Skip to content

Clarifications

Concepts

Sometimes Telha cannot pin down a fact with certainty. Rather than guess silently, it asks the right person to confirm it. This page explains the concept; the user guide explains how to answer, and the architecture page explains how the engine decides when to ask.

Why Telha asks

When Telha reads a document, it sometimes finds something ambiguous: a date that says "next quarter", two figures that disagree, an undated scan. Telha makes its best guess at lower confidence, marks the fact as pending, and asks someone who can confirm it.

The clarification engine works out who to ask from the information itself (the author, the person named, the owner), and checks that you are allowed to see the underlying document before it ever sends you anything.

What happens when you answer

  • One tap settles it. Your answer is recorded under your verified work identity.
  • Telha corrects itself from that point on, and any answer waiting on this fact updates automatically.
  • Important facts ask twice. For high-stakes items, Telha asks a second person to confirm before treating it as settled.
  • Disagreements go to a steward. If people answer differently, a steward decides, and every answer stays on record.

The architecture

Under the hood, the clarification engine runs after generation: it inspects each claim, identifies those below a confidence threshold or with conflicting evidence, and creates a clarification request linked to the generation trace. The request is routed to a user based on provenance and permission rules.

flowchart LR
    GEN["Generation"] --> CLAIMS["Claim verification"]
    CLAIMS --> AMBIGUOUS["Ambiguous claims"]
    AMBIGUOUS --> CLAR["Clarification request"]
    CLAR --> ROUTE["Route to user"]
    ROUTE --> ANSWER["User confirms"]
    ANSWER --> RECORD["Recorded as fact"]