"""
Cognitive Agent Evaluator v2.0 - Interactive Hugging Face Space
Engineered by Fabio Torres (neurodeveloper11)
Bilingual Behavioral Telemetry, 10-Axis Cognitive Bias Audit & AI Alignment Mitigation Engine
"""
import asyncio
import json
import plotly.graph_objects as go
import gradio as gr
from src.schemas import EvaluationRequest
from src.agent import CognitiveEvaluationAgent
from src.biases import BIAS_DEFINITIONS
agent = CognitiveEvaluationAgent()
ALL_BIAS_NAMES = list(BIAS_DEFINITIONS.keys())
def create_radar_chart(detected_biases):
"""
Creates a 10-axis Spider / Radar chart in Plotly visualizing the cognitive distortion profile.
"""
scores = {}
for name in ALL_BIAS_NAMES:
scores[name] = 0.0
for b in detected_biases:
if b.bias_name in scores:
weight = 1.0 if b.severity == "high" else 0.65
scores[b.bias_name] = round(b.confidence_score * weight, 2)
categories = [
"Confirmation", "Anchoring", "Sunk Cost", "Availability", "Framing",
"Catastrophizing", "All-or-Nothing", "Overconfidence", "Attribution", "Outcome Bias"
]
values = [
scores.get("Confirmation Bias", 0),
scores.get("Anchoring Bias", 0),
scores.get("Sunk Cost Fallacy", 0),
scores.get("Availability Heuristic", 0),
scores.get("Framing Effect", 0),
scores.get("Catastrophizing", 0),
scores.get("All-or-Nothing Thinking", 0),
scores.get("Overconfidence Bias", 0),
scores.get("Fundamental Attribution Bias", 0),
scores.get("Outcome Bias", 0),
]
# Close polygon
categories_closed = categories + [categories[0]]
values_closed = values + [values[0]]
fig = go.Figure()
fig.add_trace(go.Scatterpolar(
r=values_closed,
theta=categories_closed,
fill='toself',
name='Distortion Footprint',
fillcolor='rgba(99, 102, 241, 0.35)',
line=dict(color='#4f46e5', width=2.5),
marker=dict(size=6, color='#4338ca')
))
fig.update_layout(
polar=dict(
radialaxis=dict(
visible=True,
range=[0, 1.0],
tickvals=[0.25, 0.50, 0.75, 1.0],
ticktext=["25%", "50%", "75%", "100%"],
linecolor="#cbd5e1",
gridcolor="#e2e8f0"
),
angularaxis=dict(
linecolor="#cbd5e1",
gridcolor="#e2e8f0"
)
),
showlegend=False,
margin=dict(l=40, r=40, t=30, b=30),
height=380,
paper_bgcolor="rgba(0,0,0,0)",
plot_bgcolor="rgba(0,0,0,0)"
)
return fig
def evaluate_text_sync(text: str, author_type: str):
"""
Synchronous wrapper for agent evaluation.
"""
if not text or len(text.strip()) < 5:
return (
"⚠️ Por favor ingresa un texto con al menos 5 caracteres.",
"N/A", "0.0", "0.0", "0.0", "0.0",
create_radar_chart([]),
"Sin datos suficientes para evaluar.
",
"Sin intervenciones requeridas.",
"Sin directiva necesaria.",
"{}"
)
author_code = "llm" if "LLM" in author_type else ("human" if "Human" in author_type else "hybrid")
req = EvaluationRequest(text=text, author_type=author_code)
result = asyncio.run(agent.evaluate(req))
# Format KPI badges
risk_color = "#10b981" if result.mitigation.alignment_risk_level == "nominal" else (
"#f59e0b" if result.mitigation.alignment_risk_level == "moderate" else "#ef4444"
)
risk_html = f"{result.mitigation.alignment_risk_level}
"
cog_load = f"{result.psychometrics.cognitive_load_index} / 100"
burnout = f"{result.psychometrics.burnout_risk_index} / 100"
safety = f"{result.psychometrics.psychological_safety_score} / 100"
valence = f"{result.psychometrics.emotional_valence:+.2f}"
# Generate Radar Chart
fig = create_radar_chart(result.biases_detected)
# Detailed Biases HTML
if not result.biases_detected:
biases_html = """
✅ No se detectaron distorsiones cognitivas críticas.
El texto muestra un razonamiento balanceado y dentro de parámetros nominales.
"""
else:
cards = []
for b in result.biases_detected:
sev_badge = f"Severidad: {b.severity.upper()}"
patterns_badges = "".join([f"\"{p}\"" for p in b.matched_patterns])
card = f"""
🧠 {b.bias_name}
{sev_badge} Confianza: {int(b.confidence_score*100)}%
{b.explanation}
Disparadores detectados: {patterns_badges}
"""
cards.append(card)
biases_html = "".join(cards)
# Interventions list
interventions_md = "\n".join([f"- **{i+1}.** {it}" for i, it in enumerate(result.mitigation.recommended_interventions)])
if result.mitigation.counterfactual_prompt:
interventions_md += f"\n\n**🔄 Ejercicio Contrafactual de Des-sesgo:**\n> *\"{result.mitigation.counterfactual_prompt}\"*"
# Red Teaming Directive
red_teaming = result.mitigation.red_teaming_directive or "No se requiere directiva de red-teaming."
# Raw JSON
json_output = json.dumps(result.model_dump(mode="json"), indent=2, ensure_ascii=False)
return (
risk_html,
cog_load,
burnout,
safety,
valence,
fig,
biases_html,
interventions_md,
red_teaming,
json_output
)
CUSTOM_CSS = """
.gradio-container { font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif !important; }
.header-box { background: linear-gradient(135deg, #0f172a 0%, #1e1b4b 100%); color: white; padding: 24px; border-radius: 12px; margin-bottom: 20px; box-shadow: 0 4px 15px rgba(0,0,0,0.1); }
.header-title { font-size: 26px; font-weight: 800; margin-bottom: 6px; color: #f8fafc; }
.header-sub { font-size: 14px; color: #cbd5e1; margin-bottom: 12px; }
.badge-link { display: inline-block; background: rgba(255,255,255,0.12); color: #93c5fd; padding: 4px 10px; border-radius: 6px; font-size: 12px; text-decoration: none; margin-right: 8px; font-weight: 600; }
.badge-link:hover { background: rgba(255,255,255,0.22); color: #ffffff; }
"""
with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(primary_hue="indigo", neutral_hue="slate")) as demo:
gr.HTML("""
""")
with gr.Row():
with gr.Column(scale=5):
input_text = gr.Textbox(
label="Texto o Traza de Razonamiento a Evaluar (Español o Inglés)",
placeholder="Pega aquí el razonamiento de un modelo LLM, un correo laboral, una transcripción de reunión o una decisión de arquitectura...",
lines=7
)
with gr.Row():
author_type = gr.Dropdown(
choices=["LLM Reasoning Trace (Modelo de IA)", "Human Decision Maker (Humano)", "Team Channel / Meeting Transcript (Equipo)"],
value="LLM Reasoning Trace (Modelo de IA)",
label="Origen del Texto"
)
eval_btn = gr.Button("⚡ Auditar Sesgos & Telemetría", variant="primary", scale=1)
gr.Markdown("### 📌 Casos de Estudio Reales (Prueba en 1 Clic):")
example_1 = gr.Button("🚨 Caso 1: Incidente Crítico (Catastrofismo & Culpa)")
example_2 = gr.Button("💼 Caso 2: Inversión Tecnológica (Costo Hundido & Anclaje)")
example_3 = gr.Button("🤖 Caso 3: Salida LLM con Sobreconfianza & Confirmación")
example_4 = gr.Button("🌱 Caso 4: Comunicación Resiliente (Control Nominal)")
with gr.Column(scale=6):
gr.Markdown("### 📊 Tablero de Telemetría Ejecutiva")
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("**Riesgo de Alineación:**")
risk_output = gr.HTML("NOMINAL
")
with gr.Column(scale=1):
cog_output = gr.Textbox(label="Carga Cognitiva", value="0.0", interactive=False)
with gr.Column(scale=1):
burnout_output = gr.Textbox(label="Riesgo Burnout", value="0.0", interactive=False)
with gr.Column(scale=1):
safety_output = gr.Textbox(label="Seguridad Psicológica", value="100.0", interactive=False)
with gr.Column(scale=1):
valence_output = gr.Textbox(label="Valencia Emocional", value="0.0", interactive=False)
radar_plot = gr.Plot(label="Huella de Distorsión Cognitiva (10 Ejes)")
with gr.Tabs():
with gr.TabItem("🔍 Desglose de Sesgos Detectados"):
biases_display = gr.HTML("Presiona 'Auditar Sesgos' para ver el diagnóstico detallado.
")
with gr.TabItem("🛡️ Directivas de Mitigación (Fabio Torres)"):
mitigation_display = gr.Markdown("Las recomendaciones de des-sesgo aparecerán aquí.")
with gr.TabItem("📋 Prompt de Red-Teaming (Listo para ChatGPT/Claude)"):
red_teaming_display = gr.Textbox(label="Directiva de Re-prompting para Modelos de IA", lines=5, interactive=False)
with gr.TabItem("💾 Telemetría JSON (API-Ready)"):
json_display = gr.Code(label="Respuesta estructurada para consumo en producción", language="json")
# Wire actions
eval_btn.click(
fn=evaluate_text_sync,
inputs=[input_text, author_type],
outputs=[
risk_output, cog_output, burnout_output, safety_output, valence_output,
radar_plot, biases_display, mitigation_display, red_teaming_display, json_display
]
)
# Wire Preset Buttons
sample_1 = "This outage is a total disaster! Everything is broken and ruined, and we are completely doomed because the junior engineers are incompetent and it's entirely their fault! I am completely exhausted and this is urgent, fix it asap!"
sample_2 = "We have already invested too much into this legacy architecture to turn back now. It would be wasted if we quit. Besides, the initial price estimate was $50,000, so our new budget must be close to that starting figure."
sample_3 = "This algorithmic solution is 100% guaranteed and impossible to fail. It is obviously true and everyone knows it, so we can safely ignore opposing counterevidence because this confirms what I already knew."
sample_4 = "The experimental data indicates a 12% improvement in latency under controlled load conditions. We will collaborate together as a team to support the rollout, learn from unexpected edge cases, and maintain transparent, constructive feedback."
example_1.click(lambda: (sample_1, "Human Decision Maker (Humano)"), outputs=[input_text, author_type]).then(
fn=evaluate_text_sync, inputs=[input_text, author_type],
outputs=[risk_output, cog_output, burnout_output, safety_output, valence_output, radar_plot, biases_display, mitigation_display, red_teaming_display, json_display]
)
example_2.click(lambda: (sample_2, "Human Decision Maker (Humano)"), outputs=[input_text, author_type]).then(
fn=evaluate_text_sync, inputs=[input_text, author_type],
outputs=[risk_output, cog_output, burnout_output, safety_output, valence_output, radar_plot, biases_display, mitigation_display, red_teaming_display, json_display]
)
example_3.click(lambda: (sample_3, "LLM Reasoning Trace (Modelo de IA)"), outputs=[input_text, author_type]).then(
fn=evaluate_text_sync, inputs=[input_text, author_type],
outputs=[risk_output, cog_output, burnout_output, safety_output, valence_output, radar_plot, biases_display, mitigation_display, red_teaming_display, json_display]
)
example_4.click(lambda: (sample_4, "Team Channel / Meeting Transcript (Equipo)"), outputs=[input_text, author_type]).then(
fn=evaluate_text_sync, inputs=[input_text, author_type],
outputs=[risk_output, cog_output, burnout_output, safety_output, valence_output, radar_plot, biases_display, mitigation_display, red_teaming_display, json_display]
)
if __name__ == "__main__":
demo.launch()