Intelligence
FORG's “intelligence” is deterministic analysis over the metadata you already send — no model calls, no embeddings, $0 in LLM cost. Every number below is computed from your telemetry with a fixed formula you can reproduce.
Anomaly flags
As signals arrive, the ingestion engine flags a small set of per-session anomalies heuristically. These are lightweight integrity checks, not statistical baselining:
- Rapid-repeat signal — the same signal type fires again faster than the expected cadence within a session
- Extreme hesitation — a gap of more than 30 seconds between a prompt and the next action
When an anomaly is detected it is written to your Dashboard → Alerts feed (category Anomaly) and, if you have a webhook configured for alert events, delivered there too. FORG does not yet do spend-spike, token-burst, or model-shift anomaly detection — see the roadmap below.
Spend forecasting
Ask FORG can project your future spend on demand. It fits a deterministic least-squares line over your recent daily cost series and extrapolates the requested horizon, summing the predicted daily costs (floored at zero) and returning a residual-based ± band:
cost ≈ intercept + slope · day_index (basis: least_squares)This is a simple, explainable projection — not a calibrated statistical interval — and it runs with zero LLM cost. It needs at least two non-trivial days of data; with less, FORG degrades honestly and tells you it cannot fit a trend rather than inventing one.
Model efficiency
FORG records tokens and cost per session, tagged by model, so you can compare usage across developers and models. Those figures are available directly in your Usage and Sessions views:
- Tokens and cost per session, broken down by model
- Model mix over time — where spend is concentrated
FORG surfaces the numbers; it does not judge whether high spend reflects high productivity or waste. That context is for human review.
On the roadmap
Statistical anomaly detection (per-user spend/usage baselines and deviation alerts), forecast-threshold alerting, and team benchmark comparisons are planned but not shipped today. We document features here only once they are live.