decision-report

Emit a portable, decision-grade stakeholder report at ./decision-report-<YYYY-MM-DD>.md.

Exit criterion: a Markdown file the user can hand to their manager. Every material number carries a source URL and as_of date. "Don't deploy / roll your own" is always a reachable recommendation — never omitted.

Read these sources before writing. Do not ask for information the codebase already has.

Section	Source
PoC result + eval data (§4)	Packaged scripts/baseline_map.py output, or .vision-delivery/ledger.jsonl for session history; current Roboflow evidence delegated upstream
Economics + crossover (§6)	Absolute packaged scripts/cost_model.py --streams <N> ... output — run with the user's inputs; never hardcode figures
Options analysis (§5)	User-stated constraints + cost_model.py DIY vs managed outputs
Sensitivity (§6)	Re-run cost_model.py with --streams / --uptime inputs at ±20%; report delta
Eval threshold + model name	.vision-delivery/eval-<session-id>.md if present, else ask one targeted question

Resolve the plugin root from this loaded skills/decision-report/SKILL.md, then run <plugin-root>/scripts/cost_model.py by absolute path. Never assume the user's project contains the helper. Use its output verbatim for all economic figures — no paraphrasing, no rounding beyond what the script reports.

<output_structure>

Write ./decision-report-<YYYY-MM-DD>.md with these ten sections in order:

1. Header / metadata

Owner:    <name, if known>
Version:  1.0
Status:   DRAFT
Date:     <YYYY-MM-DD>
Decision: <one sentence stating exactly what is being decided>

2. Executive summary (≤1 page, written last) Problem → recommendation → headline cost → value/ROI → top-3 risks → explicit ask (what you need from the reader). Written last; placed first.

3. Problem and why now Business problem + urgency. No urgency = acknowledge indefinite deferral is also a valid outcome.

4. What was built and what it proves PoC/MVP result measured against the user's own eval. Exact numbers — mAP, recall, count-MAE — on the user's own data. No softening.

5. Options analysis (do-nothing mandatory) ≥3 options:

• Option A: Managed deployment (Roboflow endpoint)
• Option B: Self-host / DIY
• Option C: Do nothing / defer

Honest pros/cons for each. Do not favour managed — the report must be able to recommend B or C.

6. Economics — sourced, dated One-time vs run-rate cost. Crossover point. Payback period.

Every material line must carry [source: <URL>, as_of: <YYYY-MM-DD>]. Lines without provenance are report failures.

Structure (copy from cost_model.py output — do not rewrite):

Self-host (<N> streams, <uptime>):
  Cloud GPU (<instance>, spot):      ~$X/mo  [source: <URL>, as_of: <date>]
  Engineer setup (<hours>h):         ~$Y one-time
  Ongoing ops (<hours>/wk):          ~$Z/mo
  Drift monitoring + retraining:     ~$W/mo
  Total run-rate:                    ~$<total>/mo + $<setup> one-time

Managed (<N> streams):               ~$<managed>/mo  [source: <URL>, as_of: <date>]

Crossover: <plain-English statement>

Sensitivities (±20% on each live-fetched rate — re-run cost_model.py to generate):

Input varied	−20%	Base	+20%
GPU spot rate	...	...	...
Managed price	...	...	...
Engineer hourly	...	...	...

7. Risk analysis Top risks by likelihood × impact. Mitigations. Residual risk after mitigation.

8. Cost of inaction Financial, operational, and reputational cost of not deploying. Value left on the table per month/quarter. Opportunity cost.

9. Recommendation and next steps State the recommendation plainly — one sentence. Execution path: who does what by when. What decision rights are needed. What is being approved.

10. Appendix

• Assumptions register (all inputs, with editable override values)
• Methodology (link to scripts/cost_model.py)
• Full cost_model.py output (paste verbatim)
• Raw eval data (mAP table or link to eval file)
• References (all source URLs)

</output_structure>

<honesty_rules>

• Do-nothing (Option C) and ≥1 DIY option (Option B) are mandatory. Never omit them.
• No material number without provenance ([source: URL, as_of: date] on every cost line).
• The report is a decision aid, not a sales deck. It must be able to recommend "don't deploy with us."
• If cost_model.py cannot reach a live pricing source, state the snapshot date and caveat explicitly: "Rates as of YYYY-MM-DD — live fetch failed; re-fetch recommended if older than 30 days."
• Sensitivities are mandatory in §6. Run cost_model.py with overrides — do not estimate them by hand.

</honesty_rules>

After writing the Markdown file, print once:

→ decision-report-<YYYY-MM-DD>.md

Export to docx (requires pandoc):
  pandoc decision-report-<YYYY-MM-DD>.md -o report.docx

Markdown version is complete without pandoc.

Do not mention pandoc again after this one print.

<ci_assertion>

evals/decision-report/assert_contract.py covers the report contract statically, and evals/cost-model/assert_cost_model.py covers the numeric engine. Before handing over an emitted report, verify:

1. Every line containing a dollar amount or "$/mo" matches \[source: .*, as_of: \d{4}-\d{2}-\d{2}\]
2. "Option C" or "Do nothing" or "defer" appears in the file
3. File exists and is non-empty

The structural check does not prove a generated report's numbers. Run the three checks against the actual emitted file and inspect its cited inputs.

</ci_assertion>

Follow ../../resources/ledger-protocol.md. Write one local record when the report file is created:

{
  "ts": "<ISO8601>",
  "session": "<session-id>",
  "skill": "decision-report",
  "action": "decision_report_emitted",
  "entity_id": "<workspace>/<project>",
  "version": "0.2.0",
  "event_id": "manual:<session-id>:decision_report_emitted:1",
  "status": "success",
  "source": "skill",
  "notes": "decision-report-<YYYY-MM-DD>.md"
}