MIZOKI3

§ SIGNAL / MEASUREMENT FOUNDATIONS / FILED 2026

One clock. One identity. One ledger.

Meta and Google each keep their own clock, their own identity space and their own scoreboard — and both grade generously. Signal normalizes all of it into one causal accounting system, then feeds corrected values back to the platforms' own bidders.

Run the Signal Factory ▸ All capabilities

§01 · Synchronization

Platforms disagree by design. The house window doesn't.

One clock

Every conversion is re-attributed under a single house standard — 7-day click, 1-day view — regardless of which platform reported it, with empirical lag profiles per source and a recompute job the operator schedules nightly. Cross-platform numbers become comparable for the first time.

One identity

Identities stitch across platforms through SHA-256 hashes of first-party identifiers — matching without exposure. The outbound match key is deliberately unsalted, because Google and Meta both match on the SHA-256 of the normalized value and a salt would silently zero the match rate. Anything this platform stores or returns is keyed by a peppered SHA-256 instead — a held-back pepper, not a per-record salt. Raw identifiers never leave the boundary; deterministic identity is required for causal math.

§02 · The plumbing

Signal loss is an infrastructure problem. So it gets infrastructure.

RailWhat it carriesDiscipline
Enhanced Conversionshashed first-party data to GoogleSHA-256, spec-normalized
Meta Conversions APIserver-side events beside the pixelshared event-id, 48h dedup
Aggregated Event MeasurementiOS signals under App Tracking Transparency8-event priority schema
GA4 Measurement Protocolserver-side analytics eventsvalidated before send
Offline conversionsclosed-won revenue back to the clickGCLID · WBRAID · GBRAID, 90-day window

Deduplication is the quiet hero: server-side events carry the same event identity as the browser's, so nothing is counted twice — the failure mode that silently inflates every naive server-side setup.

§03 · The incrementality engine

Every number faces the experiment.

  1. Design registered before the first impression — randomized holdout, ghost bids that log the auctions a control customer would have won, or matched control geographies. Minimum detectable effect declared up front. Registration is shipped for all three; the ghost-bid execution path — the auction logger itself — is in development.
  2. Heterogeneous effects estimated per segment and region with meta-learner models — who was actually persuadable, not just whether the average moved. The shipped estimator is the DR-Learner with CUPED; the wider learner family is research-grade and in development.
  3. Automated refutation tries to break the result: placebo treatments must collapse the effect, random confounders must not move it, subset re-runs must agree. A result that fails is flagged, never shipped. The battery is built and test-pinned, but wiring it into the shipped report path is in development — until it lands, refutation is run deliberately, not automatically.
  4. Each conversion is classified caused or anticipated — against the holdout's baseline rate and the treatment rate it was measured beside — and written to the immutable ledger, the exhibit on the Signal overview. The row carries those two rates and their unit counts, plus the classifier version and a hash of its inputs, so any row can be re-derived. Effect size and interval estimation are not in the shipped ledger — design target.

Incremental return — not platform-reported return — is what the budget engine is allowed to read. That single wiring decision is most of the product.

Field notes — composite scenarios · illustrative numbers

Streaming TV is our biggest bet and our least measurable one. Finance wants to cut it. Prove it works — or lose it.

The test
The campaign ran in some markets and deliberately not in their statistical twins — the design registered before the first impression.
The ledger
Real lift in nine of fourteen markets, nothing defensible in the other five — each read against its own matched control.
The change
Doubled down on nine, cut five. The budget came out smaller and unkillable, because every remaining dollar had a receipt.

"Prove it or lose it" stopped being a threat and became the operating procedure.

Net yield · Preview · in development

Our $180 bundle is the best seller. Why is cash down while revenue is up?

The test
Every order priced at what it truly nets — components, packing, dimensional shipping, payment fees, and the returns coming back.
The ledger
The hero bundle netted between −$2 and $16 an order once a third of them boomeranged. The boring single SKU netted $31 and almost never returned.
The change
When the net-yield capability ships, the bidding signal switches from checkout revenue to net contribution — so the platforms hunt buyers who keep what they buy.

The dashboard applauded the bundle; the bank account preferred the boring one.

Stated plainly: mechanisms on this page are operating machinery except where a step says "in development" or "design target" — those are built and tested but not yet wired into the shipped path, and they are marked in place rather than left to read as live. The windows and thresholds shown are operating defaults, not promised outcomes. Rails ship flag-gated and dry-run first; live posting to ad platforms is enabled per engagement. Ledger figures shown anywhere on this site are illustrative until they are your own — produced by your experiments, on your data.

§04 · 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