post-launch-review

Post-Launch Review — /post-launch-review

Discovery without this step never learns: hypotheses get "confirmed" by analytics, solutions ship, and nobody checks whether the metric actually moved. This job is the difference between a pipeline and a feedback system.

When it fires

• The post_launch_review dependency (created at ship decision, deadline launch + 90 days) comes due — the dashboard will show it.
• Or the PM asks directly.

1. Fact vs promise

• Target from CONTEXT.md (Frame): metric, baseline → target.
• Actual: ask the PM for current numbers or /ingest the fresh export.
• Verdict: hit / partial / miss — with the honest delta, not adjectives.
• On partial/miss, reconcile with the discovery signal: which validated hypothesis over-promised, was it the evidence (source too weak), the sizing (pool smaller than estimated), or the execution (shipped thing ≠ tested thing)? That diagnosis — not the miss itself — is what the knowledge base and the next initiative need.

2. Production verdicts for hypotheses

For every hypothesis that drove shipped work: hypotheses.py set <id> --confidence 0.95 --note "confirmed in production: <actual effect>" — or downgrade/refute if production disagreed with the discovery signal. REAL production data beats everything (evidence-typing).

3. Bank the knowledge — knowledge/facts.json

The PM-level knowledge base lives at the repo root in knowledge/ (personal, gitignored automatically). Append facts a FUTURE initiative would want on day one:

{
  "fact": "connected online payment multiplies checkout CR severalfold",
  "metric_effect": "+N% CR",
  "initiative": "<slug>",
  "source": "post-launch review YYYY-MM-DD",
  "date": "YYYY-MM-DD",
  "tags": ["checkout", "payments"]
}

Format: {"version": 1, "facts": [ ... ]}. Create the file if missing. Facts must be product truths ("X drives Y"), not initiative trivia.

4. Calibration

Count from registry history across initiatives: how often did INFERRED hypotheses survive REAL validation? Report the ratio — it calibrates how much to trust the next INFERRED batch.