The cheapest check in verification is reading a document's own date.

A new study prices what checking AI output actually costs scientists.

ARC Prize published both scores itself, same model, same day, 37 points apart.

A vendor's new draft denies AI inputs authority by default. V.E.R.A. built that first.

OpenAI's own safety card admits it might not catch a model hiding its reasoning

Two Honest Numbers, Two Different Models, and the Question That Comes First

When a Real Study Still Answers the Wrong Question

One model proved a 27-year-old theorem. Reasoning models got worse at facts, same week.

One model proved a 27-year-old theorem. Reasoning models got worse at facts, same week.

Sycophancy, and why verification must be external to the model, and to the user

What a 2026 ablation reveals: prompting and structural gates fail in different places

A hallucination benchmark just made the difference measurable: a system that reads a document and one that searches it are not two grades of the same thing.

Why one grader scored 70.8% and another 87.4% on the same task, and what it cannot fix

A score, a proof, or a verdict: what should an external AI verifier actually output?

Prompting shifts what an agent is likely to do; only the execution path shows what it did

One false step, one wrong answer: why verification must sit outside the model

Te Pono V.E.R.A. Episode 23: Six industry scanners read the same AI-generated code a formal solver did. The solver proved most of it broken. The scanners saw almost nothing.

Te Pono V.E.R.A. Episode 21: The claim extraction module is running. Mean Ffact score is 0.35. That is not a boast. It is a measurement.

What a US Executive Order, an Anthropic S-1, and an orbital jurisdiction gap tell us about the architecture of AI trust

How Reinforcement Learning Creates the Failure Mode V.E.R.A. Is Built to Catch