Review Process

Every exercise and edit passes through a 3-tier AI review engine before entering the library. The system balances speed with accuracy.

Review Tiers

Tier 1: Instant

Fast model (cheap, quick)< 2 seconds

Surface-level metadata validation. Checks for obvious errors, profanity, and formatting. Auto-applies approved changes immediately.

Triggers

Minor edits to descriptions, tags, experience level, instructions, safety notes

Checks performed

  • Field format and value validation
  • Tag canonicalization (standardized tag names)
  • Description quality check (length, coherence)
  • Safety note accuracy
  • No profanity or spam

Tier 2: Expert Review

Expert model (research-capable)3-8 seconds

Deep validation by an expert AI model with web research capabilities. Verifies anatomical accuracy, equipment requirements, and exercise safety. Includes duplicate detection.

Triggers

New exercise proposals, changes to muscles, equipment, biomechanics, MET values

Checks performed

  • Anatomical accuracy (primary/secondary muscles)
  • Equipment requirements match the exercise
  • Biomechanical plausibility
  • MET value accuracy (cross-referenced)
  • Duplicate detection against existing library
  • Citation verification when provided
  • Safety concern flagging

Tier 3: Crowdsourced

No AI — user feedbackOngoing

Exercises enter the library tagged as "pending_validation." Users see a subtle feedback card and can vote thumbs up/down. After sufficient positive feedback, status upgrades to "validated."

Triggers

Exercises approved by Tier 1/2 that need real-world validation

Checks performed

  • Users rate accuracy via thumbs up/down
  • Feedback shown inline in exercise viewer
  • Notification-style prompts for review
  • Confidence score adjusts based on feedback ratio
  • Stale exercises (6+ months inactive) re-enter validation

Review Decisions

approve

Changes are applied immediately. Provenance is updated with contributor info.

approve_with_modifications

AI modifies the submission (fixes formatting, corrects minor issues) and applies.

reject

Changes are not applied. Reasoning is returned to the submitter.

needs_review

Flagged for admin review. Exercise enters pending_review status.

Confidence Scoring

Every exercise carries a confidence score (0.0 – 1.0) computed from 7 factors. Higher confidence means more trust. Exercises below 0.3 are flagged for archive review.

FactorEffect
Source base scoreAI-generated: 0.5, User-submitted: 0.6, Migration: 0.4
Citations+0.05 per citation (max +0.2)
Review count+0.05 per AI review pass (max +0.2)
Admin confirmation+0.1 if admin-reviewed
Feedback ratioUp to +0.1 or -0.1 based on thumbs up/down ratio
Usage boost+0.05 if used by 10+ users
Staleness decay-0.1 if no activity for 6+ months

Validation Lifecycle

New exercise submitted
↓ Tier 2 review
pending_validation — live but unverified
↓ Tier 3 crowdsourced feedback
validated — community-confirmed
↓ 6+ months without activity
stale — re-enters validation cycle