SYNTHETIC TENANT · NOT CUSTOMER DATA

MIZ OKI Media · Causal Growth Control · executive demo

Other tools optimize the number your ad platform reports.
We optimize the number your bank account reports.

Twenty minutes, five acts, one synthetic brand. Each act answers a question a CFO or CMO actually asks, with a control you can touch. Nothing on this screen is a customer result — it is the demo tenant, watermarked, deterministic, and labeled with what is live today and what is still being built.

Net yield and the ORACLE intent surfaces are Preview · in development. Every figure on every act is illustrative — the synthetic tenant demo-fixtures, never a customer result.

The system is designed to be trusted with authority it does not yet have — by demonstrating restraint with the authority it does.

The 90-day pilot ↗

The arc

20 min · exec cut 5 min
  • 01The Lie · Is our ROAS real?Net yield
  • 02The Signals · What do you see before the click — and refuse to see?ORACLE
  • 03The Proof · Which conversions did we cause?Causal
  • 04The Profit · Where does the next dollar go?Dosage
  • 05The Restraint · What stops it doing something stupid?DCP
  • 06The 90 Days · How do we start, and what is real today?Pilot

← → move between acts · N talk track · B boardroom type · 1–6 jump

Act 1 · The Lie · 3 min

Same campaign. 4.2× reported ROAS. Negative contribution after returns.

The systems that spend the money also grade the results. Every platform reports the conversions it believes it caused; none of them subtracts refunds, cost of goods, fees or fulfillment. Automating on top of a self-graded number automates the error.

Net yield is Preview · in development (writeback OFF). Every figure on this stage is illustrative — the synthetic tenant, never a customer result.

“Is our ROAS real?”

Profit-Leak Report · top SKUs

Net yield · in buildbank view
SKUAd spendReported ROASReturnsContribution after all costsResult

Method: revenue − refunds/returns (actuals only until one cycle per SKU) − COGS − payment fees − fulfillment − ad spend. A row missing a required cost is excluded and named; a cost is never invented to complete the arithmetic. Source: net_yield + order_economics · demo-fixtures.

Touch it: the hero SKU

SKU-DEMO-7742
Drag it. The platform's 4.2× never moves; the contribution line does.
Reported ROAS
4.2×
what the dashboard says
Contribution
−$620
breakeven at 44% returns

This month, whole catalog

SKUs analyzed
847
Advertised into a loss
23
Loss after returns
−$14,280
Median return rate
42%
Closer

We don't optimize the number the dashboard reports. We optimize the number the bank account reports — and we show you which SKUs your ads are currently paying to lose money on.

Act 2 · The Signals · 4 min

Anticipatory intent from content-free signal. Never mind-reading.

ORACLE reads how a consented session behaves — viewport deceleration, dwell, tab blur and return, form lifecycle — and forms a calibrated, expiring hypothesis about where that session is heading. It never reads what was typed, said, or looked at. What the system refuses to learn matters as much as what it learns.

Intent Engine v2 is Preview · in development — cells 33–36 are deployed and observe-only; nothing here acts. Every figure on this stage is illustrative.

“What do you see before the click — and what do you refuse to see?”

A consented session, replayed

Cells 33–36 · live · observe-onlyIntent Engine v2 · shadow · in build

    Session outcome forecast

    I-02 · shadow
    0.00 ± —

    Probability of a stage transition inside the session horizon. Served with its uncertainty and an explanation path — never as a bare number.

    stage 01Awareness
    stage 02Consideration
    stage 03In-market
    stage 04Purchase-imminent
    • ·Hesitation headhigh attention, no click — surfaces a posture recommendation, not a push—
    • ·Matched creative rank“technical proof” variant ranked #1 — logged, not applied (flag off)—
    • ·Hypothesis TTLexpires with the session; no cross-session profile is written—

    The O-1 moment

    Privacy lock · live · owner ruling

    Try to give the system something it must not have. This payload is what a keystroke-dynamics vendor would send. Press the button on stage.

    { "signal_type": "keystroke_dynamics",
  "dwell_ms": [112, 98, 143], "flight_ms": [61, 77],
  "cadence_profile": "…", "field": "card_number" }

    Bright lines — rejected at the schema, not in a policy PDF

    • ✕Keystroke dynamicsraw keyboard, entered text, dwell/flight time, typing profilesrefused
    • ✕Audio of any kindrejected at Pydantic validation — no flag can enable itrefused
    • ✕Gaze, biometrics, fine-grained locationcoarse consent-scoped region onlyrefused
    • ✕Sensitive-topic inferencehealth, sexuality, religion, financial distress, minors — never predicted, stored or surfaceddeny-list
    • ✓Content-free lifecycle onlyform_started · field_focused · form_completed · form_abandonedpermitted
    Why it sells

    A capability you can be pressured into enabling is a liability on every security review. A capability the schema refuses is an asset.

    Act 3 · The Proof · 4 min

    Anticipation is never credit. Prediction never grades itself.

    Attribution answers "which touchpoints preceded the sale?" Only a registered experiment answers "which spend caused a sale that would not otherwise have happened?" — the only question with budget implications. Register the holdout first, estimate second, refute third. An estimate that fails refutation is withheld, not softened.

    Every figure on this stage is illustrative — the synthetic tenant demo-fixtures, never a customer result. The causal core is PARTIAL; ghost bids are PROPOSED.

    “Which conversions did we actually cause?”
    CAMPAIGN

    1 · Register the holdout

    Causal core · partial

    Registered before the first impression — never retroactively. Assignment is a deterministic hash of customer + experiment; ACT cells consult the assignment before serving. Intention-to-treat: measured on assignment, not delivery.

    2 · Estimate with uncertainty

    X-/DR-learner · CATE
    Lift (pt)
    —
    95% CI —
    Platform ROAS → iROAS
    —
    the ledger's number, not the platform's

    3 · Refutation battery

    DoWhy
    • ·Placebo treatmenta fake treatment must show no effect—
    • ·Random common causeestimate must survive an added confounder—
    • ·Data subsetestimate must hold on random halves—
    • ·Deterministic identity onlyprobabilistic household edges excluded from causal math—

    Causal Credit Ledger · this campaign, this cycle

    unified.causal_credit_ledger · partial
    caused — lift CI excludes 0 and the model had not already scored in-marketanticipated — would have happened anyway

    Act 4 · The Profit · 4 min

    Your profit-maximizing spend is $4,048 a day. Not your budget.

    Net yield turns platform revenue into contribution after every downstream cost. The dosage curve then asks the marginal question: at what daily spend does the last dollar of incremental margin equal the last dollar of cost? It answers with confidence intervals — and refuses to extrapolate past the spend it has actually observed.

    Net yield is Preview · in development (writeback OFF). Every figure on this stage is illustrative — d* included; the dosage lane is advisory.

    “Where does the next dollar go?”

    Net-yield waterfall · hero campaign, month

    Net yield · in build · writeback off
    • !Discount liabilitynot declared by the tenant — row excluded, named as an onboarding gap; never estimatedexcluded

    Dosage curve · cumulative contribution vs daily spend

    continuous dosage · partial · advisory
    Profit-max spend d*
    $4,048
    τ(d)·margin = marginal cost
    At your budget
    —
    —

    Method: r_learner_local_marginal_moment_wls_v1 · observed support $1,200–$9,500/day. Outside support the curve returns no estimate. Advisory posture: NET_YIELD_WRITEBACK is off; nothing here bids.

    Closer

    Every other tool spends to the budget. This one spends to the profit-maximizing dollar — and tells you, with intervals, how much of today's budget is past it.

    Act 5 · The Restraint · 4 min

    A decision the system was proudest not to make.

    Every proposed move runs the Decision Control Plane: does a ValidationPassport exist with enough evidence, does the DEL score clear the floor, is authority granted for this action class, does any hard constraint veto? Below the floor the answer is a deterministic WITHHOLD with the constraint named. No confidence buys its way past a hard constraint.

    Every dollar figure on this stage is illustrative — the synthetic tenant, never a customer result. The DEL formula and the 80.0 platform floor are the shipped operating defaults, not scenario numbers.

    “What stops it from doing something stupid?”

    Proposed action

    DCP · threshold gate · live

    Governance challenge

    floors never lowered

    Try to drag it below 80. Tenants may raise a floor; the platform floor of DEL ≥ 80 cannot be lowered by anyone — including us.

    ValidationPassport

    VP-—
      shipped · threshold gate at the floorproposed canon · clipped shape (not claimed)

      Shipped score: 100 × (0.5 · passport pass rate + 0.3 · evidence completeness + 0.2 · verification weight), HMAC-signed at the DCP. The clipped-ReLU shape (margin-proportional authority, per-class cap) is PROPOSED CANON — drawn dashed, not claimed.

      DEL score
      —
      floor 80.0
      —

      Pick an action.

      Restraint Ledger · this month

      WITHHOLD verdicts + passports

      Decisions the system declined to make, with the reason. No competitor can show you their refusals.

      WhenProposalWithheld bySpend not moved

      F4 micro-geo reservation · live, halting at approval

      F4 · live · armed 2026-08-24

      The one frontier that cannot be observe-only — reserving a geography is an action — so its control is per-action human approval until two clean cycles. Watch a demo-tenant reservation hit every gate.

      • ·flag_gateF4_CALIBRATION is not "true" for the demo tenant — nothing submitted—
      • ·approval_gateroutes to the L2 decision queue — human approval required—
      • ·two_key_gateactuator not registered — an approved decision still cannot execute—

      Close · The 90 Days · 1 min

      Most platforms sell Gate 3 on day one. The sequence is the product.

      Autonomy is earned per account and per action class — never a global switch. The pilot's product is evidence about your own decisions; whichever exit you choose, you keep it.

      “How do we start — and what is real today?”
      Days 1–30

      Observe

      • Connect scoped media, commerce and site sources
      • Validate identities, timestamps, unit economics
      • No recommendations, no actions
      Gate: evidence-quality review. Nothing advances until the evidence base is trusted.
      Days 31–60

      Validate

      • First registered holdout; estimates produced and refuted
      • Counterfactuals including no-action
      • Humans execute everything
      Gate: recommendation-quality review scored against ground truth.
      Days 61–90

      Recommend → bounded control

      • Narrow, reversible, human-approved actions
      • Predicted vs realized scored openly
      • Executive scorecard
      Exit: Expand · Extend · Stop — decided on the record, never renewed by default.

      Autonomy gates — no override path

      Calibration
      Brier ≤ 0.20
      Discrimination
      AUC ≥ 0.72
      Durability
      ≥ 2 cycles
      stable lift, purchase cycles

      A model that misses any threshold recommends. It does not act. Today the AUC gate cannot even be scored: the live tables hold zero admissible forward labels — so the intent cell observes. That gap is on this slide on purpose.

      What is real today

      OFFERING_MAP v2.3 · §B.6
      SRPVDAL loop · Decision Control Plane · DEL scoring · audit ledgerlive
      Connector gateway · canonical envelope · O-1 privacy locklive
      F4 micro-geo calibration (pilot armed 2026-08-24 · every reservation L2-approval-gated · bounded autonomy after 2 clean cycles, 0 complete)live
      Causal proof core · credit ledger · measurement railspartial
      ORACLE cells 33–36 · intent edges (observe-only / shadow)partial
      Intent Engine v2 · Net Yield lane · F1 F2 F3 F5in build
      Ghost bids · clipped-ReLU authorization shape · ad control planeproposed

      IN BUILD rows — Intent Engine v2, the Net Yield lane, F1, F2, F3, F5 — are Preview · in development. Every label here is OFFERING_MAP v2.3 §B.6 on the day this page shipped, never this deck's opinion.

      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.

      Design a 90-day pilot ↗ The Decision Graph ↗