MIZOKI3

§ SIGNAL / BUDGET INTELLIGENCE · ReLU-GATED / FILED 2026

Noise cannot move money.

Most budget tools react to platform-reported ROAS — a number the platform grades itself. Signal's reallocation engine is gated the way its models are built: a rectified-linear gate that passes only positive, confident, well-sampled uplift. Everything else scores zero, and zero moves nothing.

Run the Signal Factory ▸ All capabilities

§01 · The gate

Reallocation earns its trigger or it doesn't fire.

The gating score — operating defaults, tunable per account

score = ReLU(uplift) × confidence × log(sample size)

GateDefaultEffect when unmet
Measured uplift≥ 5%score is zero — no proposal is generated
Confidence≥ 70%score is zero — uncertainty cannot spend
Sample sizelog-weightedthin evidence is discounted, never rounded up

Negative uplift is clipped to zero by construction — the engine is structurally incapable of chasing a loss, and a channel that merely correlates with revenue never clears the gate.

Field note — composite scenario · illustrative numbers

The CFO wants to know why we pay Google every time someone searches our own brand name.

The test
Brand-search ads switched off in a handful of matched cities, left on in their statistical twins. Six weeks. No opinions, just registers.
The ledger
Roughly four in five of those clicks slid down to the free listing one inch below — the paid ad caused a sliver, not the number on the report.
The change
Brand-search spend rightsized to what it actually causes. The CFO got a confidence interval instead of an opinion.

A channel that merely correlates with revenue never clears the gate — this is what that rule looks like in a budget meeting.

§02 · Pacing

Spend follows measured lift, inside clamps.

The pacing rule — spend can move at most ±20% per period (operating default)

spendt = spendt−1 × clamp(1 + β · (uplift ⁄ target − 1), 0.8, 1.2)

Modeβ sensitivityCharacter
Conservative0.1slow response, tightest clamps
Moderate — default0.2balanced response
Aggressive0.3fast response, still clamped

The uplift feeding this rule comes from experiments — geographic holdouts, time-switchbacks, CUPED variance reduction — never from platform-reported ROAS. Budget expansion additionally requires the lift gate: expected uplift above threshold and a confidence interval tight enough to act on.

§03 · Portfolio

Every channel priced at its margin, not its average.

Saturation curves

Each channel's response is fitted as a saturation curve — returns diminish as spend rises. The question is never "does this channel work," it is "what does the next dollar buy here."

Marginal equalization

Budget shifts until marginal returns equalize across channels. A channel with a great average and an exhausted margin gives budget to one with room left to climb.

Funnel-position bids

Bid factors follow the customer's position — 0.7× at awareness, 1.0× at consideration, 1.35× at intent, 1.75× at purchase, 1.1× at loyalty (operating defaults, each with bounded ranges).

Field note — composite scenario · illustrative numbers

Our email discounts drive half our orders. But margins keep shrinking and unsubscribes keep climbing.

The test
Signal sorted customers by what the send actually changed — not who converted, but who converted because of it.
The ledger
About a fifth of discount sends went to sure-buyers — pure margin given away — and a smaller group was being pushed toward unsubscribe.
The change
Sure-buyers stopped getting coupons for things already in the cart; the annoyed went quiet. Margin recovered without one extra send.

Cheap to send, expensive in margin — priced at its margin instead of its average, email changed shape.

§04 · Rails

Autonomy on this desk is earned in stages.

Hard limits · operating defaults

  • Daily reallocation cap — 10% per campaign per day
  • Any single step above 20% routes to a human, always
  • Idempotent decisions — the same evidence cannot double-fire
  • Every move carries its rollback token

Staged rollout · operating defaults

  • Shadow — decisions logged, nothing executes
  • Canary 10% → expansion 20% → full 70%
  • A permanent 10% holdout keeps measuring ground truth
  • Circuit breakers halt on error spikes or CPA slippage vs control

§05 · Always-on policies

Five standing rules, with cooldowns.

Between reallocation cycles, declarative policies watch every campaign. Each fires at most once per cooldown window — a system that cannot thrash.

PolicyWatches forStanding response
CPA Guardcost per acquisition running past targetbid down
ROAS Acceleratorsustained outperformance with lost budget sharebudget up
Fatigue Swapcreative click-through decaying vs baselinerotate creative
New-Segment Probehigh-lifetime-value audiences converting at targetexpand lookalike
Margin-Aware Bidproduct margin vs category averagescale bid to margin

Stated plainly: parameters shown are operating defaults, not promised outcomes. Whether a reallocation helped is decided by the credit ledger — including the permanent holdout the engine can never touch.

§06 · See it run

Mechanism on this page. Proof in the factory.

The Signal Factory desk runs the full seven-stage SRPVDAL loop on live runtime — including a deliberate guardrail block you'll watch get caught in red. No signup, no sales call.

Open the Signal Factory ▸ Signal overview Executive briefing · ~9 min