outputs.md

Back to the README · what each of the four verbs leaves on disk.

Outputs available

What a run actually leaves on disk, per verb.

/pm-requirements-v1 · the handoff package

Two halves in one directory. The record answers why should anyone believe this? The seam kit answers what exactly gets built?

Artifact	What it is
HANDOFF.md	the record: seven-part task record, done receipt, assumptions and judgment-call log, the open-questions queue with named confirmers, considered-and-deprioritised with reasons, a self-assessment close, the citation map, and a machine-actionable success-metrics section (tool, exact query, window, evaluation dates, thresholds)
spec.md	scope, out of scope, users affected, behaviour change, data touched, interfaces touched, risks, rollout notes, open items, then the acceptance scenarios
plan.md	how the work is sequenced
data-model.md	entities, fields, states, constraints
contracts/	one file per interface contract
slices.json	the dependency graph of independently shippable units, each carrying the business invariant it must not break
case-contract.md	locked constants, invariants, kill criteria, and the seal numbers in human-readable form
design-gate.json	the approval flag, always emitted approved: false, plus the integrity seal
inference-confirmed.json	the recorded ruling on the open-items queue
retro.md	written at close-out, after the final gate

Phase outputs (p1_research_base.md, p2_strategy_rationale.md, p3_product_spec.md) are canonical too: if context is lost, the run re-grounds from those files rather than from the conversation.

The machine-checkable engineering handoff is optional. With no adapter configured the seam kit is still written, marked as a fixture with no downstream configured, and the package path is printed for a human to move. That is standalone mode, and it is a first-class mode: writing the slice graph is what exposes the invariant nobody had named.

/pm-portfolio-v1 · the batch verdict

Artifact	What it is
verdict document	the batch verdict, the sprint or quarter recommendation, and the primary recommendation with the reason the two obvious alternatives lost
verdict JSON	the same verdict, machine-readable
request-form table	one row per independently decidable ask, with its own stable id, its verbatim gist, both lens scores, evidence-labelled RICE, band and classification
retro.md	what generalises from the run; a run that learned nothing writes one dated line saying exactly that

On the workbook, honestly. A spreadsheet is emitted only when the selected pack declares a workbook_form. When a pack declares none, including the shipped template pack, the verdict table is the deliverable and no spreadsheet is produced. A workbook schema is one organisation's column order, and inventing one for somebody else is worse than handing them a table they can shape themselves.

/pm-verify-release-v1 · the acceptance evidence

Artifact	What it is
acceptance report vN	tier and build stamp, attention header, TL;DR, the verdict split with the open-items queue before and after, the scenario results table, the executed receipts (tool, exact query, window, result against threshold), blockers with owners and unblock criteria, and the acceptance block the owner fills in, left empty by this verb, always
feedback-pack.md	four sections: tag movements, the PARTIAL and GAP register with cause classes, gotchas harvested, gate health. This is the first reading-list input of the next requirements cycle
retro.md	as above

Verdicts are PASS, PARTIAL, GAP or UNVERIFIED. UNVERIFIED is first class and is never forced into one of the others. Reports are versioned, never overwritten.

/pm-gtm-v1 · the plan package

Artifact	What it is
gate-state-snapshot.md	one row per gate: the state the plan assumed, the state now, the receipt, the verdict. Taken fresh before the first planning sentence, and it expires when the run closes
the plan	built phase by phase: intro shell, research, strategy, rollout
campaign architecture	per channel, per persona, gate-aware, with creative-brief pointers
channel rows	channel, objective, persona track, creative-brief pointer, tracking requirement, gate precondition, owner, state
tracking rows	platform, asset, exists-today, configuration target, verification method, state: CONFIG-PLANNED, and what blocks it
handoff record	the package plus the register the decision shape calls for, and the routed technical-candidate list
retro.md	as above

Every row is a configuration proposal. This verb never spends, never posts, never configures a platform and never touches a tag container. There is deliberately no advertising-platform capability anywhere in the package, which is precisely why the rows are proposals.