Provenance · last verified 2026-09-06

Evidence behind our results claims.

The provenance layer under BackTier's measurement claims: first-party AI-answer tests with named sources, confidence levels, last-verified dates, and causal claims flagged as unproven.

What this evidence set contains

Tests published

4

Facts with a named source

8 of 8

Facts at High confidence

7 of 8

Interpretations flagged unproven causality

3 of 4

Run date · last verified

2026-09-06 (ephemeral, 30-day re-verification window)

Platform and conditions: Google Gemini 3.6 Flash, accessed programmatically via API, no browsing tools enabled. Single completion per prompt, no retries, no location hint supplied, English. One run on one model is a single observation, not a measurement programme — it is published so the method is inspectable.

Provenance per test

The rules this page is held to

• A FACT is a direct observation with a named source and a confidence level. If we cannot name how we observed it, it is not published as a fact.
• An INFERENCE is our reading of the facts. It is labelled separately and carries no evidentiary weight of its own.
• Any statement that one action caused an AI-answer outcome is flagged UNPROVEN CAUSALITY unless a controlled test isolates it. On this evidence set, no such test exists.
• Every observation is ephemeral. Model weights, routing, and retrieval change without notice, so a TEST entry is true as of its LAST VERIFIED date and nothing more.
• The subject of this evidence set is BackTier itself. No client data, prompt set, or outcome figure appears here.

The rules this evidence is published under

Authorship, sourcing, dating, ownership, and the claims we refuse to make are stated in the editorial policy. Anything on this page that turns out to be wrong is corrected or retracted in public, with a dated entry in the corrections log — you can report an error there yourself.