MIZ OKI Media Executive Demo — presenter talk track r1.0

Copied from mizoki-exec-demo-r1.0.zip (docs/, zip sha256 40112985…58bd) on 2026-09-14 for the page served at /media/demo (slug renamed 2026-09-14; /media/executive-demo 308s there). One line trued against canon on landing (Act 2 step 5: the cells are live and observe-only, the qualifier the page's chip carries); everything else is r1.0.

Timing: 20 minutes. Executive cut: Acts 1 + 5 (5 minutes) — press 5-min cut in the top bar and the arrows skip the middle. Keys: ← → acts · 1–6 jump · N talk track · B boardroom type. Pick the tenant scale before you start so the money is in the room's units.

Open every session with one sentence: "Everything on this screen is a synthetic tenant, watermarked — the demo shows you the mechanism, the pilot gives you your own numbers."

Act 1 · The Lie (3 min) — "Is our ROAS real?"

  1. Open on What the platform reports. Ask the CFO which SKU they would scale. (Everyone picks 7742 at 4.2×.)
  2. Switch to What the bank sees. 7742 is −$620 after returns; 5527, the "worst" ROAS on the board, makes money.
  3. Drag the return-rate slider. The 4.2× never moves. The contribution line does. Read the breakeven.
  4. Point at SKU-2200: "We did not invent a cost to finish the arithmetic. That row is excluded and named. That is the discipline you are buying." Closer: We optimize the number the bank account reports.

Act 2 · The Signals (4 min) — "What do you see before the click — and refuse to see?"

  1. Press Play session. Narrate tokens as behavior, never identity: deceleration, dwell, tab-return, a form started and abandoned.
  2. Watch the forecast climb with its interval. Say it once, verbatim: anticipatory intent, never mind-reading. An intent score is never served as a bare number.
  3. Flip consent to analytics-only, play again: no consent, no event. Architecture, not a setting.
  4. The O-1 moment: press Send keystroke-dynamics payload. Let the red card sit for a beat. Then send form_started — accepted.
  5. Label honesty before they ask: cells 33–36 are live and observe-only; the Intent Engine v2 is shadow. Closer: What the system refuses to learn is the strongest thing on this slide.

Act 3 · The Proof (4 min) — "Which conversions did we cause?"

  1. Register holdout on Prospecting. Read the record aloud: estimand, power, hash assignment, before first impression.
  2. Estimate: intervals, never points. Run refutation — it passes. The ledger writes: 610 caused, 390 anticipated; 3.4× becomes iROAS 2.1×.
  3. Switch to Brand search. Platform says 6.8×. Register, refute — data-subset fails, CI crosses zero: WITHHELD. No credit, no move.
  4. Ghost bids button is disabled on purpose: "That is a plan, not a shipped technique. We do not claim it." Closer: Showing our system catch our own weak estimate outsells any passing metric.

Act 4 · The Profit (4 min) — "Where does the next dollar go?"

  1. Waterfall: name every subtraction. Then the excluded discount-liability row — excluded, not estimated.
  2. Dosage curve: drag their daily budget. Read d* = $4,048 with its band. Drag past $9,500 — "no estimate outside support."
  3. Posture: advisory. NET_YIELD_WRITEBACK is off. Nothing here bids. Closer: Your profit-maximizing spend is a number with an interval — not a budget line someone inherited.

Act 5 · The Restraint (4 min) — "What stops it doing something stupid?"

  1. Run +25% budget → audience_hv. Passport fills, policy vetoes, DEL 87.8 — and it is still WITHHELD: a DEL score never overrides a failed hard constraint.
  2. Run +12% budget → campaign_7. Authorized, bounded — and still routed to a named human for spend, then the actuator's independent re-check.
  3. Run Scale spend into SKU-4410. Stockout veto upstream of any adapter, regardless of confidence 0.91.
  4. Run Retarget the hesitation segment. No registered holdout → forbidden. Act 2 and Act 3 meet here.
  5. Governance challenge: drag the floor below 80. It refuses. Tenants raise floors; nobody lowers them — including us.
  6. Restraint Ledger: seven refusals this month, $48,900 not moved. "No competitor can show you their refusals."
  7. Submit reservation: three HALTs on a live F4 lane. Then the deploy-gate story: our own whitepaper was corrected by the gate before it could publish. Closer: Trusted with authority it does not yet have, by demonstrating restraint with the authority it does.

Close · The 90 Days (1 min)

  1. Three gates, three exits. Read the unscored-gate line on purpose — the gate is not merely missed, it cannot be scored yet (zero admissible forward labels), and that is why we observe today. Never quote 0.6884: it was the fallback scorer on one seed (LII_BACKTEST_MODEL_QUALITY_2026-08-16, recommendation 1).
  2. Status table: every label came from the Offering Map, not this deck.
  3. The ask: "Tell us the decision that costs you the most when it goes ungoverned. We scope the two pilot decision classes around it, start observe-only, and put the first registered holdout in front of you by day 31."

Objection handling (canon-safe answers)

← All docsView source on GitHub →