Explee · company spend

How to read this. Burn is a Theil–Sen slope over the readings since the last top-up, not a first/last difference: in this window a first/last estimate reports findymail burning −3623 credits/h, because a +1994 top-up lands inside it. Time to impact is hours, the only quantity comparable across every provider.
What is and is not added up. A card shows a total only when its unit is fungible across vendors, which here means a currency: a dollar of openai credit and a dollar of brightdata credit are both a dollar. The six providers whose unit is called “credits” are not summed — elevenlabs credits are TTS characters, resend credits are emails, scrapfly credits are API calls. They share a label, not a unit, so that card reports how many packages there are and which one runs out first.
Alert classes. operational_policy rules encode choices nobody specified (runway lead time 24h critical / 72h warning, unavailability tolerance 15min, postpaid floor -500). data_derived rules compute their threshold from the observed window (6 MAD of the provider's own slope distribution).
Why 15 minutes. In the reference window transient 504/429 failures lasted 1–2 polls (30–60s) and self-healing 5xx episodes lasted 10–22 polls (300–630s) — 15 of them across 10 providers in under three hours, every one self-healing. The tolerance sits above the longest outage actually measured, so a line means "longer than anything we observed", not "the API is flaky again". Freshness above turns amber at 300s so a provider going quiet is visible long before it is alerted.
Not alerts. A top-up, a package reset on its refresh date, a postpaid credit going negative, a rise that is later handed back, and a single timeout are all normal. Estimate-driven rules sustain for 300s before firing.
Lines are written on change, not on a timer. A condition produces a line when it starts and again only when it crosses a materiality band — roughly doubling steps of time-to-impact. Drift inside a band is silent, because 44 h and 43 h call for the same response. This table is the live state; alerts.jsonl is the log of changes to it.
Known measurement limit. A top-up landing in the same interval as spend is not observable: the API exposes a current value only, so the two are seen summed and never separately. Burn is a lower bound across such intervals.