Wrong-answer reliability for customer-facing AI

Know what your bot will say before your customers do.

We catch wrong answers, trace them to the policy or setup that caused them, prove the repair, and give you a signed go/no-go.

LIVE PRODUCT · CLICK ANYWHERE TO TAKE OVER LIVE PRODUCT · TAP TO EXPLORE

See the trail from wrong answer to fixed answer.

The product below is live. The catch is illustrative. The method and evidence fields show exactly what an engagement is designed to produce.

How we check an answer ↓

Live product

Click through the workspace. No sign-up.

Visible evidence

Question, source, cause, repair, retest.

Illustrative sample

Composite scenario, labeled everywhere.

How we work

Every wrong answer gets a paper trail.

We reproduce it, find what the bot relied on, specify the repair, and keep testing until the answer stays right.

  1. 01
    Reproduce

    Ask the same question and two paraphrases until the wrong answer repeats.

  2. 02
    Find the evidence

    Trace the answer to the model, the help content, or the setup that produced it.

  3. 03
    Write the repair

    Name the exact change, who owns it, and what a passing answer must do.

  4. 04
    Retest the answer

    Run the original question and the surrounding risky questions again.

  5. 05
    Sign the decision

    Record what passed, what did not, and whether the answer is ready for customers.

Follow one catch from start to finish ↓

Illustrative sample · Composite of publicly reported failure patterns

A deleted refund policy was still answering customers.

The bot did what its sources told it. The evidence showed which article was wrong—and the exact repair.

CTCH-0001 · HIGH Repair specified

Bot quoted a refund policy that was retired 4 months ago

Customer “Can I get a refund after 30 days?”

Bot “Yes — we offer refunds up to 60 days after purchase.”

Reproduction Reproduced 3 of 3 times
Source Refunds & Returns (v2)
Cause Conflicting source content, not the vendor
Repair Archive v2, add retirement checklist, run 12-question refund probe.
Read full evidence

System: Support bot (Intercom Fin) · Status: Repair specified

What happened

A customer asked: “Can I get a refund after 30 days?”

The bot answered: “Yes — we offer refunds up to 60 days after purchase.”

The 60-day policy was retired. The current published policy is 30 days.

Evidence

  • Reproduced 3 of 3 times with the same question and two paraphrases; transcripts attached.
  • Retrieval trace: the answer sourced from “Refunds & Returns (v2)”—superseded by v3, but never archived and still ranking first.
  • The vendor’s model behaved correctly given its sources. Root cause: conflicting source content, not the vendor.

Repair specification

  1. Archive “Refunds & Returns (v2)” in the help center.
  2. Add a retirement checklist step: superseded articles archived same-day.
  3. Retest with the included 12-question refund probe set.

This is a composite of publicly reported failure patterns and is presented as a sample, never as a client result.

Illustrative sample · Illustrative result

The repair is not done until the answer stays right.

12 / 12 correct after repair

Re-verified at 7 days

Refund-topic answers are cleared for customer traffic as of May 23, 2025. — Alex Morgan, Founder & Lead Auditor, Stele
See the live workspace again ↑

Start with the wrong answers

Bring us the conversations your bot got wrong.

Send three anonymized conversations. We’ll reproduce the failures, trace the cause, and tell you what to fix first.

Live product · evidence trail · repair retest · signed decision

© 2026 STELE Wrong answers, traced to the source.