Skip to content
veri—ableMy research

OUR EVIDENCE APPROACH

A clear view includes
what we don’t know.

Collection reviewed 30 September 2026. This opening research edition is a focused collection, not a complete catalogue of AI developments.

Know what kind of visual you are using.

Teaching simulations use authored rules to explain mechanisms. Our next-token probabilities are computed from toy scores; they are not commercial-model measurements. Recorded GPT-2 tokenizer examples are separate and identify their vocabulary hash. Scripted tool and agent examples show local calculations and observable actions, not private reasoning.

Go to the original material.

We use evaluator publications, papers, system cards, incident disclosures and datasets. Each story identifies who produced the evidence and links findings to the relevant source. A source being inspected does not mean we independently reproduced its results.

Keep different kinds of evidence distinct.

  • Measurements describe a model and setup under specific conditions.
  • Developer and incident reports retain their attribution.
  • External investigations have their own access and scope limitations.
  • Editorial interpretation is not an additional experimental result.

Read the chart, including its limits.

The capability chart uses METR-Horizon-v1.1 aggregate results, retrieved 30 September 2026. The source page was updated 8 May. The raw snapshot and every stored configuration were rechecked against the published file on 2 October 2026. Human-expert minutes measure task difficulty, not AI running time. The chart shows releases since 2023; the table and download contain all 26 configurations. The selected point shows the interval supplied by METR. The scale is logarithmic. Estimates above 16 hours carry METR’s reliability caution.

Source: METR’s measurements and methodology. Analysis: METR analysis repository. Citation: METR, Measuring AI Ability to Complete Long Tasks (2025), arXiv:2503.14499.

Use the right denominator.

Actions, runs and incidents are different units. A sample audit is not an exact population estimate. Reports sharing the same logs are not independent replications. A single numerical safety grade would hide these distinctions, so we do not use one.

Ask stays close to the collection.

The learning guide offers authored explanations and links into the current journey. It does not generate live answers or change your visual without a visible action. Collection search returns labelled editorial briefs. Where a synthesis service is available, AI-generated responses are labelled separately and constrained to those briefs. External web search opens separately; its results have not been verified for this collection.

Keep your research.

Saved journeys, stories, questions, collections and notes are stored in this browser. They do not automatically sync to an account. You can export them from My research. Original account records remain accessible in the previous-workspace archive.

Correct the record.

Publication dates, event dates and review dates mean different things. We keep material revisions visible and disclose missing follow-up evidence. If you find a mistake, send the source and the claim that needs correcting.