Optimism must be earned.

Signals, records and confidence

Signals are short source reports; discoveries explain a documented development, its human significance and its limits. This edition was researched and edited with AI assistance under the site owner’s direction. Review attribution is recorded with the sources; no human source review is implied. Automated ingestion prepares candidates, while publication remains an explicit editorial decision.

Maturity describes where a claim has been tested or used: experimental, emerging, demonstrated, deployed, or contested. Claim confidence is a separate assessment of support for that specific claim. Confidence is currently unassessed in the legacy collection; deployment does not make it high.

Event, source-publication, capture and last-review dates stay separate. Unknown dates are shown as unknown. Catalog dates describe Bright’s own additions and corrections.

Understand what the experience shows.

A molecular prediction, a repository signal, a forecast and an illustrative simulation are different kinds of material. Each world states its source, units and limits. Visual scale does not establish importance, confidence or human benefit. Model confidence is not experimental proof.

Open a contextual record to inspect the human problem, reported result, method, limitations and original source. The same evidence can remain open while you move between worlds. An exhibit makes a question explorable; it does not strengthen the underlying claim merely by making it look convincing.

Related records have explicitly curated thematic connections. They do not imply causation, shared authorship, or a measured dependency. The model genealogy distinguishes declared repository ancestry from editorial connections. Captured repository metadata is not a reviewed achievement, openness certification or a global census of AI progress.

Five states. No certainty score.

These are editorial classifications of specific claims, not a ladder every development must climb. Study design, replication, population, operational use, and unresolved disagreements are separate dimensions.

What we check.

We favor original papers and first-party institutional records. Each entry identifies the human problem, AI’s specific role, reported result, limitations, publication date, and source-check date. Source checked means the entry was compared with the cited record; it does not mean Bright reproduced the experiment or commissioned expert peer review.

Publication dates describe the source milestone. Catalog additions and dataset captures carry separate dates. Historical studies may have later follow-up work outside this small collection.

What changes our minds.

Corrections, retractions, failed replication, revised measurements, and credible counterevidence should change the record. Future updates must preserve prior claims and explain what changed. No correction is hidden behind a quietly rewritten headline.

How records are placed on the scale.

Bright’s front page arranges a curated edition of records by physical size, from about 10²⁵ meters to 10⁻⁸. Every placement is an editorial, order-of-magnitude choice with a written reason, never a measurement taken from the source.

• ● Size of the subject — the physical size of the thing being discussed: a heart, a carved letter, a protein.
• ○ Size of the setting — the meaningful physical environment where the work happens: a classroom, a field, a computer.
• ◐ Contextual scale — used only when a physical size clearly helps understanding and can be defended in a sentence, and always labeled with what it measures. The first station, for example, is placed by how far the gravitational wave traveled, while the AI work happened at a 4-kilometer detector.
• Off the axis — policies, agreements, software releases, institutions, programs and abstract systems with no honest physical size. They are not forced onto the scale; they live in Now, the Index and search.

Every record page shows its placement and the reason for it. The14 scales are a curated edition; all records are in the Index.

What we do not infer.

A mouse study is not a human treatment. A benchmark is not a social outcome. A predicted molecule is not a drug. A deployment is not proof of net benefit. Potential impact is written as possibility unless directly measured.