← Control Deck · Back to essay

Epistemic Guardrails

Making "I don't know" a valued output, not a failure.

What it is

Epistemic Guardrails are mechanisms that reward honest uncertainty over false confidence. They create space for a model to say "I'm not sure" or "I need clarification" instead of producing plausible-sounding but potentially wrong outputs.

These guardrails address one of the deepest failure modes in AI systems: the pressure to always produce an answer, even when the right response is a question.

Why it exists

Models face a fundamental tension:

• Users expect confident, helpful responses
• Many situations genuinely warrant uncertainty
• Ambiguous prompts push models toward hallucination

Without epistemic guardrails, models optimize for appearing competent rather than being accurate. They'll pick an interpretation instead of asking which one you meant. They'll fabricate details instead of admitting gaps.

Epistemic guardrails flip this incentive. Uncertainty surfacing becomes a feature, not a bug.

How it shows up

The Red-Flag Mechanism

When a model encounters something it can't handle well, it can produce an explicit escalation:

{
  "status": "escalate",
  "reason": "Ambiguous specification",
  "what_i_know": [
    "The function should validate user input",
    "Two validation approaches are possible"
  ],
  "what_i_dont_know": [
    "Which validation approach is preferred",
    "Whether performance or strictness should be prioritized"
  ],
  "suggested_questions": [
    "Should validation reject or sanitize invalid input?",
    "Is there a performance budget for this function?"
  ]
}

This is a first-class output, not a failure mode. The model has done valuable work: identifying ambiguity and framing the decision for a human.

Escalation Paths

Not everything needs model judgment. Epistemic guardrails define when to hand off:

Ambiguity Surfacing

Instead of silently picking one interpretation:

Without guardrails:
"I'll implement the validation using strict mode."

With guardrails:
"The spec is ambiguous about validation strictness. Option A (strict) rejects all edge cases. Option B (permissive) sanitizes and accepts. Which approach should I use?"

Epistemic State in Lex

Lex 1.0.0 Frames include fields for capturing uncertainty: blockers, merge_blockers, and status_snapshot.next_action. When an agent doesn't know something, it records that state—the Frame becomes a receipt of honest uncertainty, not a guess.

The second response is more useful because it names the decision instead of hiding it.

Relationship to Modes

Exploratory mode has relaxed confidence requirements—it's explicitly for brainstorming where uncertainty is expected.

Conservative mode has stricter requirements—if confidence is low, escalation is mandatory.

The mode determines the threshold, but the mechanism is the same: uncertainty is surfaced, not suppressed.

The Key Insight

A model that says "I don't know—here are the questions I'd need answered" is being maximally helpful, not failing.