// Parallel preview — nothing on the classic site is replacedView classic site →
// SIGNAL FACTORY · Live on the production runtime

Raw signals in. Governed decisions out.

Pick a scenario and press Start. Raw events stream onto the factory floor, normalize into canonical events, form entity signals, face the ReLU gate, and run the SRPVDAL pipeline — including one deliberate guardrail block you'll see in red.

The run answers the doorman question: which of these conversions did the ad actually cause — and which shoppers were already on their way back?

idle

1 · Event rail

Raw connector events arrive left-to-right. Each one is normalized into a canonical event — cyan seam — before anything downstream may touch it. Nothing goes straight from connector to action.

2 · ReLU gate

score = max(0, uplift) × confidence × log(1 + n) — pass requires uplift ≥ 5%, confidence ≥ 0.70, n ≥ 15 (operating defaults). Failures drop to the filtered tray with their reason, verbatim from the engine.

Signals in

max(0, uplift)

Passed — glowing cyan

3 · SRPVDAL pipeline

Seven stages light in sequence. Click any lit stage to expand its trace. Validate carries the deliberate red block.

4 · Decision card

The executed action with its full provenance chain — raw → canonical → signal → gate → guardrails → decision — and the funnel that got it here.

5 · ORACLE — anticipatory intent

The factory you just ran is the governed chassis ORACLE rides on. Latent Intent Inference reads consent-gated micro-signals and scores every person and account into four intent stages — with calibrated probabilities and explanation paths, never certainties. Anticipatory intent with proof of causal lift; not mind-reading. Preview · in development.

Stage 01Awareness
Stage 02Consideration
Stage 03In-market
Stage 04Purchase-imminent
ServeIntent Scoring API — intent vectors, predicted next interests with calibrated probabilities, stage classification, and the explanation path, designed for sub-100 ms responses (Cell 34) — design target.
ProveThe caused-vs-anticipated credit ledger — holdout, ghost-bid, and geo experiments with X-Learner / DR-Learner uplift and DoWhy refutation before any estimate is trusted (Cells 26–27, 35–36).
GovernObserve-only by default. Promotion gates: Brier ≤ 0.20 · AUC ≥ 0.72 · stable lift ≥ 2 purchase cycles. Consent-first ingestion, sensitive-topic deny-list, no audio signals — ever.
Next: run the Nexus → Enter the Counsel Room Platform walkthrough