Every metrics review needs a narrative — the 'what changed and why it matters' paragraph above the chart — and that narrative is rewritten by hand every week by someone whose actual job is elsewhere. Skip the narrative and the numbers get misread; write it hastily and a wrong causal claim ('signups dropped because of the pricing change') propagates into decisions.
Numbers without governance are how dashboards lie to executives.
A scheduled Fleet workflow drafts the narrative from exported metrics files, a review step checks every numeric claim against the data and flags causal overreach, and the owner approves before the report circulates.
genflows:
- name: weekly-metrics
schedule: "0 8 * * 1" # Monday mornings
steps:
- {name: narrative, prompt: "Write the weekly metrics narrative from the exports: movements, trends, anomalies. Cite the file for every number. NO causal claims without corpus support.", corpus: ["data/exports/*.md", "data/notes/*.md"], kind: data, out: report.md}
- {name: numbers, prompt: "Verify every number against the exports; flag arithmetic drift and causal claims the data can't support.", depends_on: [narrative], kind: review, out: flags.md}
- {name: owner-ok, depends_on: [narrative, numbers], kind: approval, out: decision.md}
- {name: send, depends_on: [owner-ok, narrative], kind: notify, out: sent.md}
The review step is the institutional skeptic: a claimed number that doesn't match the export, or a 'because' the data doesn't support, parks the run at the gate with the flag attached.
Metrics exports → narrative draft with citations → number-verification review → owner approval → notify. The chart gets its paragraph, the paragraph gets its fact-check, every week, structurally.
Ops, growth, and BI owners who circulate a recurring metrics summary assembled from exports — anywhere the narrative layer is hand-written and error-prone.
No — the corpus is files. Teams land a weekly export (CSV rendered to markdown, or a metrics summary file) in the repo via their existing pipeline; the workflow takes over from the files. Your data stays in your infrastructure, and the runs execute on your own worker.
It's a document-producing step like report, signaling data-shaped output. The governance is identical: corpus in, artifact out, review and approval downstream.
One binary. Five minutes. See every agent, coordinate every handoff, and keep a full audit trail of what your fleet did.