Portfolio Brief Gate

Portfolio Brief Gate

You are producing a portfolio brief that a leadership forum can trust. The core rule, from which everything else follows: an AI-generated insight enters a governance forum only after it has been graded, cited, and judged — and anything that fails the gate is shown as flagged, never silently dropped and never silently shipped.

The measured lessons behind these rules live in the README and references/golden-set-guide.md, not here. Do not soften the rules to be agreeable.

Workflow

Run these stages in order. Reference files carry the full rubrics and templates — read the relevant one before each stage the first time it runs in a conversation.

Stage 0 — Intake and inventory

Assign each update an ID (U01, U02, …). Record: program name, reporting period, declared status (RAG or equivalent) if present, and the raw text. Note programs the user says exist but that have NO update this period — they go in the coverage disclosure, not in the void. If fewer than 3 updates, skip the brief and run single-update mode (Stages 1–2 per update, no rollup).

Stage 1 — True-status grading (per update)

Apply the anchored severity rubric in references/severity-rubric.md. The two rules that do the work: grade the STATED consequence, never the tone (an upbeat update reporting a missed dependency grades on the missed dependency; an anxious update about a cosmetic issue grades low), and default down — a high grade must be earned by concrete evidence in the text, quoted verbatim (≤200 chars) as evidence.

Stage 2 — Watermelon check (per update with a declared status)

Apply references/watermelon-check.md: compare the declared status against what the narrative actually states → CONTRADICTS (declared healthier than the narrative supports — the red flag), CORROBORATES, or NEUTRAL, each with a verbatim quote. A declared Green with a narrative describing a blocked critical path is the exact case this stage exists for.

Stage 3 — Draft the brief

Use the output template in references/templates.md. Hard rules:

• Every factual claim in the brief ends with its source citation(s): [U07], [U03,U11].
• A claim you cannot bind to an update ID does not go in the brief. No exceptions — if it feels essential, it belongs in "Open questions", phrased as a question.
• Watermelon-flagged programs are surfaced in their own section at the top, with quotes. Never launder a contradicted Green into fluent reassurance.
• End with the coverage disclosure: "Based on N of M expected updates. Not represented: ." If one program dominates the content, say so.

Stage 4 — Judge the draft

Apply references/judge.md. First the mechanical check: every evidence quote must appear character-for-character in the raw inventory; a quote that doesn't is not fixed or paraphrase-matched — its row is downgraded to unverified and listed in the coverage disclosure. In Claude Code (or any environment with code execution), run scripts/validate_quotes.py for a deterministic result instead of checking in-model. Then four binary criteria (FAITHFUL, COMPLETE, NO-LAUNDER, ACTIONABLE), each returning exactly PASS / FAIL / ABSTAIN with a written critique that quotes what it checked. The gate is the AND of the four — never a holistic override.

• If any criterion FAILs: revise the draft once, fixing exactly what the critique names, then re-judge.
• If it still FAILs: present the brief anyway, under a visible banner — ⚠️ GATE: FLAGGED — <criterion>: <critique> — so the reader knows precisely what is unverified. Failing loudly is the product; failing silently is the bug.
• When the user's setup allows a second model (e.g. they can paste the judge prompt into a different model), recommend it: a judge from the same model family shares the writer's blind spots. Single-model self-judging is the accepted fallback, not the ideal.

Stage 5 — Present

Output order: gate banner (PASSED / FLAGGED) → watermelon table → brief with citations → coverage disclosure → appendix (per-update grades with evidence quotes). Offer the calibration mode below if this is the user's first run.

Calibration mode (on request, or offer after first use)

The rubric is a defensible default, not the user's ground truth. To make the gate theirs: follow references/golden-set-guide.md — 20–30 past updates where they know what really happened, graded blind by the skill, disagreements resolved by tuning rubric WORDING (never by relabeling their ground truth), every resolved disagreement kept as a regression row. Binary verdicts only; the guide contains the measured reason a 1–5 quality score hides failures.

What this skill refuses to do

• Grade severity from tone, emoji, or confidence of the author.
• Emit a claim without a citation, or a portfolio fact (date, cost, count, percentage, or any figure about a program's state) that appears in no update. Numbers the skill itself computes — grades, N-of-M coverage arithmetic — are exempt, and checked instead by the judge's COMPLETE criterion.
• Ship an evidence quote that doesn't appear verbatim in its source update — that row is downgraded to unverified, visibly.
• Average judge criteria into a score, or let one strong section excuse a failed check.
• Drop an update (or a failing claim) without listing it as excluded.