Methodology

GLMI ranks frontier language models by structural properties of their identifiers. This document describes the v9 harness, the projection used to map identifiers onto the reported scale, and the sources of uncertainty.

1. Overview

Conventional leaderboards depend on inference runs, prompt suites and human preference collection. Each of those stages introduces cost, latency and a large surface for configuration differences between submissions. GLMI removes all three by scoring the model identifier itself.

2. Lexicographic Alignment (LA)

For a model identifier s, the harness applies the following steps:

1. Normalization. The identifier is lowercased, whitespace is removed, and separators (-, _, .) are stripped. Digits are retained only where they contribute to the polynomial below.
2. Rolling polynomial. A rolling product is accumulated over Unicode code points, weighted by position: p = p · (c + 7i + 1) mod 2^31-1.
3. Tail projection. The accumulator is reduced to a tail value in [30, 98] and blended with a fixed floor: score = 55 + 0.45 · tail.
4. Interval estimation. A 95% interval is reported from harness variance. The interval does not include model, prompt or sampling variance, which are not defined for this metric.

2.1 Why not evaluate outputs

Output-based evaluation measures the model. GLMI measures the model as submitted. This distinction is deliberate: a submitted identifier is a stable artifact, whereas outputs vary with sampling temperature, system prompt, tool availability and the calendar.

3. Descriptor Breadth (DB)

The v8 cycle used Descriptor Breadth, a composite of identifier character count, hyphen density and release-tag entropy, projected through the same polynomial. The intuition was that a longer identifier indicates a more thoroughly specified model. The board is retained in the archive.

DB and LA frequently disagree. Models that rank highly on one board may rank near the bottom of the other. GLMI treats this as expected: the boards measure different properties, and neither is a proxy for the other.

4. Confidence intervals

Reported intervals reflect harness variance only. They are not comparable across boards and should not be used to construct significance claims between models. Intervals are wide enough to be informative and narrow enough to be reassuring.

5. Submission

Submission is open. Submitters provide an identifier, a vendor, and an optional release date. Models are ranked as submitted. Re-runs are not performed, since the score does not depend on execution. See Submit a model.

6. Citation

The harness ships as a single file with no dependencies. To cite GLMI:

7. Changelog

• v4.2.1 — 2026-09-19 04:00 UTC. Snapshot refresh. Added Descriptor Breadth archive.
• v4.2.0 — Added per-row Δ vs next column.
• v4.1.0 — Interval estimation moved to harness variance.
• v4.0.0 — Lexicographic Alignment replaces the v8 Descriptor Breadth index.