Account Rupture-Precedent Radar Automation: A Trend Never Faked Before the Data Supports It

An agency rarely loses a client without warning signs — a slow pileup of overdue deliverables, a widening gap between what the contract is worth and how much operational strain it actually costs to service, a client quietly going quiet. The trouble is that these signals live scattered across a task list, a billing spreadsheet and someone's memory of "didn't we lose someone like this before?" — never combined into one place that would let a team catch the pattern while there is still time to act on it, rather than recognizing it only after the cancellation email arrives.

Combining an operational burden score, a commercial deal-quality score, a genuine time-series trend and a precedent match against historical cancellations is a meaningfully harder engineering problem than any one of those pieces alone — most tools stop at a single burden or health score and leave the comparison-against-precedent step as something a human is expected to remember to do manually, which they eventually stop doing under normal workload. And building the trend component honestly is its own discipline: it is much easier to compute SOME number from day one and display it confidently than to build in an explicit "not enough data yet" state and actually show it instead of a fabricated line on a chart.

The radar starts from a burden score (0-100) computed today from the client's actual tasks: active-task load (capped at 30 points), real overdue deliverables — derived from the delivery or correction deadline date, never from a status field, for a documented reason explained below — (capped at 35), deliverables due today (capped at 15), high-and-urgent-priority demand (capped at 15), plus bonus weight when task volume exceeds what the client's own declared complexity level would predict. A paired deal-quality score (0-100) weighs monthly contract value, the inverse of that same burden score, the contract type, how well the client's declared complexity aligns with its actual burden band, and a bonus for zero current overdue items. Both scores are calculable immediately, for active or already-inactive clients alike, with no accumulated history required. Layered on top is a genuine day-by-day behavioral trend, read from a dedicated signal table that starts empty for every client and grows one row per day — the trend classification (rising, stable, falling, or "no history yet") only activates once at least five real days of signal exist; below that threshold, the radar states explicitly that the trend is not yet reliable rather than computing one from insufficient data. A cancellation precedent is then searched for: the engine measures a straight-line distance between the current client's (burden, deal-quality) profile and every already-inactive client who has a recorded deactivation reason, keeps the closest match, and cites it as precedent ONLY if that distance falls under a fixed threshold — above that threshold, no precedent is cited at all, however close the "best available" match happens to be, because a distant profile is not a responsible comparison no matter how much better it is than every other inactive client. The final composite risk score weighs burden most heavily, followed by inverse deal quality, real overdue count, a rising trend, and precedent when one is genuinely close — producing one of four risk levels (low, attention, high, critical) alongside a plain-language list of exactly which factors drove that specific score for that specific client.

An account's burden score climbs to 68 over several weeks as its task backlog grows and two deliverables slip past their correction deadline, while its deal-quality score sits at 32 — a mismatch between what the contract pays and what it now costs to service. The account's signal history has only three days recorded so far, so the radar reports the trend honestly as "not enough history yet" rather than guessing a direction from three data points.

Searching against clients already marked inactive, the engine finds one whose own (burden, deal-quality) profile at the time of cancellation was a close match — well under the distance threshold — and its deactivation reason ("client cited unsustainable back-and-forth on scope") gets surfaced directly as the cited precedent. A second inactive client had cancelled for unrelated reasons entirely, with a very different operational profile; its distance is well above the threshold, so it is never mentioned, even though it is technically the second-closest match available. The account manager sees one composite score, one risk level, and one specific, genuinely comparable precedent — not a list of every past cancellation ever recorded.

A pattern that used to live only as a scattered set of individually-visible symptoms — rising task load, a slipping deadline, a client gone quiet — becomes one composite, explained score an account manager can act on before the relationship actually breaks. A senior team member's "this reminds me of a client we lost" hunch becomes a structured comparison anyone on the team can see, institutional memory encoded in data rather than trapped in one person's recollection. And a team can trust the trend component specifically because it is willing to say nothing rather than say something false — the same honesty standard the rest of this pillar's Run pages already hold to.

A brand-new operation with no clients yet marked inactive, or with too little signal history across the board, will see mostly "not enough history yet" trend states and no precedent citations — that is the system working correctly, not a missing feature; the radar is built to wait for real data rather than manufacture confidence it has not earned.

A team member's memory of a similar past cancellation is real signal, but it is unstructured, hard to pass on to a new hire, and impossible to verify against the actual numbers. Centriu Run's radar turns that same kind of comparison into a repeatable, capped, explainable calculation — one that refuses to cite a precedent that is not genuinely close, and refuses to compute a trend before enough real days of data exist to support one.