A worked example: the nightly pass on one phantom deal

Abstract rules are easy to nod along to and hard to picture, so walk one deal through the loop.

The deal. "Acme Corp , Platform Renewal", 48k, stage 4 of 5 ("Negotiation"), forecast category Commit, close date last Friday, 90% probability. On the board it looks like one of the strongest deals in the quarter.

What the nightly pass finds. The stage-integrity rule fires: stage 4 but zero meetings logged in 23 days and the last email thread went one-way, two unanswered notes from the rep. The decay rule fires: the close date is four days in the past and has been pushed three times across two quarters. The dedupe rule fires: there is a second "Acme Corporation" account carrying its own 48k renewal, the same deal, double-counted.

What the agent does autonomously. It merges the duplicate account, stripping a phantom 48k out of the pipeline total, and logs the merge with both prior records attached so it can be undone. It standardises the stage label (the second copy read "negotiation" lowercase). It does not touch the probability or the forecast category, those are judgement.

What it surfaces to the human. A single line in the morning report: "Acme renewal carries Commit at 90% but shows no two-way engagement in 23 days and a thrice-slipped close date; recommend moving to Best Case or booking a re-engagement call. A duplicate Acme account was merged (-48k phantom)." The number the founder forecasts on just dropped by 48k of fiction and flagged 48k more of fragile optimism, before any human spent a minute on it.

Run that across a 60-deal pipeline every night and the compounding is obvious. The agent is not predicting anything clever here. It is refusing to let the pipeline lie, deal by deal, while everyone sleeps. That is the whole job, and it is why the forecast on top of it can finally be trusted.