Table of Contents

Designing Intent Models

Why you're reading this page: This page explains how to derive intent from behavior and how to design intent models: intent ≠ label, signals, confidence, policy. It is useful when setting up your first intent model or turning existing rules into intent-focused ones.

How meaning emerges from behavior.

1. What is intent?

Intent is:

  • not an event
  • not a state
  • not a rule

Intent is the direction behind observed behavior.

  • Events → happened
  • Behavior → observed
  • Intent → inferred

Intent is computed, not defined.

2. Intent ≠ Label

This is critical.

Wrong:

Intent = "Fraud"

Right:

Intent = probability distribution over hypotheses

Example:

  • AccountRecovery: 0.62
  • SuspiciousAccess: 0.28
  • Unknown: 0.10

Intentum's Confidence concept comes from this.

3. How intent models work

Intentum does not impose a single method.

There are three levels:

Level 1 — Heuristic intent models

Simplest and fastest. Intentum provides RuleBasedIntentModel for this: pass a list of rules (each returns a RuleMatch or null); first match wins. Use ChainedIntentModel to try rules first, then fall back to an LLM when confidence is below a threshold. Intent can include Reasoning (e.g. which rule matched or "Fallback: LLM").

var rules = new List<Func<BehaviorSpace, RuleMatch?>>
{
    space => space.Events.Count(e => e.Action == "login.failed") >= 2 && space.Events.Any(e => e.Action == "password.reset")
        ? new RuleMatch("AccountRecovery", 0.8, "login.failed>=2 and password.reset")
        : null
};
var model = new RuleBasedIntentModel(rules);
var intent = model.Infer(space); // intent.Reasoning set when a rule matches

Pros: deterministic, explainable, fast
Cons: does not scale, limited pattern variety

Level 2 — Weighted signal models

This is where real-world use begins.

  • login.failed → +0.3 risk
  • captcha.passed → -0.2 risk
  • device.verified → -0.4 risk

Confidence: confidence = Σ(weights × signals)

Pros: flexible, tolerant, tunable
Intentum's default model is close to this.

Level 3 — AI-assisted intent models

Here AI enters the loop.

  • embeddings
  • clustering
  • LLM reasoning
  • hybrid scoring

But:

AI does not produce intent.
AI produces intent signals.

The final decision stays in Intentum.

This is a deliberate design.

4. What confidence is (and is not)

Confidence is:

  • not accuracy
  • not truth
  • not correctness

Confidence is:

"How much do we trust this intent hypothesis?"

So:

  • 0.6 can be enough
  • 0.9 can be suspicious

In Intentum:

  • High confidence + wrong intent = 🚨
  • Low confidence + correct intent = 🟢

5. Intent anti-patterns

❌ Confusing intent with outcome

Intent = "Blocked" — wrong. Intent is before the decision.

❌ Tying intent to a single event

if (login.failed) → Fraud — that is a reflex, not intent.

❌ Using confidence like a boolean

if (intent.Confidence.Score == 1.0) — in the AI era, this line is a red flag.

6. When is intent "correct enough"?

In Intentum, the question is:

"Is the decision we make from this intent acceptable for the system?"

That depends on:

  • risk tolerance
  • business context
  • user
  • time

So intent is contextual.

7. Intent model lifecycle

Observe → Infer → Decide → Learn → Adjust

Intentum leaves this cycle open on purpose:

  • learning can live outside
  • AI can be added later
  • human feedback can be included

8. Intentum's philosophical boundary

Intentum does not:

  • claim that intent is "true"
  • simulate the human mind
  • promise absolute correctness

Intentum does:

Make reasonable decisions under uncertainty possible.

Next step: When you're done with this page → Scenarios or Architecture.