Imported from haldco-io/strategic-agentic-architecture (plugins/strategic-agentic-architecture/skills/decision-memo/SKILL.md). Install upstream with npx skills add haldco-io/strategic-agentic-architecture --skill decision-memo. Copyright stays with the author.

Decision-Memo

Purpose & position in the engagement

A decision memo exists to get a specific decision made, not to summarize the analysis. The failure that wastes a good engagement is a memo that recommends without a decision ask (the reader finishes unsure what they are saying yes to), advocates without showing the rejected alternatives (a recommendation with no visible options is advocacy, and a competent board rejects it), asserts a return the upstream artifacts do not carry, buries or omits the downside, and offers no owned next step so the decision, even when made, never happens. This skill converts the Storyline and the supporting artifacts into a Decision Memo: the recommendation stated first as a falsifiable claim with its confidence, the exact decision ask, the options and why the recommended one wins against each measured against the do-nothing baseline, the economics honestly ranged, the material risks with their residual and the explicit "what would change our recommendation," the load-bearing assumptions with their validation status, and the owned path after yes.

It is skill 20, the capstone. It consumes the Storyline from narrative-builder (the answer-first structure, governing thought, and MECE support) and every supporting artifact the recommendation rests on: the Option Set (strategic-option-generation), the Business Case (business-case), the Risk Register + Mitigation Plans (risk-mitigation), the Assumption Register (assumption-audit), the Stakeholder Map + Alignment Plan (stakeholder-alignment), and the KPI Tree + Measurement Plan (kpi-architecture). Downstream, the memo is the deliverable the decision-maker acts on; consulting-operating-system (skill 21) routes engagements to it as the terminal step. Because it sits at the end of the chain, the memo is where chain inconsistencies surface — a number in the Storyline that does not match the Business Case, a risk in the register the narrative ignored — and reconciling them before the board sees them is a core, non-optional step of this skill, not a courtesy.

Modes: match the ceremony to the stakes

Pick the mode before starting; a fifteen-page board paper for a $2M call the CEO makes on Thursday is its own failure.

• Full memo (default for board decisions, investment-committee papers, capital allocation, irreversible commitments): all workflow steps (0 through 11), a board-ready memo with the full options comparison, economics, risk, path-after-yes — delivered layered per Step 11: the 2-page plain-language board render on top, the audit layer intact in the appendices.
• Decision brief (a fast decision, a pre-read, or a single gate where the reader needs the recommendation and the ask, not the full apparatus): Steps 0 (recommendation as a claim), 1 (assemble and reconcile — never skipped), 2 (the decision ask), a collapsed options line (name each option and why the recommended one wins in one sentence each), the economics as a range with payback, the single top risk and the "what would change our mind," and the first owned next step. One page. Label it Decision Brief. The brief is itself the reader layer; its reconciliation log and provenance notes ride as an attachment, never on the face. Offer the full memo when the decision is irreversible or the economics land near the hurdle. Mode-adjusted kills in references/red-team.md apply.
• Gate memo (the decision was already made; this is a go/no-go at a pre-committed gate): restate the original ask and the pre-committed kill criteria, report the gate metric against its threshold, and recommend proceed / hold / kill with the arithmetic of the comparison. Do not relitigate the original case; test it against what the gate now knows. Keeps the provenance, reconciliation, and what-would-change-our-mind kills unreduced. Like the brief, the gate memo is itself the reader layer: the recap, the gate-metric-vs-threshold arithmetic, and the proceed/hold/kill land on the face in plain language; the reconciliation log and provenance notes ride as an attachment, never on the face.

The reconciliation step (Step 1), the decision-ask precision, the options-visible requirement, the provenance discipline, and the "what would change our recommendation" honesty apply in every mode — no fast path escapes them, because a memo that hides the alternatives or the downside fails whether it is one page or fifteen.

Inputs required

Establish six things before drafting a line (plus a seventh when the recommended move was contested and war-gamed). If any is missing, elicit it or draft-and-label — a memo that does not know the decision it serves recommends into a vacuum.

1. The decision and the decision-maker. Exactly what will be approved or killed, who has the authority to say yes (a named person or body, not "leadership"), by when, and to what evidence bar (directional / investment-grade / board-defensible). If the decision is fuzzy, the memo cannot state an ask; elicit it before writing. Route back to problem-framing if no decision has ever been defined.
2. The Storyline (from narrative-builder): the governing thought, the MECE supporting arguments, the answer-first structure. This is the memo's spine. If absent, the memo has no argument to carry; run narrative-builder first, or — because a missing upstream ARTIFACT is the §10 degraded mode, not a missing client fact — draft a minimal governing thought as a proxy and label it, and every claim resting on it, DEGRADED — gap: 19 (narrative-builder) (this exact string; the red-team gate string-matches it). Spawn the re-anchor obligation with the flag: the memo owner is the named re-anchor owner, and when narrative-builder later runs, that owner reconciles this proxy governing thought against the real Storyline in a governed restatement (proxy and restated wording shown side by side, arithmetic reproducible) and amends the pointer in every superseded section, so no board reader meets a stale governing thought unwarned; if narrative-builder never runs, the obligation closes as "producer never ran — proxy remains labeled." The retired "unvalidated input" label is not used. (A missing client FACT — a number the Storyline would have carried, not the Storyline itself — is the OTHER degraded mode: estimate and label it "model estimate — unverified," not DEGRADED.)
3. The Option Set (from strategic-option-generation): the real options considered, including Option 0 (the momentum / do-nothing baseline). A memo with no visible alternatives is advocacy; if only one plan was ever generated, route to strategic-option-generation before writing.
4. The Business Case (from business-case): the NPV range, EV, payback, IRR, the load-bearing driver and its breakeven, and the staged commitment. The economics are traced here, never re-asserted. If absent, elicit at least the investment, the value range, and the payback and draft them as a proxy — a missing upstream ARTIFACT is the §10 degraded mode, so label the proxy economics and every figure resting on them DEGRADED — gap: 10 (business-case), with the memo owner as the named re-anchor owner (reconcile against the real Business Case when it runs; if it never runs, close as "producer never ran — proxy remains labeled"). A specific client FACT missing from an otherwise-present case is the OTHER mode: estimate and label it "model estimate — unverified."
5. The Risk Register + Mitigation Plans (from risk-mitigation) and the Assumption Register (from assumption-audit): the material risks with residual and reserve, and the load-bearing assumptions with their validation status. These are where the downside and the evidence quality come from; a memo without them hides what could go wrong.
6. The Stakeholder Map + Alignment Plan (from stakeholder-alignment) and the KPI Tree + Measurement Plan (from kpi-architecture): who decides and what they will object to, and how success will be measured after yes. If the Stakeholder Map is absent, at minimum name the decision-maker and the one objection most likely to sink the memo. The Stakeholder Map must also carry the decision-maker's residual-risk tolerance — the aggregate residual the decision can absorb — because the reconciliation step tests the recommendation's residual against it, and a tolerance that traces to no artifact is a fabricated bar that passes its own check by construction. Source it from the Stakeholder Map (stakeholder-alignment elicits it); when it was never elicited, do not invent one silently — label it explicitly with skill 18's form, "not elicited — assumed: [bound + basis]," and carry that label into the reconciliation log to validate with the decision-maker. The Alignment Plan may also hand over an embargo-dependent flag (skill 18's form): a stakeholder "yes" that holds only until an information embargo lifts — a pending earnings release, an unannounced deal, a regulatory ruling not yet public. Ingest it: the affected recommendation or approval is recorded conditional-on-embargo-lift, and the memo names the embargo-lift date that skill 18 registered with skill 03 as a re-validation trigger (cite it — provisionally as "candidate [embargo-lift trigger], handed in Alignment Plan, row ID pending" until absorbed, then by row ID; skill 03 is the sole writer, so the memo cites, never re-writes the register). This is information-release conditionality, a separate mechanism from the residual-risk tolerance label above — keep both; do not conflate them. Where the embargo-lift date itself was never pinned, label the date with that same tolerance form ("not elicited — assumed: [bound + basis]"), never a new label.
7. The War Game Readout (from war-gaming) — required only when the recommended move is one a competitor could contest and it was war-gamed: the competitive-response evidence behind the recommendation (the equilibrium, whether the move still creates value after the competitor's best response, which options survived the counter). It often arrives folded into the Storyline; when narrative-builder did not carry it in, pull it directly, because a recommendation whose value depends on surviving a competitor's counter must show the counter was played out, not assumed away. If the move was contested but never war-gamed, say so and label the competitive-response case DEGRADED — gap: 16 (war-gaming) (the missing-upstream-artifact mode), with the memo owner owning the re-anchor when war-gaming runs.

Elicitation batch when inputs are thin: see references/templates.md Template 4. When a supporting artifact is missing, the memo may still be written on a drafted, labeled minimal version — but it states, in a provenance line, which inputs are drafted proxies (each carrying its DEGRADED — gap: NN (skill-name) flag) and which are model estimates ("model estimate — unverified"), because a board that learns after yes that the economics were never modeled does not extend a second decision. The retired "unvalidated input" label is not used anywhere in this skill; a missing upstream artifact is DEGRADED, a missing client fact is a model estimate.

Workflow

Step 0 — State the recommendation as a day-1 falsifiable claim

Before assembling anything, write the recommendation the way it will lead the memo: a single claim naming the option, the ask, and the confidence, with the specific evidence that would change it. Format: "We recommend [the decision-maker] approve [Option X — the specific commitment], because [the one governing reason]. Confidence: [level]. We would change this recommendation if [specific, observable evidence — the kill]." This is the BLUF the memo opens with and the falsification device Step 10 attacks. A recommendation nobody could disagree with ("pursue value-creating growth") is an automatic kill, because a claim with no conceivable counter-case recommends nothing. When the recommendation is a conditional posture, the kill is rarely one parameter — it is a kill-tree with multiple branches; Step 0 records the branches, and Step 6 will hold their single canonical in-memo statement (the flip block, authored board-plain) that the BLUF then quotes verbatim. Output: one recorded recommendation-claim with a named kill condition. Quality check: the kill condition names evidence a real gate or analysis could return, not "if it turns out badly."

Step 1 — Assemble the chain and reconcile inconsistencies (the reconciliation step)

Pull each piece of the memo from its named source artifact, then reconcile the pieces against each other before the board sees them, because the memo is the first place the whole chain is read together and a contradiction the board catches destroys the memo's credibility in one question. Pull:

Then run the reconciliation cross-checks and resolve every conflict, do not paper over it:

• Number agreement. The value in the Storyline, the NPV in the Business Case, and the value-at-risk in the Risk Register must be the same numbers. If the narrative says "clears the hurdle comfortably" while the Business Case shows a thin, terminal-value-dependent NPV, the narrative overclaims — correct the memo to the Business Case, not the reverse.
• Recommendation vs. evidence. The option the Storyline recommends must be the one the Business Case values highest against the baseline — and the highest-value check must run on ONE stated basis: rank base-case NPVs against base-case NPVs, or EVs against EVs, but never a ranged EV silently against point NPVs; when the Business Case carries the alternatives only as base-case point NPVs, state the check as a base-case comparison and confirm the ranking survives on EV too. The recommended option must also be the one whose residual sits inside the decision-maker's risk tolerance, and the reconciliation must cite the tolerance's source: the Stakeholder Map entry that records it, or its explicit "not elicited — assumed: [bound + basis]" label (skill 18's form). A residual tested against a tolerance no artifact carries is reconciliation theater; the bar was invented to be passed. When the tolerance is assumed rather than elicited, the residual-vs-tolerance cross-check is provisional — log it "✓ provisional, tolerance assumed — validate," never a clean pass. If the recommended option is not the highest-value or its residual breaches the sourced tolerance, the chain disagrees with itself; resolve before drafting.
• Assumption vs. claim. Every load-bearing number in the recommendation must be at least as well-validated as the Assumption Register says it is. If the Business Case base case assumes a volume the Assumption Register has only partly validated, the memo presents the recommendation on the validated-plus-gated basis, not the optimistic base.
• Competitive response (contested moves). When the recommended move was war-gamed, the option the Storyline recommends must be one the War Game Readout shows still creates value after the competitor's best response — not one that dies on the counter. If the readout's surviving-option set excludes the recommended option, the chain disagrees with itself; resolve before drafting. If the move was contested but never war-gamed, flag the competitive-response case DEGRADED — gap: 16 (war-gaming) rather than presenting it as settled.
• Embargo-dependent alignment (information-release conditionality). If the Alignment Plan (skill 18) carries an embargo-dependent flag — a stakeholder "yes" that holds only until an information embargo lifts (a pending earnings release, an unannounced deal, a regulatory ruling not yet public) — the memo must surface the affected recommendation or approval as conditional-on-embargo-lift, not settled, and name the embargo-lift date skill 18 registered with the Assumption Register (skill 03) as a re-validation trigger. Cite that trigger, do not re-write the register: skill 03 is the sole writer, so cite the row by ID once absorbed, or provisionally as "candidate [embargo-lift trigger], handed in Alignment Plan, row ID pending" until then. Keep this distinct from the analytical flip-block triggers (volume, price): those are read at a gate, this reverses on information release. It is also a separate mechanism from the residual-risk tolerance label "not elicited — assumed: [bound + basis]" — do not conflate them. An embargo-dependent flag the Alignment Plan carries but the memo does not surface is an automatic kill.
• Calendar feasibility of every embedded gate — two prongs, timing AND cleanliness. For every gate, tripwire, or checkpoint criterion the memo embeds (the staged release, the flip conditions, the path-after-yes gate): does the evidence physically exist by the gate date, and is the read uncontaminated? Test it on the anchored, additive formula (§12), not a factor product: readable(N) = [first clean post-event period close] + (N−1) × cadence + latency ≤ gate date, where a clean read is a full measurement period that (i) begins after the event being measured — a period straddling the event is contaminated, not clean — AND (ii) is free of any known one-time exogenous event landing inside the measurement window that would materially distort the metric being read (a mega-event flattering sell-through, a stimulus, a one-off duty recovery, a launch-timing spike). A period correctly timed but containing such an event is contaminated too, and the disposition is to adjust the read for the event, require a confirming clean period before the criterion fires, or explicitly discount the print — never to treat the contaminated period as the clean signal. A windowed trigger ("two consecutive quarters down") adds its worst-case detection lag. The check names any exogenous event inside a gate's window. The memo is the LAST artifact before the board signs, so this is the last chance to catch a criterion that cannot be read — for timing OR for cleanliness — by its own deadline. An unreadable-by-timing criterion is a design defect with FOUR dispositions — re-date the gate, reduce the read-count (state the downgrade), substitute a faster proxy (state the downgrade), or, when the date is immovable and no proxy exists, restructure the ask itself (shrink the tranche to what an evidence-free decision justifies, or present the commitment as explicitly unstaged and labeled) — routed back to the artifact that set it (business-case for restaging); a contaminated read is fixed the same way at the source (adjust, add a confirming clean period, or discount the print); never a caveat. An infeasible or contaminated gate shipped in the ask is an automatic kill.
• Currency and horizon. All values on the same discounting, horizon, and units. A mismatch here is the quiet error that makes two true numbers contradict.
• Register absorption and citation match (the terminal-artifact check) — value AND subject. The memo is the chain's terminal artifact; when consulting-operating-system (skill 21) does not run its Gate D, this reconciliation owns the §11 register check. Verify four things: (a) the Assumption Register absorbed every threshold, flip point, kill criterion, and convention candidate the chain handed in — none is still cited by an ad-hoc name; (b) every memo citation quotes its register row's operative value and matches it by row ID (a value mismatch is then mechanically detectable); (c) the cited row's SUBJECT actually supports the claim it is attached to — inspect the subject, not just the number, because a row cited for a proposition it does not concern is a MISLABEL (e.g. citing the row that carries the outreach-volume breakeven as the source for the price floor, or a cost-base row as the source for a demand claim), a kill in its own right and distinct from citing a superseded row: the value can match to the digit and the citation still be wrong because the row is about something else; (d) the flip-block rows carry their register IDs in the AUDIT layer only, never above the appendix line. The flip block is not a rival canonical home — it is the register's flip/threshold rows stated once inside the memo. An unabsorbed candidate cited by ad-hoc name, a quoted value that does not match its row, or a row whose subject does not support the claim it is attached to (a mislabel), is an automatic kill.

Record each conflict found and its resolution in the reconciliation log (Template 3). Output: the assembled elements, each tagged with its source, and a reconciliation log naming every inconsistency and how it was resolved. Quality check: point at any number in the memo and name the artifact it came from; if two artifacts disagreed, the log shows which one governed and why. An unreconciled contradiction is an automatic kill.

Step 2 — State the decision ask with precision

Write, in one bounded block, exactly what the decision-maker is being asked to approve, because a memo where the reader finishes unsure what they are saying yes to has failed regardless of how good the analysis is. The ask names five things: what (the specific commitment — the amount, the mandate, the scope), who approves (the named authority), by when, the staged structure (what releases now vs. at a gate, if the commitment is staged), and what a yes authorizes (and, as important, what it does not — a yes to a $9M design phase is not a yes to the $52M build). Output: the decision ask as a standalone block a reader could act on without the rest of the memo. Quality check: read only the ask and answer "if I sign this, what have I committed to, and what have I not?" If either is unclear, rewrite it.

Step 3 — Lead with the answer: BLUF and the governing thought

Structure the memo answer-first from the Storyline, because a decision-maker reads the first paragraph and decides whether to trust the rest. Open with the recommendation-claim from Step 0 (bottom line up front), then the governing thought — the single sentence that, if accepted, makes the recommendation follow — then the two-to-four MECE supporting arguments that carry it. Depth comes after, never before. Do not open with context, method, or a chronology of the analysis; the reader wants the answer and will descend into support only if the answer earns it. Output: the memo's opening — recommendation, confidence, governing thought, MECE support — in one page or less. Quality check: a reader who stops after the first page knows what is recommended, what it costs at a headline level, and why; the rest is defense, not news. Where the headline value is soft (majority terminal-value/arbitrage), the opening leads with the hard-vs-soft composition from Step 5, never a single soft aggregate presented as "what your yes buys."

Step 4 — Show the options and why the recommended one wins

Present the real options considered, including the do-nothing baseline, and show why the recommended one beats each, because a recommendation with no visible rejected alternatives reads as advocacy and a good board rejects it on sight. For each option from the Option Set, state what it is, its value against Option 0, and the specific reason it loses to the recommended option (lower return, unacceptable risk, wrong capability, larger outlay for less). Measure everything against the do-nothing baseline (Option 0), not against today, because the momentum case — often a decline the client is already living — is what makes the recommended option's value real. Show the rejected options fairly, not as strawmen; a board trusts a memo that steelmans the alternative it rejected. Output: an options comparison (Template 2) with each option's value-vs-baseline and the reason it was not chosen. Quality check: the rejected options are ones a competent incumbent would actually consider, and each loses for a stated, quantified reason, not "it wasn't strategic."

Step 5 — Carry the economics, honestly

State the value, the investment, the return, and the key sensitivity, traced to the Business Case and never re-asserted or rounded up. Give the value as a range (downside / base / upside NPV with the expected value), not a point, because a single deterministic number is false confidence and a board that has seen one bad point estimate distrusts all of them. State the investment (and its staged structure), the payback, and — critically — the one driver the case lives or dies on with how far the base case sits from its breakeven, so the reader knows the single number to watch.

Split the headline into hard vs. soft — do not front a single aggregate. A headline value is dishonest when it is a single number that is mostly perpetuity, terminal value, or arbitrage/undervaluation, because "worth about +$X" then reads as money the yes buys when most of it is a fragile bet the board never consciously took. Decompose the headline into its hard and soft parts and LEAD with the composition, not the aggregate: the hard, near-term, non-perpetuity, non-arbitrage value (what the move is worth if the fragile driver never delivers past what is already proven) stated first, then + $X of perpetuity/terminal value on [the fragile driver], then + $Y of undervaluation/arbitrage bet where one exists. When the hard part is roughly breakeven and the aggregate is soft, say so plainly — and then carry the reversibility/optionality argument as the actual basis for the recommendation: the recommendation does not rest on the soft aggregate being real, it rests on the ask being reversible (the cheap first tranche prices the soft bet, the large tranche commits only if the gate clears — the board is buying an option on the soft value, not underwriting it). A headline that leads with a single aggregate value that is majority terminal-value/arbitrage without decomposing hard vs. soft is an automatic kill (Step 11's render and Step 3's BLUF carry the same decomposition, never a re-aggregated single number). If the case is thin or terminal-value-dependent, this composition IS how you say so.

Output: the economics block — investment, NPV range with EV, payback, load-bearing driver and its cushion, and the hard/soft decomposition of the headline — every figure tagged to the Business Case. Quality check: every number reconciles to the Business Case (Step 1); the memo presents the risk-adjusted or gated figure where the base case overstates validated confidence; and the headline leads with the hard-vs-soft composition, with the recommendation's basis being the reversibility/optionality where the aggregate is soft, never a bare aggregate.

Step 6 — State the risks and what would change the recommendation

Present the material risks from the Risk Register with their residual after mitigation and the reserve ask, then make the falsification executive: state plainly what would change our recommendation, because a memo that hides the downside gets one shot with the board and loses trust when the downside arrives unannounced. Give the material few (not the full register), each with its residual and its tripwire, and then the honest kill condition from Step 0, tied to a real observable: "If [the load-bearing driver] comes in below [value] at [the gate], we would not release the main tranche and would revert to [the fallback option]." This is the "what would change our mind" the board most wants and most memos omit.

When the posture is conditional, the flip machinery is rarely one numeric parameter and one flip — it is a kill-TREE: multiple branches, conjunctions, different actions per branch. Sanction the shape, but discipline the statement: the flip conditions are stated ONCE, in canonical form — the flip block, a small table or tree of condition → threshold → action. Write the block in BOARD-PLAIN LANGUAGE here, by construction, so it needs no later translation: because the canonical block is already plain, every other section (BLUF, board render, ask, path after yes) QUOTES it verbatim or points to it ("per the flip block"), and NONE re-words it. There is no licence to re-state the conditions "in the reader's language," because the reader's language is already the canonical wording. Re-worded restatements drift: a threshold softens in the summary, a price qualifier inverts in the risk section, and the board reads two different kill criteria in one memo. The block is subordinate to the Assumption Register — it is the register's flip/threshold rows stated once inside the memo (not a rival canonical home), and its rows carry their register IDs in the audit layer only. Flip conditions stated in more than one wording is an automatic kill.

Information-release conditionality is a distinct branch of the conditional posture, not an analytical flip-block trigger. When the Alignment Plan hands over an embargo-dependent flag (Step 1's reconciliation, skill 18's form), state — as its own board-plain condition, kept separate from the flip block's volume/price triggers — that the affected approval holds only on the currently public information and reverses (or must be re-validated) when the embargo lifts, naming the embargo-lift date registered with the Assumption Register as a re-validation trigger (cited provisionally until absorbed; skill 03 is sole writer). The distinction is load-bearing: the flip block's triggers are things a named person READS at a named gate; the embargo condition is an INFORMATION-RELEASE event that flips a stakeholder yes when news breaks. Discipline it exactly as the flip block — state it once, board-plain, and every later mention (BLUF, board render) quotes it verbatim or points to it, never re-words it. It is a separate mechanism from the residual-risk tolerance label "not elicited — assumed: [bound + basis]"; both can be present; do not conflate them.

Output: the material risks with residual and reserve, the canonical flip block, and — when the Alignment Plan carries one — the embargo-dependent condition surfaced as a distinct information-release branch. Quality check: each flip condition is specific enough that a named person could observe it at a named gate and act; the residual risk traces to the Risk Register, not a fresh guess; the flip conditions exist in exactly one wording — board-plain by construction — with every other mention a pointer or a verbatim quote, never a re-wording; and any embargo-dependent flag from the Alignment Plan is surfaced as conditional-on-embargo-lift (distinct from the analytical triggers), not presented as settled.

Step 7 — State the assumptions and the evidence quality

Surface the load-bearing assumptions from the Assumption Register with their validation status, and state the evidence bar the memo meets against the one the decision-maker requires, because a board-defensible decision made on directional evidence is a decision made on sand. Name the two or three assumptions the recommendation rests on, each with its status (validated / partially validated / unvalidated) and its evidence grade, and state honestly whether that clears the decision-maker's bar. Where it does not, do not hide the gap — make it the reason for the staged structure (approve the phase that buys the missing evidence, gate the rest). Output: the load-bearing assumptions with validation status, and an explicit "evidence bar met vs. required" statement. Quality check: if the recommendation rests on an unvalidated assumption, the memo says so and the decision ask is structured to reflect it, not to obscure it.

Step 8 — Specify the path after yes

Define what happens the moment the decision-maker says yes, because a decision with no owned next step does not happen. Name the decision rights (who now has authority to act on what), the immediate next steps (the first two or three, with owners and dates), the first gate or milestone and its go/no-go criteria (from the Business Case's staged kills), the named owner of execution (a person, not a function), and how success will be measured (the leading and outcome metrics from the KPI Tree). Output: the path-after-yes block — decision rights, first steps with owners, first gate with criteria, execution owner, success metrics. Quality check: a named person owns the first step with a date, and the first gate has a pass/fail criterion tied to the load-bearing driver; a path that says "the team will develop a plan" is not a path.

Step 9 — Frame for the decision-maker and pre-empt objections

Tune the memo to the actual decision-maker and address the known objections from the Stakeholder Map, without concealment, because a memo that ignores the CFO's known scar from the last failed program invites the objection to arrive as a surprise in the room. For each material objection the Stakeholder Map records, state it fairly and answer it in the memo body (not in an appendix nobody reads), using the recommendation's actual structure — the staged gate answers the capital-risk objection, the change plan answers the disruption objection. Framing is not spin: never suppress a risk or an alternative to make the memo land; a board that later finds a concealed downside treats every future memo as suspect. Output: the known objections named and answered inside the memo. Quality check: the strongest stakeholder objection has a fair statement and a real answer in the memo, and no objection was answered by hiding the fact that raised it.

Step 10 — Falsification pass: the strongest case against

Before delivering, argue the recommendation down, because a memo only its author could believe fails the board and the pre-mortem. Do three things with real content, not straw men:

1. State the strongest honest case against the recommendation — the world in which the rejected option was right, or in which no should have been the answer — and then state why the recommendation still holds, or the specific conditions under which it would not.
2. Address the dissent, do not suppress it. If a stakeholder, an analyst, or an artifact disagrees with the recommendation, put the disagreement in the memo and answer it. A suppressed dissent that surfaces in the room is fatal; an addressed dissent builds trust.
3. Confirm the recommendation survives its own kill condition today. Re-read the Step 0 kill: is any part of it already true on the current evidence? If yes, the recommendation is not ready — revise before delivering, do not footnote.

If any attack materially weakens the recommendation, revise the memo before delivering; only residual, non-material doubts survive into the memo's honesty section. Output: the strongest case against, the dissent addressed, and confirmation the kill condition is not already met. Quality check: the case against is one a hostile board member would actually make, and the memo answers it on the merits.

Step 11 — The board-render pass: separate the board-facing document from the audit layer, THROUGHOUT

The memo that survives Steps 0–10 is, by construction, an audit object — reconciliation logs, provenance tags on every figure, gate records, basis notes, register IDs woven through the body — and an audit object is a mediocre reading experience. The failure this step exists to kill is a BIMODAL instrument: a clean 2-page render bolted onto a body that got HEAVIER with audit scaffolding, so a director who reads past page 2 hits an avalanche of internal machinery ("R-2 reconciliation-log," a "close-leg sweep," a "C1–C8 convention table," "DEGRADED — gap: 04," "candidate … row ID pending," "readable(1) ≈ Sept 30," "not elicited — assumed," a runtime red-team-gate paragraph) that reads as a measurement program, not a decision. Separating the layers on page 1 alone is not enough. Do not delete any of the machinery; LAYER the WHOLE document.

The line is between the board-facing document and the audit layer, and it runs the full length of the memo:

• The board-facing document = everything a director reads: the 2-page render on top PLUS the rationale below it (the options, the economics, the risks, the assumptions, the path after yes, the honest case against). ALL of it is written board-plain, end to end — not just the top two pages. No audit-layer notation anywhere above the appendix line: no R-IDs or reconciliation-log labels, no C-IDs or convention-table IDs, no register row IDs (AR-nn), no DEGRADED — gap: NN strings, no provisional-citation strings ("candidate … row ID pending"), no readable(N) formulas or calendar arithmetic, no "not elicited — assumed" labels, no red-team-gate / "why this passes its own gates" prose, no skill names, mode labels, or bracket tags ([RCA-U]). If a director would read a phrase aloud with a raised eyebrow, it belongs below the line. The board-facing rationale conveys the SUBSTANCE of all of that in plain language (the downside, the validation gap, the flip conditions, the reversibility case, an honest "these economics were drafted, not modeled — validate before the build commits" where a proxy was used), but never the notation.
• The audit layer = everything below the appendix line, mandatory, intact, findable: the reconciliation log (R-IDs), the provenance ledger mapping every board-facing claim to its source artifact, register extracts (AR-nn rows, the flip-block row IDs, absorbed candidates and the sanctioned "candidate … row ID pending" provisional-citation strings), the DEGRADED — gap: NN (skill-name) flags with their named re-anchor owners, the readable(N) calendar-feasibility arithmetic (including the cleanliness check for exogenous events), gate records, basis notes, and the gate-compliance record. The traceability is NON-NEGOTIABLE and complete — it just lives below the line. The canonical flag strings the earlier gates string-match (DEGRADED — gap: NN, "candidate … row ID pending," AR-nn) still EXIST, in the audit layer, so §10/§11 are satisfied there while the board face stays plain.
• The fidelity gate (board-facing vs. audit): this is the LAST pass and it rewrites the reader-facing layer, so readability is not the only test — truth is. Every claim and every number in the board-facing document (render AND rationale) must be ENTAILED by a specific figure in the audit layer, checked claim by claim (the diff discipline of the flip-block rule, extended to the whole board-facing document). A plain-language gloss may not assert what the audit layer does not support: "evidence grade C against a required B" does not become "the evidence is nearly there"; an EV of +$28.6M does not round to "about $30M"; "half the demand the case plans on is signed" is legal only because 0.9M is 50% of the 1.8M plan — mutating "plan" into "need" (0.9M is 64% of the 1.4M breakeven the case needs) is truth drift and an automatic kill. Point at each board-facing sentence and name the audit-layer figure that entails it; a sentence that entails nothing, or overstates what it entails, is rewritten or cut.

The quality bar: a non-technical director gets the decision in five minutes from the top layer alone, and can read the entire board-facing document — render and rationale — without meeting a single piece of audit-layer notation. "Five minutes" is not itself falsifiable, so test the render by its proxies — all five enumerated elements present and readable (the ask, the recommendation, the three numbers as the hard/soft composition, the flip conditions, the boundary of the yes), the top layer at or under two pages, and no notation above the appendix line. And test the whole board-facing document by a second proxy: scan every line above the appendix line for the banned notation classes (R-IDs, C-IDs, AR-nn, DEGRADED-gap strings, "candidate … row ID pending," readable(N), "not elicited — assumed," red-team-gate prose) — one hit above the line fails. The audit layer remains mandatory and complete — this step is about layering, not deletion; a memo that deletes its scaffolding to read well fails exactly as one that shows it all on the face. Output: the two-layer deliverable — board-facing document (render + plain rationale) on top, audit appendices below (Template 7). Quality check: read the whole board-facing document as a cold, non-technical director — the five render elements land, the render is ≤2 pages, nothing above the appendix line requires the engagement's internal machinery or carries its notation, and every board-facing claim traces to an audit-layer figure that entails it.

Quality bar

The Decision Memo is ready only when all of these hold:

• The recommendation leads the memo as a falsifiable claim with a stated confidence and a named condition that would change it.
• The decision ask names what is approved, by whom, by when, the staged structure, and what a yes does and does not authorize — a reader knows exactly what they are signing.
• The real options are visible, including the do-nothing baseline, and the recommended one beats each for a stated, quantified reason measured against Option 0, not today — the highest-value check stated on one named basis (base-case NPVs, or EVs), never a ranged EV silently ranked against point NPVs.
• The economics are a range with EV, traced to the Business Case, with the investment, payback, and the load-bearing driver and its breakeven cushion stated; no number is re-asserted or rounded up from its source.
• The headline value leads with its hard-vs-soft composition, never a single aggregate: the hard, near-term, non-perpetuity, non-arbitrage value stated first, then the perpetuity/terminal value on the fragile driver, then any undervaluation/arbitrage bet. Where the hard part is roughly breakeven and the aggregate is soft, the recommendation's stated basis is the reversibility/optionality of the ask (the cheap tranche prices the soft bet; the large tranche commits only if the gate clears), not the soft aggregate being real. A headline that fronts a single aggregate value that is majority terminal-value/arbitrage without decomposing hard vs. soft fails.
• The material risks carry their residual and the reserve ask from the Risk Register, and the memo states explicitly what would change the recommendation, tied to an observable at a named gate; the flip conditions exist in exactly one canonical wording — the flip block, written board-plain by construction (condition → threshold → action) — and every other mention, the board render included, quotes it verbatim or points to it, never re-words it. The flip block is subordinate to the Assumption Register (the register's flip/threshold rows stated once in the memo; its rows carry register IDs in the audit layer only), not a rival canonical home.
• The residual is tested against a risk tolerance that traces to the Stakeholder Map or carries the explicit "not elicited — assumed: [bound + basis]" label (skill 18's form) — never against a bar no artifact carries; when the tolerance is assumed rather than elicited, the residual-vs-tolerance check is marked provisional, not a clean pass.
• When the Alignment Plan carries an embargo-dependent flag, the memo surfaces the affected approval as conditional-on-embargo-lift in the conditional-posture section — a distinct information-release branch, kept separate from the analytical flip-block triggers — names the embargo-lift date as a re-validation trigger registered with the Assumption Register (cited provisionally until absorbed; skill 03 is sole writer), and carries it into the board render so the board sees the yes reverses when the embargo lifts. It is a separate mechanism from the residual-risk tolerance label, not a substitute for it.
• Every gate, tripwire, or checkpoint criterion the memo embeds passes the calendar-feasibility check on both prongs — timing AND cleanliness — on the anchored additive formula (readable(N) = first clean post-event period close + (N−1) × cadence + latency ≤ gate date), where a clean read both begins after the event being measured AND is free of any one-time exogenous event landing inside its window that would materially distort the metric. An unreadable-by-timing criterion is fixed at the source by one of four dispositions (re-date, reduce read-count, faster proxy, or — date immovable and no proxy — restructure the ask and route to business-case); a contaminated read is adjusted for the event, deferred to a confirming clean period, or explicitly discounted at the source; neither is shipped as a caveat.
• Because the memo is the terminal artifact, its reconciliation runs the §11 register check when skill 21's Gate D does not: the Assumption Register absorbed every threshold, flip, and convention candidate (none cited by ad-hoc name); every memo citation quotes its register row's operative value and matches it by row ID; AND the cited row's subject actually supports the claim it is attached to (a row cited for a proposition it does not concern is a mislabel — a kill distinct from citing a superseded row — caught by inspecting the row's subject, not just its value).
• The load-bearing assumptions carry their validation status, and the memo states honestly whether the evidence bar met clears the decision-maker's required bar.
• The path after yes names decision rights, the first steps with owners and dates, the first gate with go/no-go criteria, the execution owner, and the success metrics.
• The known stakeholder objections are stated fairly and answered inside the memo, without suppressing the fact that raised them.
• The strongest honest case against the recommendation is in the memo, the dissent is addressed not hidden, and the recommendation survives its own kill condition on today's evidence.
• Every number and claim traces to a named source artifact or is labeled an assumption; the reconciliation log records every chain inconsistency found and how it was resolved. A cold reader can trace any figure in the memo back to the artifact it came from and find no two artifacts contradicting.
• The board render passes the five-minute test, tested by its proxies: a non-technical director gets the decision in five minutes from the top layer alone — the five enumerated elements (the ask, the recommendation, the three numbers that matter presented as the hard/soft composition, the flip conditions, the boundary of the yes) all present and readable, the top layer ≤2 pages, and no internal label or chain shorthand above the appendix line — with all audit scaffolding (reconciliation log, provenance ledger, gate records, basis notes) intact below it. Every claim in the top layer is entailed by a specific audit-layer figure, checked claim by claim; a plain-language gloss that asserts what the audit layer does not support is truth drift and fails.
• The board-facing document is board-plain END TO END, not just on page 1: the render AND the rationale below it (options, economics, risks, assumptions, path after yes, honest case against) read plain, and no audit-layer notation appears anywhere above the appendix line — no R-IDs or reconciliation-log labels, no C-IDs or convention-table IDs, no register row IDs (AR-nn), no DEGRADED — gap: NN strings, no provisional-citation strings ("candidate … row ID pending"), no readable(N) formulas, no "not elicited — assumed" labels, no red-team-gate / "why this passes its own gates" prose, no skill names or mode labels. The audit trail still EXISTS and is complete below the line (the canonical flag strings the earlier gates string-match live there); the board never has to open it to get the decision. One piece of audit notation above the appendix line fails.

Output format

Default deliverable is the Decision Memo — full template in references/templates.md. Structure: (1) the recommendation and the decision ask, up front (BLUF); (2) the governing thought and MECE support from the Storyline; (3) the options and why the recommended one wins vs. the baseline; (4) the economics, ranged and traced; (5) the risks, residual, and the canonical flip block; (6) the load-bearing assumptions and evidence quality; (7) the path after yes — decision rights, next steps, first gate, owner, metrics; (8) the honest case against and the dissent addressed; (9) the appendices — source artifacts, the reconciliation log, the provenance ledger, gate records, basis notes. Lead with the ask and the answer, never with context or method. Deliver it layered per Step 11: the board-facing document — the 2-page plain-language render on top (Template 7) and the 2–4-page rationale below it — is board-plain END TO END, and the audit layer (reconciliation log, provenance ledger, register extracts, gate records, basis notes, and every piece of audit notation: R-IDs, C-IDs, AR-nn rows, DEGRADED-gap flags, "candidate … row ID pending" strings, readable(N) arithmetic) sits complete below the appendix line — moved, never deleted, and never on the board face. In decision-brief mode, deliver the one-page version labeled Decision Brief.

Red-team gate

Before delivering, run the full checklist in references/red-team.md against the draft memo and fix every failure — it is a runtime gate this skill executes every run, not authoring documentation. Automatic kills (deliver nothing that violates these): no clear decision ask, or an ask a reader cannot act on without asking a follow-up; the recommendation not stated as a falsifiable claim with a confidence and a kill condition; no options shown, or the do-nothing baseline absent, or the recommended option not measured against Option 0; a recommendation not traced to the Business Case (economics re-asserted rather than sourced, or a value the Business Case does not carry); risks or the downside buried or absent, or no explicit "what would change our recommendation"; flip conditions stated in more than one wording (the flip block — authored board-plain — is the only statement; every other mention, the board render included, quotes it verbatim or points to it, never re-words it; the block carries its Assumption-Register row IDs in the audit layer and is not a rival canonical home); an embedded gate, tripwire, or checkpoint criterion that cannot be read by its own deadline OR is read from a contaminated window (anchored additive test, two prongs — timing AND cleanliness: readable(N) = first clean post-event period close + (N−1) × cadence + latency vs. the gate date, where a clean read both begins after the event measured AND is free of any one-time exogenous event inside its window that would distort the metric) shipped as a caveat instead of fixed at the source (timing: re-date, reduce read-count, faster proxy, or restructure the ask when the date is immovable and no proxy exists; contamination: adjust the read, defer to a confirming clean period, or discount the print); the headline value fronted as a single aggregate that is majority terminal-value/arbitrage without decomposing hard vs. soft (lead with the hard/near-term value + the perpetuity on the fragile driver + any arbitrage bet, and carry the reversibility/optionality as the recommendation's basis where the aggregate is soft); a top-layer claim not entailed by a specific audit-layer figure (a plain-language gloss asserting what the audit layer does not support — truth drift); the Assumption Register not shown to have absorbed every threshold/flip/convention candidate, or a memo citation whose quoted value does not match its register row by ID, or a citation whose row SUBJECT does not support the claim it is attached to (a mislabel — inspect subject, not just value; a kill distinct from citing a superseded row) (the §11 terminal-artifact check the memo owns when skill 21's Gate D does not run); a residual-risk tolerance used in reconciliation that traces to no artifact and carries no "not elicited — assumed: [bound + basis]" label (skill 18's form); an embargo-dependent flag the Alignment Plan carries but the memo does not surface as conditional-on-embargo-lift (the yes presented as settled when it reverses on information release), or surfaced but with no embargo-lift date registered with the Assumption Register as a re-validation trigger (cited provisionally until absorbed; skill 03 is sole writer); a missing upstream artifact (the Storyline, the Business Case, the War Game Readout) drafted as a proxy but not carrying the exact canonical flag DEGRADED — gap: NN (skill-name) with a named re-anchor owner, or carrying the retired "unvalidated input" label instead; the load-bearing assumptions' validation status omitted, or the evidence bar met/required comparison absent; no owned path after yes (a first step with a named owner and date, and a first gate with criteria); hedging that obscures the recommendation (a memo that recommends three things is recommending nothing); a stakeholder objection or a dissent suppressed to make the memo land; an unreconciled inconsistency between two source artifacts; audit-layer notation anywhere in the board-facing text — render OR the rationale a director reads below it, not just page 1 — where the notation classes are R-IDs / reconciliation-log labels, C-IDs / convention-table IDs, register row IDs (AR-nn), DEGRADED — gap: NN strings, provisional-citation strings ("candidate … row ID pending"), readable(N) formulas, "not elicited — assumed" labels, red-team-gate / "why this passes its own gates" prose, skill names, mode labels, or bracket tags like [RCA-U] (the board-facing document must be board-plain end to end; all of this moves below the appendix line where it still exists and is complete), or the audit layer deleted rather than moved to appendices, or a top layer a non-technical director cannot get the decision from in five minutes (tested by its proxies: the five elements present and readable, ≤2 pages, no notation above the appendix line); unsourced precision anywhere, or any citation naming a source not actually observed this session (numbers recalled from training carry "model estimate — unverified," never a filing/comp/benchmark tag). In decision-brief and gate modes, apply the mode-adjusted kills defined there.

Handoffs

• consulting-operating-system (skill 21) routes an engagement to this memo as its terminal step and, if the decision is deferred pending evidence, re-sequences the chain to close the gap the memo surfaced.
• business-case (skill 10) is re-run when reconciliation (Step 1) shows the recommended option is not the highest-value against the baseline, or when the decision-maker's counter-case moves a driver — do not adjust the memo's number, fix the case and re-pull it. When the calendar-feasibility check finds an embedded gate criterion that cannot be read by its date, route the fix to the artifact that set it (business-case for staged kills, kpi-architecture or transformation-roadmap for measurement gates) — re-date, change the read-count, substitute a faster proxy, or, when the date is immovable and no proxy exists, restructure the ask (shrink the tranche to what an evidence-free decision justifies, or label it explicitly evidence-free) and route to business-case for restaging — at the source; never caveat it in the memo.
• risk-mitigation (skill 15) is re-run when the "what would change our recommendation" condition reveals a material risk the register did not size, or when the decision is staged and the gate needs a tripwire the register lacks.
• stakeholder-alignment (skill 18) is re-run when the falsification pass surfaces an objection the Stakeholder Map did not anticipate, when the memo must be re-tuned for a different decision-maker, or when the decision-maker's residual-risk tolerance was never elicited — the memo can run on a labeled "not elicited — assumed: [bound + basis]" tolerance (skill 18's form, with the residual-vs-tolerance check marked provisional), but skill 18 closes that gap. Skill 18 also hands off the embargo-dependent flag when a stakeholder yes holds only until an information embargo lifts; the memo receives it, records the affected approval as conditional-on-embargo-lift, and cites the embargo-lift date skill 18 registered with skill 03 — re-run skill 18 if the embargo-lift date shifts or a new embargo-dependent yes is discovered after the map was built.
• narrative-builder (skill 19) is re-run when reconciliation shows the Storyline overclaims relative to the artifacts — the argument, not the memo, needs correcting.

Offer the handoff explicitly: name the Decision Memo as the artifact produced and, when a reconciliation surfaced an upstream defect, name the artifact that must be fixed before the memo goes to the board.