1
Rulebook
Description and objective, team composition, format and progression, run structure and timings, permitted and prohibited actions, penalties, appeals, and definitions.
Events / The Black Box
The Doon School · 08–10 December 2026
The Black Box
You write down where you think the model will fail before you train it. Accuracy alone is not enough: the marks are in the diagnosis.
AI Skills
The dataset is broken on purpose. Find out how.
Failure modes in play
Phase 4 gives you ten targeted photographs. Spend them on the top two failure modes.
Rules
How it runs
- Audit, debug and fix a severely flawed image dataset so a computer vision model can accurately classify unseen test images. Two participants per school (one is permitted, at their own risk), open to students in any grade; no substitutions.
- A 10-minute setup window comes first: power on, verify internet connectivity, open the browser and check the camera or webcam. The sealed dataset envelope or folder may be opened only on the lead judge’s start signal.
- Phase 1, Autopsy (45 minutes): an online audit of a dataset filled with missing samples, bad lighting and heavy background bias, and a written failure prediction of where the model will break, and why.
- Phase 2, Baseline (30 minutes): upload the original dataset to The Black Box AI Deployed Model, train a baseline model, and log its metrics, baseline accuracy and confidence scores.
- Phase 3, Hunt (45 minutes): test the model on 30 unseen images from the judges, fill in the confusion matrix, tag each error by cause (colour, background, angle and so on), and identify the top two failure modes.
- Phase 4, Override (45 minutes): shoot up to 10 new targeted photos to fix those flaws, retrain the model and re-test it, logging the hypothesis and the before/after comparison.
- Phase 5, Presentation (10 minutes per team): a 6-minute presentation of the audit, predictions, confusion matrix findings, override strategy and before/after metrics, followed by a 4-minute Q&A with the judges. Slides: PDF or PPTX, at most 5.
- Deliverables must be submitted at each phase deadline to remain eligible for later phases. Custom code, image editing software, AI assistants and search engines are prohibited, internet reaches only the official submission portals, and the original training images may not be deleted or altered.
Error categories used in Phase 3
- Colour
- Background
- Angle
- Other causes the team identifies
Event details
- Discipline
- Machine learning
- Participants
- 2 per school
- Control
- Five phases
- Event head
- Abin Mukherjee
- Platform
- The Black Box AI Deployed Model, in the browser with no login. Chrome, Edge or Firefox only. No programming languages.
Judging
- Scored out of 100 across all five phases
- Penalties: 2 points per minute late (up to 5 minutes per phase), 5 points per photo over the 10-photo limit, and 10 points per edit to the original training set
- Ties: highest accuracy on the 30 test images after Phase 4, then the Phase 1 score, then the Phase 3 score, then the lowest total time to submit deliverables across Phases 1 to 4
Judging rubric
How the marks fall
| Criterion | Weight | Marks |
|---|---|---|
| Phase 1: Autopsy & Failure Prediction | 20 | |
| Phase 2: Baseline Model Execution | 10 | |
| Phase 3: Failure Hunt & Confusion Matrix | 25 | |
| Phase 4: Override Strategy & Post-Patch Accuracy | 30 | |
| Phase 5: Presentation & Q&A Defense | 15 | |
| Total | 100 |
Weights are those in the event’s published documents. Scoring levels, tie-break order and who scores what are set out in full in the document pack below.
Documents
The The Black Box pack
4 of 5 documents available
These are the documents your team will receive for The Black Box. Between them they set out everything a team needs in order to prepare.
2
Platform & Environment Specification
Languages and libraries permitted, devices brought versus provided, internet access, policy on AI tools, provided datasets and platforms, submission format.
4
Judging Rubric
Criteria with point values summing to a stated total, scoring levels, tie-break order, who scores what, and when scores are published.
5
Inspection & Safety
Pre-competition inspection checklist and schedule, consequences of failed inspection, safety requirements during runs, spectator boundaries, and the emergency stop procedure.
6
Support Material
A worked example, a starter guide to core techniques, the concepts a team should learn, common mistakes, and an FAQ.
Documents marked “Available soon” will appear here as soon as they are released.
Flow
How The Black Box works, step by step
The event from start to finish, as the rules above set it out. The detailed documents for The Black Box will appear here once they are released to registered schools.
-
10-minute setup; the dataset is opened on the start signal
-
Phase 1: Autopsy (45 minutes)
Audit the flawed dataset online and write failure predictions.
-
Phase 2: Baseline (30 minutes)
Upload the dataset, train the baseline model, log its metrics.
-
Phase 3: Hunt (45 minutes)
Test 30 unseen images, fill the confusion matrix, tag every error, name the top two failure modes.
-
Phase 4: Override (45 minutes)
Up to 10 targeted photos, retrain, re-test.
-
Phase 5: Presentation (10 minutes)
Six minutes to the judges, then four minutes of questions.
-
Scored out of 100 across all five phases
Who runs it
The people behind The Black Box
Abin Mukherjee
Event Head
Abin is an 11th-grade student pursuing the IB Diploma Programme, with a keen interest in Art and Economics. He serves as the Event Head of The Black Box and Head of Design for the IPSC IT Fest. He is also an editor and the Chief of Design for The Infinity, the School’s premier mathematics publication, and a Senior Designer for The Echo, the School’s science magazine. A senior member of the School Orchestra, he plays the violin. He is honoured to serve as Event Head of The Black Box and Head of Design for the IPSC IT Fest 2026.
Next
Terra Incognita
Robots map an unseen arena cell by cell, push a block into the target green spot, and finish as close as they can to where they started, fully autonomously.