ds-report

ds-report — Communication

Overview

Converts the evidence from /ds-evaluate and /ds-explain into a narrative a stakeholder can act on: a recommendation, its evidence, and its honest limitations.

When to Use

• After /ds-evaluate and /ds-explain have both produced their artifacts.
• Whenever asked to write up, present, or summarize DS results for a non-technical audience.
• NOT for: packaging the model for reuse (that's /ds-handoff) — this stage is about the narrative, not the artifact.

Core Process

1. Gate check: confirm .last-ds-mile/stages/07-evaluate.md includes slice or subgroup performance, not only an aggregate number. If it doesn't, compute the slice performance yourself now, say plainly that you did, and continue — never write the report around an incomplete evidence base, and never stop the task to send the user back to /ds-evaluate separately (see ds-method's discipline-gate handling).
2. If the ask is specifically to confirm the model is good to ship (not just to write it up), delegate a full-pipeline sanity check to the ds-reviewer agent before concluding — it checks baseline, validation, slice performance, and metric choice end to end in one pass, cheaper and more reliably than re-deriving that checklist inline.
3. Lead with the decision this informs (from /ds-frame), not with model architecture.
4. State the recommendation plainly, then the evidence: baseline comparison, slice performance, calibration.
5. Translate the metric lift into /ds-frame's original cost/business terms, not just metric units. /ds-frame required a success metric tied to a real decision cost (a false negative costs $Y, a 1-point AUC move is worth $Z); if that translation was done once at framing time and never carried forward, the report ends up repeating "RMSE improved by 0.04" or "recall is 0.87" with no stated dollar or operational impact at the actual chosen operating point — the exact "success metric is a pure ML metric with no tie to a business cost" Red Flag /ds-frame exists to catch, resurfacing here instead. State the lift's real-world size (e.g. "the median prediction error corresponds to roughly $X, down from $Y for the baseline" or "at the frozen decision threshold, this catches N more true positives per 1,000 cases than the baseline, at a cost of M more false alarms").
6. If the recommendation implies intervening on a feature — targeting a segment for a changed offer, pushing customers toward an option, recommending a policy change — rather than just using the model's score to rank or prioritize, check it against causal-vs-predictive before it ships. A ranking/scoring recommendation ("use the score to prioritize outreach") only needs predictive validity, already established in /ds-evaluate; an intervention recommendation needs a causal argument the analysis may not have made.
7. List assumptions and limitations explicitly — what the model does not cover, and where it's known to underperform (from the slice table).
8. Write to .last-ds-mile/stages/09-report.md: the narrative, the recommendation, the cost/business-terms translation, and the assumptions/limitations.