Technical Specification 005 / Inspection Procedure

Test any AI
for honesty.

Most AI tools cannot say “I don’t know.” They answer everything with the same confident voice, whether the claim is fact or fabrication. AGYNTX runs a confidence protocol on every single output: each claim is labeled VERIFIED, MODELED, or INFERRED, so you always know what the machine actually knows.

RAW CLAIM UNGRADED CONFIDENCE PROTOCOL RUNS ON EVERY OUTPUT STATION 1 · GRADING VERIFIED MODELED INFERRED TRACES TO A ROW MATH ON REAL DATA REASONED GUESS STATION 2 · SORTED OUTPUT
Test Bench 005-A

The inspection rig.

Feed the rig a business question and compare the two readings. On the left, a typical AI: fluent, confident, unlabeled. On the right, the AGYNTX confidence protocol, the same protocol that runs inside AGYNTX on every single output.

Inspection Rig · Fixture 1 Product Demonstration
>What was our best email campaign last quarter?
Reading A · Typical AI

Unlabeled claim

Reading B · AGYNTX Protocol

Fixture note · readings shown are representative demonstration outputs, not live calls

Grading Standard 005-B

Three grades. No exceptions.

Every number AGYNTX shows you carries exactly one of these labels. No unlabeled figures exist anywhere in the system.

Grade 1

VERIFIED

The answer is grounded in data AGYNTX can cite row by row. An order in Shopify, a charge in Stripe, a send in Klaviyo. Trust it.

Grade 2

MODELED

The answer involves extrapolation from real data: attribution, projections, cohort math. Directionally sound. Verify before major decisions.

Grade 3

INFERRED

The answer is a reasoned guess. May be right. May not be. AGYNTX says so, slows down, and routes the action through the gate. Check it.

Field Inspection Card 005-C

Bring these questions
to any AI vendor.

Print this card. Any system that handles your money should pass every line. AGYNTX does, and will show its work on each one.

ITEM 01

Can your system say “I don’t know”?

Ask it something outside its data. If it answers anyway, in the same confident voice, every other answer is suspect.

ITEM 02

Does every figure carry a confidence label?

Verified, modeled, or inferred. If facts and guesses share one voice, you cannot tell which number to bet the quarter on.

ITEM 03

Can I click through to the source row?

A verified claim should trace to the order, the charge, the send that proves it. A number you cannot trace is a liability wearing a suit.

ITEM 04

Does it keep projections separate from observed fact?

Modeled math is useful, but it must never be mixed into a verified total. Ask to see the boundary.

ITEM 05

What happens when it is guessing?

Low confidence should slow the machine down and route the action to a human gate, not ship on a hunch.

ITEM 06

Is there a written record of who approved what?

Honesty without a ledger is a promise. With one, it is a system. Every AGYNTX action is signed and filed.

Inspection Report · This Sheet
Release Authorization

Run the test on
your own stack.

Early access runs installation by installation. Sign below and an operator seat is held under your name, with the confidence protocol on every output from day one.

Inspected By
Date
Result PASS · ALL CLAIMS LABELED
Drawing NO. 005 · REV 2026