Threat Engine Overview

Typed triggers, a three-role debate, and a materiality bar turn many noisy inputs into a small number of evidenced signals matched to your holdings in dollars.

What is the threat engine?

The threat engine is the core of Telumin. It continuously reads six source types — SEC filings and their item codes, company fundamentals, news, price history, social chatter, and operational events — and lands them in a local warehouse. Declared watch-theses with typed triggers decide what deserves investigation.

Unlike single-model systems that emit unverified summaries, Telumin uses a three-role debate: an Advocate argues the threat is real, a Skeptic attacks it, an Adjudicator rules. Only signals that clear the materiality bar ship — and every one ships with its evidence: sources, timestamps, excerpts.

How a threat gets from data to your dashboard

The engine runs every 5 minutes. Each cycle processes all six source types and routes candidates through three gates.

GATE 1

Typed trigger fires

A declared watch-thesis specifies what triggers an investigation: an 8-K item code, a margin threshold crossing, a drawdown percentage, or a keyword. When the trigger matches incoming data, an investigation opens.

SEC filingsFundamentalsNewsPrice historySocialOperational events
GATE 2

Three-role debate

Three model roles run on the same evidence. The Advocate argues the threat is real and material. The Skeptic attacks every assumption and demands source quality. The Adjudicator weighs both sides against the materiality bar and issues a verdict with a severity score.

AdvocateSkepticAdjudicatorMateriality barSeverity score
GATE 3

Ships with evidence

Signals that clear the materiality bar are matched to your holdings in dollars and pushed to the dashboard with their full evidence chain: sources, capture timestamps, and excerpts. If Telumin cannot show its work, it does not ship the signal.

Evidence chainDollar exposureImpact horizonDecision windowWebhook fan-out

Engine Characteristics

Cycle cadenceEvery 5 minutes
Debate modelAdvocate / Skeptic / Adjudicator (3-role)
Filing → signal latency<3 minutes (8-K / 10-Q / 13F)
Earnings call distillationSame trading day
Materiality barSet by Adjudicator; cannot be bypassed
Evidence requirementSource, capture timestamp, and excerpt ship with every signal
Audit trailFull investigation record retained

Signal Lifecycle

Every investigation carries a status:

TRIGGEREDA watch-thesis trigger matched incoming data; investigation opened
DEBATINGThree-role debate running: Advocate / Skeptic / Adjudicator
PUBLISHEDCleared the materiality bar; signal in dashboard, API, & webhooks with full evidence
DROPPEDFailed the materiality bar; investigation record retained but signal not shipped

Why not a single model?

Single-model failure modes

A single LLM hallucinates, lacks real-time market context, and cannot reliably reject its own outputs. Telumin solves this by separating roles: the Advocate must argue the case, the Skeptic must attack it, and the Adjudicator must rule. No role can shortcut the debate. If the Skeptic’s objections are not resolved, the signal does not ship.

Dive Deeper