Iterating post-launch

The launch plan survives contact with reality for about four days. Post-launch product judgment, in its recurring shapes:

• Week one is forecast-vs-actual, not vibes-vs-hopes: the miss-path rates against your scope contract's forecasts, acceptance by segment against the prototype's promise, the support team's qualitative read (schedule it — they know things the dashboard doesn't yet). Expect one genuine surprise; the discipline is chasing it to a mechanism (the Module 4 drill) before reacting to it. Dana's real week one: acceptance strong overall (~74%), but the first-time-customer segment sits far lower — ~48% accept, versus ~78% for returning customers — mechanism, after twelve transcript reads: customers with no prior Harbor Lane relationship read an unsolicited 'we noticed a problem with your order' as suspicious rather than helpful, and want a plainer identity signal before a one-click fix. That's a copy-and-first-run fix (say who we are and why we're reaching out), not a model fix — the layer sorting, live.
• The fix-vs-feature allocation is the PM's weekly knife-edge: the improvement backlog (Module 4's harvest) competes with the roadmap's next feature for the same engineers. The honest heuristic: while any quality-contract dimension is under bar or any miss rate is out of band, the loop outranks the roadmap — a probabilistic product below its bars is spending trust budget every hour it runs. Above bars, ship the roadmap and let the loop run on its standing budget line. Teams that invert this — new features atop an under-bar core — build the AI products users describe as 'impressive and unreliable,' which is the genre's epitaph.
• Scope expansion is earned by data, granted in steps: v1's tolerant position (propose, human accepts) earns v2 autonomy (auto-resolve under $X for tenured customers) only when the numbers argue it — acceptance rates near-verbatim, edit distance near zero, incident-free at volume in that segment. Expand one rung at a time, instrument the new rung like a fresh launch, keep the sampling (the finance course's threshold-history discipline: every loosening documented with the evidence that earned it). The ladder is slower than ambition and faster than rebuilding trust.
• And sometimes: kill it with grace. If the value hypothesis fails — resolutions accepted but tickets don't drop; CSAT flat; the segment that loves it too small to matter — the feature earns the strategy course's process-success kill: announce plainly, thank the users in the exposed cohort, write the postmortem that makes the next bet smarter, and keep the instrumentation learnings (the episode schema and eval assets transfer; sunk features leave usable estates). PMs who can kill their own launches get trusted with bigger ones. That's not consolation; it's the actual career mechanics.