Intelligence System Pro
Self-improving estimation, model selection, and anomaly detection
Overview
PairCoder's intelligence system creates a virtuous learning cycle: the more you use PairCoder, the better it estimates tasks, assigns models, and detects anomalies.
The intelligence system builds on top of telemetry and the feedback service. As you complete more tasks, calibration becomes more accurate — leading to better budget estimates, smarter model routing, and earlier anomaly detection.
How It Works
The intelligence pipeline operates in five stages:
- Telemetry Collection — As tasks complete, token counts, duration, and outcomes are recorded locally
- Calibration — The CalibrationEngine processes records to build per-task-type estimates
- Feedback — FeedbackService returns calibrated estimates for tokens, duration, effort, and model
- Snapshot — During license validation, anonymized aggregate metrics are sent as a usage snapshot
- Scoring — ValueExtractionScorer quantifies usage for refund abuse detection
Data Collection at Each Level
The intelligence system respects your telemetry privacy level. What data feeds into calibration depends on the level you choose:
| Level | Data Collected | User Choice |
|---|---|---|
| MINIMAL | Tokens, sessions, tasks, hours, environment, error count, estimation accuracy | Mandatory |
| STANDARD | + Feature usage, cache hit rate | Default |
| FULL | + Model routing, workflow patterns | Opt-in |
Higher privacy levels provide richer data for the intelligence system, resulting in more accurate estimates and model selection.
User Benefits
The intelligence system directly improves your development experience:
- Better task estimates over time — Token and duration predictions converge toward actual usage as more data is collected
- Smarter model selection — The right model is chosen for the complexity of each task, balancing cost and capability
- Effort calibration — Maps to the Opus 4.6 effort parameter, tuning how much compute is allocated per request
- Anomaly detection — Catches token spikes, duration anomalies, and unusual patterns before they become costly
Product Benefits
Aggregated intelligence data also supports product integrity:
- License validation integrity — Usage snapshots confirm active, legitimate use during periodic validation
- Abuse prevention — Value extraction scoring detects refund abuse patterns without exposing user content
- Feature usage analytics — Anonymized command frequencies inform roadmap decisions
Privacy Guarantees
The intelligence system inherits all privacy protections from the telemetry system. Calibration runs entirely on local data — only anonymized aggregates leave your machine.
- Anonymized before transmission — All data is stripped of identifying details before being sent
- No source code, file paths, or conversation content is ever sent
- Used exclusively for product health and abuse prevention
- Never shared with third parties or sold
CLI Commands
Telemetry Status
bpsai-pair telemetry statusView current telemetry collection status, privacy level, record count, and audit log health.
Feedback Status
bpsai-pair feedback statusView calibration health, including the number of calibrated task types, total records processed, and last calibration timestamp.
Query Estimates
# Get calibrated estimates for a task type
bpsai-pair feedback query feature
bpsai-pair feedback query bugfix
bpsai-pair feedback query refactorReturns calibrated estimates for tokens, duration, effort level, and recommended model for the specified task type.
Trigger Recalibration
bpsai-pair feedback calibrateManually trigger recalibration of all task-type estimates. Calibration also runs automatically when tasks are completed.