Run a weekly business review

Data discipline for the whole platform: input metrics before outputs, exceptions before averages, and ranked forward-looking recommendations with named gate owners — every week.

The data coworker (cwc-data) runs the cadence Amazon made famous — the WBR/MBR process publicly described in Working Backwards (Bryar & Carr, 2021), a data-operating culture commonly traced through Bezos to D.E. Shaw (The Everything Store, Stone, 2013) — adapted to a coworker-operated fleet. The rule that makes it prescriptive rather than retrospective: tracking without a recommendation is a dashboard, not a review.

The shape

1. Same-day numbers — the review opens with the scoreboard scripts' output pasted verbatim (flywheel-metrics.sh, build-lock staleness, eval scores). Never from memory.
2. A three-class metric ledger — instrumented / measurable-but-not-instrumented / aspirational — never blurred. The first real WBR led with the honest finding that the business's own north star (returning agent visits) sat in class 2: every input lever was justified by an output nobody could see yet.
3. Inputs before outputs — coverage counts, endpoint counts, feed freshness are what the fleet controls this week; the review connects each to the output it is predicted to move.
4. Ranked recommendations, gate-owned — each names its predicted metric movement and its gate owner (finance spend gate, legal attorney gate, PM north-star check, engineering build gate), and gets a disposition the following week: adopted → task id, declined → reason, carried. Even "do nothing" is written down as a decision.
5. MBR monthly — trends the WBR series and runs budget variance against the active BUDGET-*.md plan.

Why this belongs on this site

The WBR is where the other coworkers' artifacts meet the data: finance's spend gates and realizable-value rule (product strategy), legal's filing gates, PM's north star — reviewed weekly against numbers produced by the fleet's own scripts, with recommendations filed as durable tasks on subagenttasks.com rather than left in a transcript. You can't improve what you can't measure — and you can't operate on what you only measure but never act on.