Groundwork
Enter workspace
Methodology

Where is the evidence this can be trusted?

This is the question behind every score Groundwork produces. This page shows how we find the evidence, weigh it, and report it.

01

The six modules

Each module checks a different stage of the AI product lifecycle.

Use Case Readiness
Checks if an AI use case is worth building before you build it.
Launch Risk
Checks what can break after launch, before it breaks.
Platform Risk
Checks if an AI vendor or platform is safe to adopt.
Automation Trace
Checks where AI decisions escalate to a human, and where they don’t.
Agent Grounding
Checks if an AI agent’s outputs are grounded in real, verifiable data.
UX Audit
Checks if the interface tells users the truth about what the AI is doing.
02

The confidence scale

Every finding gets one of four confidence tags. No finding gets a free pass.

strong evidenceweight 1.00Independently verified. Confirmed by a source outside the vendor’s own claims.
limited evidenceweight 0.67Some support found. Not fully confirmed.
inferredweight 0.33No direct evidence. Based on pattern or context.
no signal foundweight 0.00We looked. We found nothing. We say so.

If we can’t back a finding, we don’t hide that. We label it.

03

How the score is calculated

Each confidence tag carries a weight. A report’s overall confidence score is the average of those weights across every rated pattern.

Two patterns that both read as “nothing found” take opposite paths. A pattern that is not rated leaves the calculation altogether; a pattern rated no signal found stays in it at weight 0.00, adding to the denominator and pulling the score down.
  • A pattern that is not rated is left out of the average entirely. It neither drags the score down nor inflates it.
  • no signal found is different: it is a real tag carrying a weight of 0.00, so it counts toward the sample and lowers the average. Looking and finding nothing is itself a result.
  • Every report shows its sample size, so you always know how many patterns the score rests on.

This is not a single number pulled from nowhere. It is a calculation you can check.

04

Evidence types: claim vs. verification

A vendor saying “we are SOC 2 compliant” is a claim. A SOC 2 report confirming it is verification. These are not the same strength of evidence, and Groundwork does not treat them as the same.

Self-reported

From the vendor’s own materials, marketing, or website.

Independently verified

From a third-party audit, public filing, breach disclosure, or other outside source.

The confidence tag on each finding reflects this gap. A claim alone does not earn “strong evidence”.

05

Our rule on regulatory flags

Groundwork never tells you a system is compliant. It never tells you a system is non-compliant.

When a regulatory issue is found, Groundwork does three things:

  1. Names the issue in plain language.
  2. Cites the specific article or clause it relates to.
  3. Flags it for legal review.

That is the full extent of the call we make. Compliance is a legal determination. We are not a law firm.

06

What Groundwork does not do

  • Groundwork does not replace legal review.
  • Groundwork does not certify compliance.
  • Groundwork does not guarantee outcomes.

A diagnostic tool that pretends to do more than this is not trustworthy. We would rather state our limits than overstate our value.

07

Sources behind the scoring

Groundwork’s judgments are grounded in named frameworks, not a generic checklist.

  • The six failure patterns from The Interface Is Not the System (Nelly Pourmehr)
  • The four ControlGap failure modes: silent failure, overtrust, misinterpretation, feedback loop collapse
  • Compliance reference material: SOC 2, GDPR, ISO 27001, HIPAA, PCI DSS, FedRAMP

Every score traces back to one of these sources. Nothing is scored on vibes.