outcome-review

name	outcome-review
description	Grade shipped product bets against what their discoveries predicted — pull each registered metric from its source of truth, walk the watch items, verdict every falsification line. Use when (1) user says "/outcome-review", "did it work", "grade our bets", "outcome review", (2) a month has passed since the last review, (3) any outcomes-register row's first check date has passed.

Outcome Review

Every discovery ships with predictions: success metrics with baselines, falsification lines ("we're wrong if…"), watch items. This skill is the half of the job that happens after shipping — without it, a wrong build verdict and a right one look identical forever, and the next verdict can't be calibrated by the last one.

Bets are registered at ship time by /sdlc and /adhoc-fullstack-feature; this skill grades them.

Inputs

• The register: {{root}}/docs/discovery/outcomes-register.md — every row whose status is watching and whose first check date has passed. Never check a bet before its registered date: an early read measures the novelty spike, not the truth.
• Per feature: its discovery folder (DISCOVERY.md §§12–13, 16, or DISCOVERY-LITE.md's Design Decisions bets + "How we'll know it worked").

Method — per due bet

1. Pull the source of truth the register names.

   Customize this section for your product — each bet names one of these metric stores:

   • {{YOUR_SOURCE_ANALYTICS}} — product analytics events (the codebase logs one on every state-changing action)
   • {{YOUR_DB_QUERY_COMMAND}} — read-only production queries (the /local-testing and /observability patterns)
   • {{YOUR_LOG_SOURCE}} — error/exception monitoring

   If a source can't be pulled in-session, name the exact query/export needed and grade inconclusive — blocked on {x}; an unmeasurable bet is a registration defect to name, not to skip.

2. Compare against the registered baseline and expectation. At small scale, evidence is named-actor behavior — "both operators opened it daily for three weeks", "2 of 5 self-serve users who saw it acted" — never percentages over tiny denominators.

3. Walk the watch items and falsification lines. Each falsification line gets an explicit verdict: fired / didn't fire / can't tell.

4. Answer three questions, in writing: What happened? What did we learn? Did the feature everyone was sure about deliver what the discovery promised?

5. Grade the bet: confirmed / falsified / inconclusive — and one mandated line: what we got wrong, even on confirmed bets (a scope cut, a missed edge case, a metric that measured the wrong thing). A review reporting only wins is suspicious by base rate — even elite experimentation shops find most ideas don't move their target metric; if every bet reads confirmed, the first hypothesis is that the measurement is too soft.

Writes

• {feature}/OUTCOMES.md in the feature's discovery folder — dated, the per-bet detail from the method above.
• The register row: status updated, grades appended.
• {{root}}/docs/proposals/DECISIONS.md: the original verdict's row gets → confirmed / → falsified (date) so future intakes see how past verdicts aged.
• Falsified bets feed the corrections loop: when the miss generalizes (a class of assumption that keeps failing, an instrument that keeps overstating), encode it into the relevant {{root}}/.claude/skills/ux-discovery/references/*.md file as a BAD/GOOD pair and close with "Encoded: {file}" — the same mechanism requester corrections use. A falsified bet is a result the pipeline paid for; wasting it is the defect.

Report

Anomaly-first digest in chat: bets graded this run (falsified first, with what they imply — iterate / remove / leave), what's blocked and on what, which watching bets come due next, and any register hygiene problems: sweep {{root}}/docs/discovery/*/ for discoveries whose feature shipped (merged PRs) with no register row — features shipped outside /sdlc and /adhoc-fullstack-feature miss the registration step, and this sweep is what catches them. Where a falsified bet traces to a specific requester's ask, say so plainly and kindly: the bet didn't pay, the raising was still right — the intake pipe lives on people staying willing to raise things.