Reference

This section centralizes the framework’s reference material: the numerical baselines used for calibration, the published circuits used as ground truth, the controlled vocabulary used throughout the site, and the bibliographic sources. The pages in this section are dense reference tables rather than narrative.

What this section is for

The framework requires calibration of every reported score against published reference values. Those values are scattered across the literature; this section consolidates them into a single table with citation links, so a reader can apply the calibration step of the audit without re-collecting the references each time.

The known-circuits page serves the same purpose for circuit-recovery comparisons: it lists the head and MLP sets identified in the canonical IOI, greater-than, subject-verb agreement, and gendered-pronoun circuits, with citations.

The glossary defines every term the framework uses in a fixed way; the bibliography records every citation in the framework’s pages.

How the reference pages are structured

Each reference page is a table with a fixed set of columns: quantity, value, task, model, source paper, figure or table number, notes. Where a quantity has been re-measured by the framework on a matched distribution, the re-measurement is reported in a second column alongside the source value. The two columns let a reader see at a glance whether the published value reproduces.