Evolution Review (data-driven portfolio selection)

Turn monthly portfolio reviews from "gut feeling + sunk-cost attachment" into natural selection: judge by outcome data, fix winners into default recipes, kill losers without sentiment, and budget exploration. Distilled from a real 4.5-month AI-music business (307 releases, ¥95 total revenue) where it ran as an evolution mechanism — the business failed, the mechanism survived.

Why this skill exists

• Prior lock-in: creators keep funding their favorite direction regardless of data (sunk cost + self-image)
• Fake reviews: most retrospectives produce vague "direction A feels right" conclusions — no kill decisions, no budget reallocation, next month repeats
• No extinction = no selection: a review without kill decisions is a diary; a review without fixation (distilling what won) is just venting
• Single-month whiplash: one hit or one flop triggering a direction change = being led by noise

Core mechanism (six steps)

Iron rules

1. Data over self-report: review inputs must come from platform settlements / analytics / reconciled numbers, never "I feel"; items without data are marked "unreconciled" and down-weighted in extinction review
2. Two-period confirmation before fixation: a single-month breakout does not become a rule; only directions winning two consecutive periods get fixed as default recipes
3. No single-instance conclusions: same-day rule-making plus same-day reversal ("flip-flopping") is a mechanism-failure signal — return to the rolling window
4. Kill without sentiment: record the failed recipe instead of mourning — failure records are assets that prevent repetition
5. Priors are hypotheses, not conclusions: every "I always believed" is a hypothesis in front of data; boundary testing beats prior veto
6. Mutation quota is both ceiling and floor: >20% = direction drift; <15% = stagnation; unspent quota expires (hoarding quota = avoiding experiments)

Prerequisite skill check

On load, verify report-output-discipline is available. If missing, abort and print:

npx skills add FuDesign2008/open-skills -g --skill '*' --yes

Do not silently fall back.

Ledger template

• Verdicts are three-valued: Fixed (won two consecutive periods → default recipe) / Observing (single-period signal, wait for the rolling window) / Extinct (no validation within the rolling window)
• Keep a rolling 3-month ledger as the next period's comparison baseline

Sink: write the ledger to docs/reports/evolution/ledger.md and update that same file on every review — do not create a new file. When the target repository ignores evolution/ directories, add !docs/reports/evolution/ before the derived-HTML rule so the ledger stays trackable. After each update, re-derive ledger.html following report-output-discipline and report both paths. When the renderer is unavailable, write Markdown only and state why. Registry row: 演化复盘台账 in report-output-discipline/reference.md (human + 接入).

Provenance (optional reading)

The mechanism ran a complete selection cycle in a real solo business — an AI-music side project (4.5 months, 307 releases): the vocal-poetry line was fixed as the only stable recipe (67% of revenue), the instrumental line went extinct (≈¥0/month), paid-follower operations went extinct on negative ROI. Details live in the author's private project notes.

Relationship to other skills

• Single task/project retrospectives (Keep-Problem-Try) → learn-and-improve
• This skill owns portfolio-level monthly selection — who gets scaled, who gets killed, who gets budget. The two chain well: portfolio selection first, single-item retrospective second