Roadmap
SATI is a foundation, not a finished product. Here's where we're heading and the hypotheses driving development.
Current State
The system today provides pattern recognition and guided feedback for jhana practitioners based on text descriptions. It works, but with meaningful limitations (64% top-1 accuracy, 23% false positive rate).
Phase 1: Data Foundation
Status: In progress
What We're Building
- Session persistence with device-based identity
- Cross-session pattern tracking
- Practitioner outcome logging
- Feedback collection on response helpfulness
Key Hypothesis
Recurring patterns across sessions are more predictive than single-session classifications. A practitioner who shows "over-efforting" in 4 of 5 sessions has a different profile than one who shows it once.
Success Metrics
- 500+ logged interactions with outcomes
- Pattern recurrence data across 50+ multi-session practitioners
- Correlation analysis between pattern profiles and breakthrough timing
Phase 2: Calibration Refinement
Status: Planned
What We're Building
- Feedback-driven calibration adjustment
- Per-practitioner confidence weighting
- Pattern co-occurrence modeling
Key Hypothesis
Calibration multipliers should be learned from outcomes, not set manually.Current multipliers are based on intuition about detection difficulty. Real practitioner feedback will reveal which patterns we're actually good at detecting vs. which we overestimate.
Success Metrics
- Confidence-accuracy correlation above 0.80 (currently 0.71)
- False positive rate below 15% (currently 23%)
- Top-1 accuracy above 75% (currently 64%)
Phase 3: Proactive Guidance
Status: Research
What We're Building
- Pattern trajectory prediction
- Suggested practice experiments based on profile
- Personalized session preparation prompts
Key Hypothesis
Pattern succession is predictable. When over-efforting resolves, grasping often becomes visible. When grasping resolves, fear of depth often emerges. If we can predict the next likely obstacle, we can address it before it fully manifests.
Success Metrics
- Next-pattern prediction accuracy above 60%
- Reduced sessions-to-breakthrough for practitioners using predictions
- Practitioner-reported value of proactive suggestions
Phase 4: Teacher Integration
Status: Vision
What We're Building
- Teacher dashboard showing practitioner patterns
- Session summaries for teacher review
- Teacher feedback integration for model improvement
- Handoff protocols for complex cases
Key Hypothesis
AI pattern recognition amplifies teacher effectiveness rather than replacing teachers. A teacher who can see "this practitioner has shown fear-of-depth in 3 of their last 4 sessions" before a call can provide more targeted guidance in less time.
Success Metrics
- Teacher time per practitioner reduced while outcomes improve
- Higher practitioner satisfaction with teacher sessions
- Teachers report the tool surfaces useful patterns they would have missed
Open Questions
These are genuinely uncertain and will shape direction:
- How much personalization is appropriate? Should the system learn individual practitioner patterns, or does this risk overfitting to their stated (vs. actual) experience?
- When should the system defer to humans? Beyond safety escalation, are there complexity thresholds where AI guidance becomes counterproductive?
- Can we validate meditative states? Practitioners report experiences, but we can't independently verify them. Is self-report sufficient, or do we need other signals?
- What's the right interaction cadence? Real-time during sits? Post-session reflection? Periodic check-ins? Different practitioners may need different rhythms.
Non-Goals
To stay focused, we're explicitly not building:
- Guided meditation audio. There are good solutions for this. We focus on pattern recognition, not session guidance.
- Gamification or streaks. Jhana practice is not a habit app. Extrinsic motivation tends to reinforce the wrong patterns.
- Community features. Not our core competency. Better to integrate with existing communities than build another.
- General meditation support. We're specifically focused on jhana practice. Depth over breadth.
Timeline Assumptions
We avoid specific dates because they depend on data collection velocity and what we learn. General sequencing:
- Phase 1 provides the data foundation everything else requires.
- Phase 2 can begin once we have 200+ interactions with feedback.
- Phase 3 requires validated calibration and clear pattern trajectories.
- Phase 4 requires a working system worth integrating with teachers.
Each phase informs whether the next makes sense. We're building to learn, not executing a fixed plan.