The problem worth investigating
Static difficulty bores advanced learners and overwhelms beginners.
This innovation is treated as a falsifiable engineering hypothesis. The goal is not to prove that an advanced technique is impressive; it is to determine whether it produces a measurable improvement over a simpler control.
Experimental architecture
- 01Mastery state
Instrumented independently so the experiment can reveal which stage contributes value. - 02Challenge metadata
Instrumented independently so the experiment can reveal which stage contributes value. - 03Difficulty policy
Instrumented independently so the experiment can reveal which stage contributes value. - 04Reward model
Instrumented independently so the experiment can reveal which stage contributes value. - 05Review scheduler
Instrumented independently so the experiment can reveal which stage contributes value.
Algorithms under study
Elo-inspired updates
Hypothesis. Maintains a compact estimate of learner skill and challenge difficulty that updates after each attempt.
Risk. Assumes a simplified relationship between ability and item difficulty.
Evidence. Prediction calibration, ranking accuracy and learning progression.
Bandit selection
Hypothesis. Uses feedback to allocate more traffic to promising choices while reserving some exploration for alternatives.
Risk. Biased or sparse rewards can push the policy toward a locally attractive but globally poor choice.
Evidence. Reward lift, regret, exploration coverage and stability over time.
Spaced repetition
Hypothesis. Schedules review around memory strength so practice is concentrated where forgetting risk is highest.
Risk. A poor mastery estimate can schedule reviews too aggressively or too late.
Evidence. Recall after delay, review efficiency and mastery calibration.
Mastery thresholds
Hypothesis. Mastery thresholds is included because it addresses a specific measurable part of the system rather than being added as decoration.
Risk. The component must be compared with a simpler baseline and removed if it adds complexity without measurable value.
Evidence. Task-specific quality metric, latency, reliability and failure-case analysis.
What data the experiment needs
- Answers — recorded with enough context to reproduce and audit the result.
- Time — recorded with enough context to reproduce and audit the result.
- Hints — recorded with enough context to reproduce and audit the result.
- Retries — recorded with enough context to reproduce and audit the result.
- Topic — recorded with enough context to reproduce and audit the result.
How CortexLab decides whether the idea survives
- Learning gain
- Drop-off
- Mastery calibration
Results should be compared against a control, segmented for failure cases and repeated across enough observations to avoid promoting noise into product behavior.
What can go wrong
Wrong reward
Optimization can improve the metric while making the actual experience worse.
Overconfidence
Small or biased samples can make experimental gains look more certain than they are.
Distribution shift
A policy that works on past users may degrade as topics, traffic and behavior change.
Next research milestones
- 01Personalized pacing
- 02Cross-domain transfer
- 03Longitudinal learning models
Questions still open
- Which simpler baseline must this beat before deployment?
- How should uncertainty be calibrated and communicated?
- What evidence would make us reject the idea?
- How do we prevent reward hacking or accidental optimization of engagement alone?
- Which decisions must remain human-reviewed?