Who Does the Law Make Answerable? A Classification Framework and Coding Schema for AI Governance Instruments

AI governance instruments are catalogued by what they cover. A registry records an instrument's jurisdiction, legal form, date, sector, and topic. That tells a reader what an instrument is about. It does not tell them what the instrument does, and in particular it does not tell them the one thing that determines whether the instrument will work when something goes wrong: who it makes answerable, for what, to whom, from what point, with what means of discharging the obligation, and with what recourse for the person affected. This article proposes classifying instruments by that structure instead. It names it Answerability Architecture and specifies the AA-12 schema: twelve coded dimensions, each with a controlled vocabulary and an explicit value for absence, organised into four groups — what the instrument is; who answers, when, and for what; to whom, how, and with what proof; and what the affected person can do and how the instrument changes over time. Two instruments on the same topic can have opposite architectures, and two on unrelated topics can share one. Topic classification hides both facts; architecture classification exposes them. The schema operationalises every construct developed across the preceding eight papers of The Answerability Series. Accountable-Party Resolution grades how precisely an obligation locates who answers, from a named natural person to nobody — and the decisive step is between naming an organisation and designating a role within it. Trigger Position records whether answerability is constituted before an act or assigned only after harm. Method Supply records whether the obligation supplies a means of compliance or only an outcome. Delegation Handling records whether the obligation addresses what happens when a task is delegated to an automated system, the dimension the article predicts will most often be empty. The unit of coding is the obligation rather than the instrument, and five design principles govern the schema: codable from text alone; absence is a value; resolve ambiguity toward the less specific reading; topic-independence; and an evidence grade on every coded value. Coding rules for eight common hard cases are specified, together with a six-step reliability protocol reporting agreement per dimension — including an honest prediction of which dimensions will need revision cycles, and a statement that if Delegation Handling cannot be made reliable it should be redesigned rather than retained on faith. The schema is demonstrated on three constructed archetypes — an outcome-based statute, a voluntary code, and an internal policy — showing that the binding instrument makes an organisation answerable with no owner inside it, the voluntary code supplies method but no recourse, and the internal policy has the strongest internal architecture and answers to nobody outside. Six hypotheses are stated before any coding, each with a refutation condition, including the counterintuitive prediction that binding instruments supply methods less often than voluntary ones. The article codes no real instrument and cites none. A registry is only as good as the verification of its entries, and verification belongs to whoever codes the corpus; the contribution here is the scheme. It is released for immediate use: a regulator can code a draft before enactment, an organisation can code the obligations that bind it and find where the law names the organisation and nobody within it, and a researcher can code a corpus and test the hypotheses. Paper 9 of 10 in The Answerability Series. Manuscript ID AA12-SCHEMA-2026-09. 38 pages, 9 figures, 21 tables, six appendices. Supplementary files deposited with this record: the full AA-12 codebook (77 coded values) as CSV, a blank obligation coding sheet as CSV, and a JSON schema that validates a coded record — all released under CC BY 4.0. Scholarly analysis; not legal advice.