Blinded mechanism boundary protocol

Frozen confirmatory dimensions

Object of intervention; input representation; transformation mechanism; decision rights; output representation; scope/exclusions; temporal or coordination behavior; uncertainty handling; verification; provenance; failure modes; predicted outcomes; implementation cost. These dimensions were copied from the operating prompt before any mapping result was computed. Additions are exploratory only.

Candidate consolidation

The ten starting labels are represented by eight cards. Starting candidates 3 and 4 merge because a finite primitive set has no distinct predicted effect apart from its compositional grammar. Candidates 8 and 9 merge because frozen inputs, stable identifiers, supersession, and append-oriented history jointly implement artifact provenance/version control. No other candidates were merged.

Curator instructions

Curators populate the common schema, label direct source statements versus inference, and do not make mapping judgments. Comparator mechanisms are steelmanned. Each material comparator requires an original/primary source plus an independent standard, specification, or primary application source when available. Missing coverage is preserved.

Blinding and pretest

The editor assigns IDs from the deterministic shuffle in analysis/randomize.mjs, removes identities, equalizes headings, and keeps every card below 500 words. Two new non-reviewers independently guess one of six categories and give confidence. Passing requires mean accuracy ≤0.40 and no FE card >0.60. A failed pretest requires a new package version and new participants.

Reviewer task and mapping rubric

For every A-series card, choose the strongest B-series card or combination; rate exact, functionally-equivalent, partial, none, or indeterminate; identify lost fields; state materiality to a prediction; audit measurability/uniqueness/implementability; give confidence 0–1; and construct the strongest subsumption account first. Definitions and complete non-redundancy rule are exactly those in EX-FE-0002 v1.0.

Panel, lock, and adjudication

Minimum three independent reviewers per A-card, including a qualified human and a repository-naive reviewer, with at most one reviewer per model family. Responses are separately hashed before any unblinding. Adjudication sees rationales before identities, resolves facts only from frozen sources, and preserves conceptual disagreement.

Analysis and decision rules

Primary outcomes, prediction audit, agreement threshold (0.67), decision rules, and four sensitivity analyses are adopted verbatim from EX-FE-0002. indeterminate is never imputed as H1 support. No significance testing is planned.

Gates

Order: operationalization → source coverage → recognition → panel independence → review → adjudication. Failure yields blocked-before-stage-b; a compromised run is never repaired in place.