NASA FIRMS detects thermal anomalies. SentinelCore decides what they mean locally: the decision model turns detections, forecast weather, land/fuel context, source quality, uncertainty and community feedback into ranked, explainable wildfire intelligence.
• Deterministic state machine — six decision modes with prohibited claims per mode; elevated fire-weather can never be worded as a detected fire.
• Evidence graph — every run links detections, forecast, context and source quality to the output with weighted, reasoned edges.
• Explicit calibration — method, reliability band and confidence interval ship with every score; an uncalibrated region says so in the report.
• Counterfactuals — each decision names what would change it, computed from the model’s own documented thresholds.
Six bundled scenarios run through the full decision engine with zero credentials — each must land in its expected decision mode.
Running scenarios…
Backtest Studio — demo run
Running demo backtest…
External worker & job queue
checking…
Heavy compute (archive imports, large backtests, calibration, raster features) never runs in a web request: it is queued as a sentinel_jobs row and executed by the external worker — npx tsx scripts/workers/sentinel-worker.ts — locally, in GitHub Actions, or on Cloud Run. Safe deterministic jobs can also run in-process:
Source quality & model effect
Checking sources…
Export judge report
A deterministic full SentinelCore report (bundled scenario, fixed clock) — copy it into slides or download the machine-readable JSON.