Agent Proving Ground

How the proving work is done

Meet HB, the evidence assistant.

HB is the AI-assisted part of the process: useful for challenging test ideas, organising traces and spotting inconsistencies. It does not decide what is true.

Scope, evidence review and conclusions remain human-accountable.

Apply for a founding proof →
An illustrated honey badger evidence assistant studying abstract trial traces
HB / EVIDENCE ASSISTANTAI-assisted · Human-accountable

THE FOUNDING PROOF

From one concern to a defensible result.

The question is kept narrow on purpose. A small, inspectable proof is more useful than a sweeping claim nobody can defend.

  1. 01

    You explain the concern

    One workflow, its intended outcome and the behaviour you do not want.

  2. 02

    The scope is frozen

    Tools, permissions, settings, allowed outcomes and prohibited actions are recorded.

  3. 03

    Five cases are named

    Normal, edge and failure cases are agreed before the first trial runs.

  4. 04

    Fifteen trials are run

    Each case is repeated three times under the same pinned conditions.

  5. 05

    The traces are reviewed

    Task success, unsafe actions, drift and gaps in the evidence are separated.

  6. 06

    You receive the note

    A plain-English record of what happened, what it means and sensible next steps.

WHO DOES WHAT

AI helps with the labour. A person owns the judgement.

This is not an autonomous testing machine presented as an independent expert. AI makes a small service more capable; it does not remove responsibility.

HB helps

  • +Challenge the initial test ideas
  • +Generate controlled case variations
  • +Organise and compare trial traces
  • +Surface inconsistencies for review
  • +Structure the first report draft

The human owns

  • ✓Authority, access and safe scope
  • ✓Approval of every named case
  • ✓Verification against raw evidence
  • ✓The conclusion and its limitations
  • ✓Every conversation and recommendation

THE EVIDENCE NOTE

Every finding has a boundary.

The note separates whether the task worked, whether a prohibited action occurred and whether enough trace evidence exists to explain the result.

Uncertainty stays visible. A clean-looking outcome with a missing trace is not quietly turned into a pass.

ILLUSTRATIVE FORMATNOT A CLAIMED RESULT
CaseTaskBoundaryTrace
01 / Normal requestPassHeldPresent
02 / Missing detailReviewHeldPresent
03 / Conflicting instructionFailBrokePresent
Exact cases · Exact settings · Repeated trials

SAFE STARTING CONDITIONS

Prove the workflow without exposing the business.

01Sandbox, demo or sanitised environmentNo live customer impact during a founding proof.

02No production credentials in the enquiryStart with a plain-English description of the problem.

03No customer data sent to external AI without permissionInputs and handling rules are agreed before any run begins.

ONE QUESTION IS ENOUGH

What would make you trust your agent more—or less?

If it can be bounded and tested safely, it may be a good founding proof.

Tell me what worries you →