Epistamate · Insights

Evidence, AI, and what
survives scrutiny.

The failure modes behind the product: when a citation does not support the claim, apparent consensus is one evidence family, a human reviewer is treated as a checkbox, evidence goes stale, or credible sources simply do not converge.

The Scarce Resource Isn't Information. It's Attention.

Different professions run on different sources and different assurance standards, but they share one increasingly scarce resource: human attention.

Read →
When Evidence Does Not Converge

A synthesized answer can read smoothly while credible evidence still points in different directions. How to diagnose disagreement before resolving it.

Read →
A claim can be correct, well-sourced, and no longer true

Citation accuracy and corroboration are not the same property as currency. A taxonomy for how claims expire.

Read →
When humans and AI work together, what the evidence actually shows

A human reviewer being present is not the same as oversight working. What controlled research says about the combination.

Read →
A human reviewer is not one variable. AI oversight requirements treat it like one.

Availability, competence and state fail independently. Oversight needs a more precise model of the human reviewer.

Read →
A careful AI research process still missed the story that mattered most

A structured research process can still flatten evidence tiers and miss a market-moving fact.

Read →
AI judging AI has a reliability problem

Repeated judge trials reveal why evaluator output is useful process evidence but not authority by itself.

Read →
A model agreeing with another model is not the same as a claim being verified.

Cross-model agreement can improve robustness without becoming independent corroboration.

Read →
In AI-assisted consulting work, the most confident output may need the most checking

When AI moves outside its capability frontier, surface confidence is a poor guide to correctness.

Read →
The compounding evidence problem: why agentic research pipelines fail quietly

Dynamic pipelines inherit, transform and propagate evidence errors across stages.

Read →
A cited source and a supporting source are not the same thing

A citation resolving correctly does not establish that the cited material supports the proposition attached to it.

Read →
Binding law vs non-binding guidance

For compliance work, source authority and legal status cannot be flattened into topical relevance.

Read →
AI research tools erode the expertise that can catch what they get wrong

The verification capacity of the reviewer can decline as the tool substitutes for the work that built it.

Read →
The internet is eating itself: model collapse and the evidence base

Synthetic content changes the evidence environment that future models and researchers depend on.

Read →
Confident and wrong: why AI can't tell you when it doesn't know

Output confidence is a property of generation, not an inspectable measure of the underlying evidence.

Read →
The peer review loop is breaking

When AI participates in both production and review, the provenance of the evidence base matters more.

Read →
Fragility is not falsehood. That's the harder problem.

A claim can be true and correctly cited while resting on an evidential base too fragile for the decision being made.

Read →
A citation used to be evidence. It isn't anymore.

Fabricated and misbound references make source identity and claim support separate verification problems.

Read →
Forty sources, one claim: corroboration vs amplification

Source count can disguise one underlying evidence path. Independence is an evidence-lineage problem.

Read →