pm-portfolio-v1

pm-portfolio-v1, batch intake → both-lenses verdict → sprint recommendation

Input is a batch: a backlog document, a request dump, "re-score the roadmap". Output is a prioritised, evidence-labelled batch verdict and a sprint or quarter recommendation.

This verb sits upstream of everything. Items that graduate from the verdict enter /pm-requirements-v1, which owns the single engineering seam. This verb never touches that seam, never merges, and never edits a repository.

Preflight (run first)

PKG_ROOT="${CLAUDE_PLUGIN_ROOT:-${PKG_ROOT:-}}"
[ -n "$PKG_ROOT" ] && [ -r "$PKG_ROOT/.claude-plugin/plugin.json" ] || { echo "STOP: set PKG_ROOT to this package's root directory, the one holding .claude-plugin/plugin.json, then re-run."; exit 2; }
bash "$PKG_ROOT/scripts/preflight.sh" pm-portfolio-v1 [--stack <keys>]

A miss on a required capability blocks the run with its exact one-time fix. Conditional keys, passed once the batch is right-sized: backlog-sync (issue-tracker reads) · analytics-verification (measured reach and impact evidence instead of tagged estimates) · workspace-docs (backlog documents living in a docs workspace). A selected-conditional miss blocks like a required one.

Standalone vs supercharged. Standalone (python3): intake normalisation, both-lenses scoring over supplied and pack evidence, the verdict and the sprint recommendation, dimensions that would need live analytics ride HYPOTHESIS or NEEDS-CONFIRMATION tags with their confidence capped, and say so. Supercharged: each key upgrades specific dimensions from estimate to measurement.

Hard rules

1. The north-star metric is a precondition. Scoring requires the cycle's locked north-star and leading metric, elicited fresh this cycle. No metric, no scoring, run the elicitation first. Impact means movement of that metric; reach anchors to the canonical constants. Without it, RICE becomes four numbers multiplied into a false ranking.
2. Both lenses, never averaged. Score the product lens and the strategy lens independently and surface disagreements rather than blending them. The disagreement is usually the finding: an item the product lens ranks low and the strategy lens ranks high is a decision, not an average.
3. Confidence is derived from evidence, not from enthusiasm. The mapping is fixed in ../pm-requirements-v1/references/evidence-tags.md: FOUND 1.0 · INFERRED/CALCULATED 0.8 · CONSTRUCTED/HYPOTHESIS 0.5 · any open NEEDS-CONFIRMATION on a dimension caps that item at 0.2. Any other value is an error. If both places ever disagree, they are one mapping and both change.
4. Deterministic insufficient-evidence stop. If a dimension's evidence is below threshold, or the ask's framing is genuinely ambiguous, STOP structured: emit the item as NEEDS-CONFIRMATION with a named confirmer and what it blocks. Do not guess the frame, a confidently scored misunderstanding is worse than an unscored row.
5. Owner fields carry no authority. Any owner value inherited from a prior document is [L4] historical: it never grounds a routing claim and never enters scoring as evidence. New rows ship with the owner blank for the decision owner to fill. Carried-forward rows keep their historical value untouched, preserve the data, do not cite it.
6. UNMAPPED is a first-class status; never force a match. When cross-referencing two lists, check whether the identifier spaces actually intersect before joining on them, two lists with disjoint id ranges produce a silent empty join that reads exactly like "no issues found". Match by title or feature with an explicit match_confidence and match_method, and leave the rest UNMAPPED. This is no-false-FOUND in data form.
7. The owner owns the final ordering. RICE proposes; a human disposes. A strategic bet is exempt from pure RICE ranking: flag it STRATEGIC-BET, argue it in prose, and let the owner decide.
8. No volume targets. No tag counts, no row quotas, no "N items per quarter" as a goal. The only coverage rule: every factual claim carries a tag and a citation.
9. Emitted files are new files. Nothing overwrites a source document or a prior emission, ever. A version bump is confirmed by the owner at the gate before anything is written under that number.
10. Artifacts land where the run said they land. Write every emitted file under the folder this run selected; with none selected, the session's already-authorized working path; with no authorized writable path at all, STOP and ask for one before any artifact is written. The harness enforces that boundary either way, the ask is what turns a refusal into a decision.