Portfolio PDLC — Advance

Outcome

One card measurably closer to its next confidence milestone: the stage's missing work produced, the pending decision packaged for humans, or — when a decision is recorded — the stage transitioned and logged. Advance means growing confidence toward a decision, not pushing cards rightward.

Pick-one rule

Work exactly one card per invocation (the loop's leverage table usually picked it; if called ad-hoc, take the named card). Everything else you notice goes to the loop log or improvements/, not into scope.

Stage-appropriate moves

Card sits in	Advance means	Compose
explore	Frame the problem for real: draft/complete the lightweight canvas; identify the riskiest DVF assumptions; recommend investigate-further / park	`intent-framing-chooser`
discovery	Drive the timebox to a verdict: run/summarize the experiments against the riskiest assumptions, tag results evidence/opinion, produce a proceed / de-risk-further / stop recommendation	`sniff-test` (readiness check)
plan-commit	Build the commit package: Outcome-Oriented Roadmap with an explicit confidence range, resourcing picture, what's still unproven at commit
execute	Keep confidence honest: read leading indicators, check slices are validating the hypothesis, surface drift (scope creep on an "experiment", indicators flat) as a steer-or-escalate note
rollout	Turn "shipped" into "realized": adoption/hypercare status, instrument the outcome measures, draft the Value Realization read
bau	Run the measure-and-learn questions (hypothesis achieved? metrics teach what? diminishing returns? still a constraint?) and recommend keep / enhance / close — plus the baseline-vs-actual watermelon check on exit

Improvement cards advance the same way; their discovery is a probe, a one-review pilot, or a portfolio-pdlc-simulate scenario.

Decision boundaries — prepare, don't cross

Entering plan-commit or execute (and any invest / pivot / kill / freeze call) is human territory:

Write the decision brief to reviews/YYYY-MM-DD-<slug>-decision.md: what you can rely on (evidence-backed), what's still open (opinion-tagged), options framed as double down / de-risk further / kill-pause (or invest/skip-discovery at the earlier boundary), a recommendation with reasoning, and what would make us say no.
Set the card's next_decision (date — decision — owner) and stop there.
Only when a dated human decision exists in the card's Decision log: update stage + stage_entered, re-run the board script (it logs the transition), and confirm the flow-log row appeared.

Overdue-decision cycles (leverage row 1) refresh the brief, restate the cost of waiting (items queued behind it, aging), and surface it — they never pile more work behind the stuck decision.

Rules

Confidence forward, not cards rightward: a proceed/stop recommendation or an honest "stay and de-risk" is a successful advance.
Respect optionality: discovery and rollout are skippable by decision — prepare the skip decision when risk is genuinely low; flag habitual skipping.
Evidence vs opinion tagging in everything you produce; unevidenced confidence stays a question mark in the brief.
Timeboxes are real: a lapsed discovery gets a verdict from what was learned, not an extension by default.
Update the card's Evidence log with what this advance produced.

Quality Gates

Exactly one card touched; its frontmatter, Evidence log, and (if transitioned) flow log all agree.
Any boundary transition has a dated, named Decision-log entry — or didn't happen.
The decision brief (when one was due) states evidence, open risks, options, a recommendation, and "what would make us say no."
Board regenerated after the move.

References

Stage expectations inline above; depth: skills/sniff-test/references/ (this repo).
Umbrella: portfolio-pdlc (leverage table, loop mechanics).