cdp-atlas: Structural Analysis of the CDP Corporate Questionnaire (Module 7, 2024–2026) — Rights-Safe Aggregate Release

This dataset release presents a structural analysis of the CDP corporate questionnaire — Module 7, the climate / GHG-emissions module — across the 2024, 2025 and 2026 cycles. The questionnaire is analysed as an institutional computation device: a structure that converts corporate environmental behaviour into things that can be observed, compared and scored, and that leaves other things unobservable. The unit of analysis is the datapoint (a single cell, field, option or attachment slot a responding company can fill in). No company response data is used anywhere in this study; the sources are public CDP documents only (questionnaire/guidance, scoring methodology, and official change-tracking artifacts). Coverage: 2,958 datapoints modelled across the three cycles (2024: 1,001 · 2025: 983 · 2026: 974; the 2026 cycle is portal-only, so its structure is reconstructed from CDP's official change-tracking artifacts — a documented provenance exception, not a silent inference). Each datapoint is classified on three axes under the ratified rubric CDP-SNE v1.0 — S (Substance), N (Narrative), E (Enforceability) — with honest evidence tiers: 1,949 human-ratified, 496 AI-reviewed (kept below the ratified tier), and 513 held (uncertainty is a first-class outcome, excluded from analysis rather than forced into a label). CDP-SNE's E axis is a domain-specific operationalisation and is not interoperable with the general SNE canon's E = Expectation. Six derived structural indices (structural observability, evidence enforceability, narrative discretion, route comparability, temporal comparability, disclosure burden) are computed per year with formulas shipped verbatim; where evidence is insufficient the value is reported as not_evidenced or partial_lower_bound, never silently 0 or 1. This study is scoped to one link in the raw-data-to-investor disclosure chain: whether CDP's own questionnaire, as designed, structurally demands or enables externally verifiable, machine-traceable answers, as opposed to accepting curated self-report — not whether companies' actual submissions are raw-sourced or narratively processed, which is out of scope for this rights-safe v0.1 (see methodology.md §2). Only the E axis is a structural demand for auditable provenance, and it is a small minority of the instrument in every cycle studied; ratifiable coverage itself fell from 79.5% (2024) to 74.6% (2025) to 43.1% (2026), with the 2026 structure portal-gated and reconstructed rather than independently extractable. This is the rights-safe v0.1 release: aggregate statistics, derived indices, methodology, schema documentation, a build-provenance manifest, and a synthetic example illustrating the internal data model. It contains no text from any CDP document and no company response data; every public artifact is built from scratch on an allowlist and passes a hard-gate audit (substring scan of every internal document-derived string against every public file). Row-level data (question identifiers paired with classifications) is withheld pending a licensing clarification with CDP. "CDP" is a trademark of CDP Worldwide, used nominatively to identify the object of study; this is independent academic research, not affiliated with, endorsed by, or approved by CDP. Generative AI tools were used for extraction tooling, row-level review labour and document assembly under the audited pipeline described in methodology.md (AI recommendations were audited on a stratified sample before batch confirmation); research design, the ratified rubric, all ratification decisions and conclusions are the author's own.