Signal turns ad platforms, analytics, your site, and inventory into governed evidence: what is about to happen, what actually caused what, and which spend earns its keep. The deepest division today — and the home of the media-buying stack.
Calibrated probabilities and proof of causal lift. Targets labeled as targets, and a credit ledger that separates what marketing caused from what it merely anticipated.
"Mind-reading." "This account will buy." Audio capture — ever. Predictions are probabilities, not promises.
Every division runs the same 7-Stage Governed Decision System, under the same Safety Guardrail Engine, writing to the same Compounding ROI Memory. Here is Signal's slice of it — stated plainly, so you can hold us to it.
Not a mockup — the live desk runs the production runtime, seeded and replayable, and this is one of its decisions.
Pixel fires drop 35% overnight; orders are flat. A naive automation slashes budgets against the broken number. Signal quarantines the metric, evaluates the ROAS floor on verified server-side evidence, and holds spend steady — the correct move is refusing the panic cut.
Illustrative scenario data · deterministic and seeded — replay it yourself on the live desk.
Every figure below is read from the Boss modules that actually run — the Signal Factory engine, its guardrail set, and the simulator — or labeled as platform spec where the capability ships observe-first. The test suite asserts this page against the code: if an engineer changes a floor, this page fails the build until it tells the truth again.
The gate is max(0, uplift): weak evidence is zeroed, never averaged in. A segment must clear all three floors — measured uplift, decision confidence, and sample size — before it may carry budget.
Operating defaults — uplift floor 5% · confidence floor 0.70 · sample floor n = 15Budget moves only through cleared gates, and swings are capped in the Validate stage. The demo deliberately requests a +25% swing so you can watch the guardrail block it — the part most demos hide.
Operating defaults — budget swing cap ±20% · bid swing cap ±30%Every proposed action is a value object, not a hunch: it carries expected value, decision confidence, and the conversions supporting it. In the seeded run, the winning move is +12% on campaign_7 — because the numbers say so.
Operating parameters of the seeded run — winning action +12% campaign_7 · expected value $8,400 · confidence 0.86 · support n = 48Fatigue is a verdict, not a reflex. When creative CTR is flat while post-click conversion falls, the cause is the site, not the creative — the simulator's Reason stage makes that call in front of you. Real-time fatigue tracking ships in the Meta & Paid Social module.
diagnostic rule flat CTR + falling post-click CVR → not creative fatigueSpend paces to modeled incremental lift per segment — X-Learner and DR-Learner estimate who is actually persuadable — so budget concentrates where dollars change outcomes, not where platforms claim credit.
estimators X-Learner · DR-Learner · optimization target iROAS, not platform ROASLookalike expansion with the upgrade classic lookalikes lack: every candidate carries a calibrated intent probability and stage — awareness, consideration, in-market, purchase-imminent — so you expand toward people approaching in-market, not people who merely resemble your seed.
method two-tower embeddings + nearest-neighbor retrieval · stages 4, calibratedMulti-touch attribution calibrated by experiments — holdout, ghost-bid, geo — with every causal estimate stress-tested by DoWhy refuters before it is trusted. And when a feed lies, it is quarantined: the demo's e-mail scenario refuses proxy-open inflation at the gate; the simulator defends the ROAS floor on verified 3.1× evidence when pixel fires drop 35%.
refuters placebo-treatment · random-common-cause · data-subset · ledger caused vs anticipatedAnticipatory models ship observe-only and are promoted to autonomy only past explicit calibration gates. Micro-signal ingestion sits behind a hard consent gate — no consent, no ingestion, no exceptions. Outcome figures on this site are design targets, labeled as targets, never outcome promises.
promotion gates Brier ≤ 0.20 · AUC ≥ 0.72 · stable lift ≥ 2 cycles · serving sub-100msPulled from the modules that run the demos and the simulator. The suite's sync test imports the runtime and fails if this table drifts from the code.
Live-in-demo figures are executable — same seed, same numbers, every run. Platform-spec figures come from the owner's Signal Intelligence specification and ship observe-first behind the promotion gates above. Design targets stay labeled as targets.
The desk narrates its own run: the checks, the gates, and the moment it commits or holds. Same runtime that ships.