Sentinel scores every active client 0–100 on each pipeline run, classifying the account into one of four health states.
| Score Range | State | Meaning |
|---|---|---|
| 75–100 | Healthy | Strong engagement, timely payments, positive sentiment |
| 40–74 | Drifting | Declining engagement, late payments, neutral/negative tone shift |
| 0–39 | At Risk | Multiple compounding negative signals — non-response, overdue invoices, disputes |
| 75+ AND expansion signal | Expansion Ready | An overlay, not a separate score band — see below |
A corrected design decision: an earlier version of the rubric placed expansion_ready as its own band below healthy (60–79 vs. 80–100 for healthy) — meaning a client's best accounts, showing genuine expansion appetite, scored lower than merely stable ones. This was identified as an inverted-semantics bug and fixed: expansion_ready is now an overlay condition requiring a healthy-range score plus an explicit signal (an expansion ask, a referral, a mentioned budget, or a high NPS). health_state is computed in code from the score and this flag — the model's own state label is advisory only, so the rubric can never be silently violated by the LLM.
The scoring prompt explicitly encodes the domain principle that makes this useful: a single late invoice is a minor event. A late invoice, plus a cancelled meeting, plus reduced email engagement, is a serious compound signal. The model is instructed to reason across the full signal set holistically rather than scoring each event independently.
Without damping, an LLM re-scoring the same 30-signal window from scratch on every run can swing a client's score wildly between adjacent runs with no real change in the relationship — alert fatigue that would kill trust in the system within two weeks of real use.
The fix: if the raw score delta from the previous run exceeds 20 points and no genuinely concerning signal justifies it, the movement is clamped to ±20 and the rationale explicitly notes the dampening. health_state is recalculated after damping, not before.
Each scoring run persists: the score, state, delta from the previous score, a plain-English rationale, up to three key risk signals, up to three key positive signals, and whether expansion signals were detected — giving every score a full audit trail rather than an opaque number.