A.6.1 adds durable two-way offline synchronization and one-click assignment-to-guided-session practice while preserving Firebase role access and on-device MediaPipe/MudraNet inference.
Camera inference runs locally in this browser.
Session progress
Keep practicing for a clearer trend.
Choose a student
Prototype learning metrics are calculated from locally stored practice attempts.
Active assignments
Achievements
Validated-core mudras
Recurring corrections
Recent attempts
Skill weakness profile
Multi-student validation
Track practice coverage, review workload, AI/teacher agreement and the mudras that need the most instructional attention.
Mudra validation coverage
Correction heat map
Percentage of image-backed practice attempts where each coaching area needed correction.
Practice assignments
Reusable correction templates
Teacher wording can override the built-in level-aware coaching language when mudra, level, correction area and action match.
Students
AI attempt review
Review the captured student image alongside the AI result, confirm the mudra, and assess whether the corrections match your teaching judgment. Older attempts without a review image are excluded.
Attempt queue
Automatically ranked by Ask Teacher, unhelpful feedback, uncertainty, low score and repeated corrections.
Human-validated AI monitoring
These metrics use teacher-confirmed attempts. They are separate from the original offline model-validation results.
Teacher-confirmed confusion matrix
Confidence calibration
Correction review
Student feedback
Confidence calibration
Model A/B testing
Engineering health dashboard
Inspect dataset coverage, reference readiness, prediction gates and feature-level scoring without changing the trained model.
Dataset health by mudra
Latest inference / attempt
Feature-level scoring
28-hasta curriculum scaffold
Five mudras are AI-active. A.5 opens the next five for structured teacher data collection while recognition remains locked until a 10-class model passes held-out validation.
Next-five dataset readiness
Each new mudra targets 30 approved samples, 4 performers, both hands, and 10 labeled-error samples before 10-class training evaluation.
Pataka
One correction at a time
The next retry focuses on the most important correction first.
Before → After
Geometry checks
Your response helps the teacher identify corrections that need review.
Capture Pataka
Hold the hand steady, then save either an approved reference or an intentionally incorrect example.
Error labels are only required when saving an intentional mistake. They will become supervised correction labels later.