§ SIGNAL / THRESHOLD INTELLIGENCE · ReLU / FILED 2026
Modern delivery algorithms are neural systems, and their activation behaves like the ReLU function that powers them: output = max(0, input). Signals below the activation threshold are filtered to zero — spread-thin budgets never register. Signal's threshold agents find the activation point by structured search, then concentrate spend so your campaigns live above the line.
§01 · The threshold reality
Delivery systems score every ad set continuously. Scores below the activation threshold are cut to zero — not reduced, removed. The auction never sees them. Ten ad sets at $15 a day can each sit below the line and buy collectively nothing that compounds.
The same total spend, concentrated, can cross the activation point and start compounding: better delivery earns better signals, which earn better delivery. The difference between the two spending shapes is structural, not luck.
The thinking behind this page is written long-form in the ReLU Lens essay — why activation thresholds, learning phases and budget concentration decide delivery outcomes before creative ever gets a vote.
§02 · Discovery
Guessing the threshold wastes weeks. Signal's discovery agent runs a binary search against the live system and treats a sharp performance cliff as the signature of the activation point.
Discovery procedure — operating defaults, tunable per account
The output is not a feeling about bids. It is a number, with the evidence trail that produced it.
§03 · Amplification
No ad set runs below $50 a day — below that, the signal an ad set generates is too weak for the delivery system to learn from. Fewer, stronger ad sets beat many weak ones.
80% of budget flows to the top performers rather than being spread evenly (operating default). For an activation push, concentration can narrow to a single ad set.
Where a steady budget would hover near the line, deliberate bursts push a campaign across it, then hold with less — momentum is cheaper to keep than to create.
§04 · Compounding
Once a campaign is activated, the agent shifts from searching to exploiting: 85% of decisions reinforce what is working, 15% keep exploring adjacent thresholds (operating defaults) so the system never goes blind. Momentum is evaluated on a 24-hour window, compounds at a bounded rate, and decays the moment it stops being fed — nothing is assumed permanent.
The last phase matters as much as the first: when momentum decays into fatigue, the baton passes to Creative Intelligence for rotation instead of riding a declining asset down.
§05 · Playbooks and rails
Discovered thresholds, concentration patterns and pulse cadences are stored per vertical in the playbook pattern store. Every new campaign starts from the pattern its vertical has already earned, not from zero.
Threshold discovery and amplification propose; the decision gates dispose. Bid and budget moves pass identity, policy and constraint gates before execution, land in the audit trail, and stay reversible. Autonomy starts at observe.
Stated plainly: the parameters on this page are the system's operating defaults, not promised outcomes. What a threshold strategy returns is measured the only way Signal accepts — against a holdout, in the causal credit ledger.
§06 · See it run
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.