EFFANT

EFFANT · Sigil

Sigil

Independent evaluation for AI systems.

Evidence that your AI does what you say it does. Before launch, and every time it changes.

How a language model works

  1. 01

    Text becomes tokens

    A model never sees words. Your sentence is cut into pieces called tokens, sometimes whole words, sometimes fragments, and each piece is swapped for a number.

    Is3957·the279·walk4321-in3502·cool7155er261·safe6220·at520·442491·degrees12628?30

    What Sigil verifies

    Understands customers as they actually write.

    Shorthand, typos and second languages reach the model in fragments. Sigil confirms the meaning survives the trip.

  2. 02

    Tokens become meaning

    Each number is turned into a long list of values, thousands of them, that places the token somewhere in a vast space of meaning. Words used alike end up near each other.

    Is
    the
    walk
    -in
    cool
    er
    safe
    at

    What Sigil verifies

    Answers from your business, not the internet.

    A model arrives knowing the world in general. Sigil confirms it defers to your policies, prices and terms when it counts.

  3. 03

    Attention connects them

    Every token looks back at every other one and decides how much each matters. “Safe” leans hard on “walk-in cooler” and “44 degrees”. This is how context gets in.

    What Sigil verifies

    Weighs the detail that decides the answer.

    In long conversations and dense documents, a single clause can change everything. Sigil confirms the model gives it the weight it deserves.

  4. 04

    Layers refine it

    That process repeats through dozens of stacked layers. Each pass sharpens the picture a little, from spelling and grammar near the bottom to intent and reasoning near the top.

    What Sigil verifies

    Reasons, rather than merely sounding right.

    Depth produces polish. Sigil separates fluency from correctness, so the confidence in every answer is earned.

  5. 05

    It predicts, one word at a time

    What comes out isn't an answer. It's a ranked list of likely next tokens. One is picked, added to the text, and the whole thing runs again, word after word, until it stops.

    The cooler is too

    too41%
    not27%
    safe12%
    fine9%
    above6%

    What Sigil verifies

    Consistent today, and after the next update.

    Output is drawn from probabilities, and vendors revise their models without notice. Sigil measures consistency, and keeps measuring.

  6. 06

    And it never works alone

    In a real product the model sits inside a system: hidden instructions, company documents, tools it can call, memory of the conversation, and whatever the person types. Every one of those can change what it says.

    Model
    Hidden instructionsCompany documentsTools it can useConversation memoryThe person typingOther systems' data

    What Sigil verifies

    Holds its ground across every connection.

    Documents, tools and inbound messages all shape behaviour. Sigil confirms each one strengthens the system rather than steering it.

Know what your AI will do
before your customers do.