repo_id=REPO_ID,
            repo_type="space"
        )
        return "✅ Observation submitted and saved permanently!"
    else:
        return "⚠️ Saved locally but HF_TOKEN not set — won't persist after restart."

initial_df = load_data()

with gr.Blocks(title="Coherence Gap Explorer") as demo:
    gr.Markdown("# Coherence Gap Explorer")
    gr.Markdown(
        "Interactive browser for the observational dataset from the Coherence Gap paper."
    )

    with gr.Tab("Browse & Filter"):
        with gr.Row():
            platform_dd = gr.Dropdown(
                choices=get_choices(initial_df, "platform"),
                value="All", label="Platform"
            )
            category_dd = gr.Dropdown(
                choices=get_choices(initial_df, "behavior_category"),
                value="All", label="Behavior Category"
            )
            confidence_dd = gr.Dropdown(
                choices=get_choices(initial_df, "interpretive_confidence"),
                value="All", label="Interpretive Confidence"
            )
        search_box = gr.Textbox(
            label="Search",
            placeholder="Search across all columns...",
            lines=1
        )
        result_count = gr.Markdown(
            f"*Showing {len(initial_df)} of {len(initial_df)} observations*",
            elem_classes=["result-count"]
        )
        table = gr.Dataframe(
            value=initial_df,
            label="Observations",
            interactive=False,
            wrap=False,
            elem_classes=["table-wrap"]
        )
        inputs = [platform_dd, category_dd, confidence_dd, search_box]
        for component in inputs:
            component.change(fn=update, inputs=inputs, outputs=[table, result_count])

    with gr.Tab("Stats"):
        refresh_btn = gr.Button("Refresh Stats", variant="primary")
        stats_display = gr.Markdown(value=build_stats())
        refresh_btn.click(fn=build_stats, inputs=[], outputs=stats_display)

    with gr.Tab("Submit Observation"):
        gr.Markdown("### Add a New Observation")
        with gr.Row():
            f_date = gr.Textbox(label="Date", value=str(date.today()), placeholder="YYYY-MM-DD")
            f_platform = gr.Textbox(label="Platform")
            f_model_version = gr.Textbox(label="Model Version")
        with gr.Row():
            f_memory = gr.Dropdown(choices=["Yes", "No", "Unknown"], label="Memory Enabled")
            f_window = gr.Textbox(label="Window Type")
            f_prompt_class = gr.Textbox(label="Prompt Class")
            f_category = gr.Textbox(label="Behavior Category")
        with gr.Row():
            f_raw_prompt = gr.Dropdown(choices=["Yes", "No", "Unknown"], label="Raw Prompt Available")
            f_raw_response = gr.Dropdown(choices=["Yes", "No", "Unknown"], label="Raw Response Available")
            f_confidence = gr.Dropdown(choices=["Low", "Medium", "High"], label="Interpretive Confidence")
            f_repro = gr.Textbox(label="Reproducibility Status")
        f_description = gr.Textbox(label="Description", lines=3)
        f_evidence = gr.Textbox(label="Evidence", lines=3)
        f_alternative = gr.Textbox(label="Alternative Explanations", lines=2)
        f_screenshots = gr.Textbox(label="Screenshot IDs")
        f_notes = gr.Textbox(label="Coder Notes", lines=2)
        submit_btn = gr.Button("Submit Observation", variant="primary")
        status = gr.Markdown("")
        submit_btn.click(
            fn=submit_observation,
            inputs=[
                f_date, f_platform, f_model_version, f_memory,
                f_window, f_prompt_class, f_category,
                f_description, f_evidence, f_screenshots,
                f_raw_prompt, f_raw_response, f_confidence,
                f_alternative, f_repro, f_notes
            ],
            outputs=status
        )

demo.launch(css=custom_css)