Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/elenacarino-max/Pildora4_ext_StarWars/llms.txt

Use this file to discover all available pages before exploring further.

Once a team completes all three missions, the activity enters its final and most pedagogically rich phase. Instead of writing more code, students are now asked to act as data scientists who must justify a model choice to a stakeholder audience. This phase has two parts: a comparator that surfaces the tracked metrics and artifacts for all three runs side by side, and a defense submission that requires students to document their reasoning in a structured, evidence-grounded form. Completing the defense seals the team’s record and awards the final +25 points.

The comparator

The comparator view presents all three mission runs — Mision_Tatooine, Mision_Coruscant, and Mision_Mustafar — alongside each other so teams can identify differences and trends before committing to a recommendation. For each run the comparator shows: Tracked metrics:
MetricDescription
accuracyOverall fraction of correctly classified digits
precision_weightedClass-weighted precision across all 10 digits
recall_weightedClass-weighted recall across all 10 digits
f1_weightedHarmonic mean of precision and recall, class-weighted
recall_1Recall specifically for digit 1 (fraction of true 1s correctly identified)
recall_8Recall specifically for digit 8 (fraction of true 8s correctly identified)
Tracked artifacts:
  • matriz_confusion.png — a visual confusion matrix showing per-class prediction patterns
  • classification_report.json — the full per-class precision, recall, and F1 breakdown as structured JSON
Both artifact types can be opened directly in the comparator view. The confusion matrix is especially useful for spotting which digits the model confuses with each other; the JSON report provides the precise numbers to quote in the defense text.

Orden 66 — Receiving the holocron

After the comparator, each team presses the Orden 66 button to reveal their unique holocron challenge card. Holocrons are assigned deterministically based on team join order within the session, cycling through all seven challenges so that different teams in the same classroom receive different analytical lenses. The instructor can reassign a holocron from the admin panel if needed. The holocron card states a specific task and a series of investigation steps that frame how the team should approach their defense. All seven challenges require evidence from the tracked runs, but each emphasises a different concern — from audit reproducibility to non-technical communication to per-class recall analysis.

The seven holocron challenges

◈ Consejo desconfiado

Focus: evidencias — Convence a un Consejo que no acepta una recomendación basada únicamente en accuracy. Requiere al menos dos métricas, un artefacto y una limitación explícita.

◒ Recursos limitados

Focus: complejidad — Decide si la mejora de Mustafar (300 árboles) compensa su mayor coste frente a runs con 50 ó 100 árboles. Evalúa rendimiento frente a número de estimadores.

⌁ Auditoría imperial

Focus: trazabilidad — Prepara una auditoría que permita a otro equipo identificar los cinco parámetros, cuatro métricas, dos artefactos y el modelo registrado, y que indique qué faltaría para reproducción perfecta.

◎ Senado galáctico

Focus: comunicación — Presenta la decisión al Senado sin jerga técnica. Traduce la métrica principal a lenguaje llano y usa la matriz de confusión para ilustrar un error concreto.

◐ El lado oscuro de la métrica

Focus: recall_8 — Investiga si una accuracy alta oculta fallos en el dígito 8. Compara recall_8 entre las tres runs y localiza en la matriz con qué otros dígitos se confunde el 8.

✦ Rescate prioritario

Focus: recall_1 — El rescate depende de detectar tantos dígitos 1 reales como sea posible. Prioriza recall_1 incluso si ello implica sacrificar accuracy global; justifica el riesgo aceptado.

⬡ Misión en producción

Focus: producción — Decide qué run está lista para producción. Combina rendimiento, coste (número de árboles), trazabilidad (parámetros y modelo registrado) y propone un control adicional antes del despliegue.

Submitting the defense

The defense form has seven required fields. Every field must be filled before the team can seal the defense; leaving any field blank will show a validation error. Once sealed, the defense is stored permanently against the team record and the +25 points are applied.
1

selected_run

The name of the run the team recommends (one of Mision_Tatooine, Mision_Coruscant, or Mision_Mustafar). This must be different from discarded_run.
2

evidence

The metric or metrics used to justify the selection — e.g. “Mustafar tiene el mayor f1_weighted (0.9692) y recall_1 (1.0)”. Teams should refer to specific numeric values from the comparator.
3

artifact

The artifact consulted during the analysis — either matriz_confusion.png or classification_report.json. Teams should note what specific information the artifact revealed.
4

discarded_run

The name of the run explicitly rejected. Must differ from selected_run.
5

discard_reason

An evidence-based explanation of why the discarded run was not chosen — e.g. “Tatooine tiene recall_1=0.9167, significativamente inferior a Mustafar”.
6

limitation

One honest acknowledged limitation of the chosen model or decision — e.g. “recall_8 permanece en 0.8571 en las tres runs; el modelo no mejora para el dígito 8 independientemente del número de árboles”.
7

recommendation

The final recommendation text as it would be delivered to a stakeholder — a concise, professional statement that integrates evidence, context from the holocron challenge, and the acknowledged limitation.
The selected_run and discarded_run fields must contain different run names. If both are set to the same value, the form will reject the submission with the message “La run elegida y la descartada deben ser distintas.”
Completing and sealing the defense awards +25 points — the largest single reward in the activity. Combined with a perfect run through all chambers (6 × 20 = 120) and all three missions (3 × 30 + 15 = 105), the maximum total score is 250 points.

Build docs developers (and LLMs) love