A citation and documentation standard for the era in which machines write.
AMOS stands in the tradition of the Chicago Manual of Style and Evidence Explained, and is built for what those works predate: authorship in which machines research, draft, translate, process, and verify alongside people. It treats AI participation as first-class metadata rather than an editorial afterthought.
Each empty slot above has a reason, and the reason is part of the record.
Publication. An AMOS edition is a published snapshot built from a tagged release, containing only entries with status adopted or amended. No edition has been tagged. The first target is edition/1.0; the eventual home is amos.adaptorhouse.com. This page gives the manual a citable public address ahead of that, and publishes the adopted entries only. Drafts are named below but their rules are not stated here, because a draft is not yet the standard.
Id. AMOS produces no records of its own yet, so it has no declared id space. An id is not withheld here; there is nowhere it could correctly go. Declaring that space is a decision held by a person, not by a session, and it is open.
A citation answers who. AMOS also requires an answer to how — and in AI-assisted work, how is where integrity lives or dies. A single author field cannot distinguish a claim a scholar wrote unaided from one a model drafted out of three retrieved documents the scholar never opened. A reader, an editor, and a court each need exactly that distinction.
Existing standards omit it not by oversight but by architecture: they were built when the path from source to sentence ran through one human mind and needed no recording. AMOS records the path, because the path is now the contested thing.
The manual is organized as numbered entries. An entry is one rule, its rationale, its exemplars, and its lifecycle metadata — the atomic unit of the standard. Entries carry coordinates of the form E-<part>-<seq>, and they move through a status lifecycle: draft → adopted → amended → deprecated. Entry numbers are never reused.
AMOS serves seven audiences, from editors and historians through to AI systems themselves — Audience VII, memory-integrated generative systems applying the standard autonomously.
The machine-readable layer exists so that Audience VII is a first-class reader rather than an afterthought: the same principle the standard applies to authorship. This page publishes its own machine-readable index at /works/amos/amos.json.
AHS-01-0001 is the constitutional standards record of Adaptor House: nine principles covering position, source, distinction, challenge, transformation, witness, AI, remainder, and access. It states what the house commits to.
AMOS is constitutional before it is stylistic: it governs how knowledge is transmitted across people, languages, media, and human-machine systems.
— AHS-01-0001, Foundation
AMOS is the manual that resolves those commitments into numbered, testable entries. Where AHS-01-0001 says claims should remain traceable to their sources, AMOS says which fields a citation must carry and what makes one invalid. The standards record is the position; the manual is the operative apparatus.
These six entries are adopted and therefore normative. Each is reproduced here as its rule; the full entry — rationale, exemplars, resolution, and lifecycle metadata — lives in the standard's source of truth.
The house is written Adaptor House. The spelling Adaptor Haus is historical (2007–2020) and appears only inside verbatim quotation of historical material, where it is preserved exactly and never corrected.
When two or more words act together as a single modifier of a noun, AMOS hyphenates them when the compound precedes the noun — the attributive position — and leaves them open when the compound follows the noun or stands alone. Thus AI-assisted work but the work was AI assisted; retrieval-grounded claim but the claim is retrieval grounded.
One exception holds in both positions: a compound whose first element is an adverb ending in -ly is never hyphenated, because the -ly already marks it as modifying what follows — a highly cited paper, never "highly-cited". The test is grammatical function, not the token.
Every AMOS citation carries an archival number as its id. Archival numbers follow the grammar AMOS-<NAMESPACE>-<SEQ> and are minted only by the ledger registered for their namespace. An archival number that does not appear as assigned in its ledger does not exist, and a citation carrying it is invalid.
Every AMOS citation, in any form, carries a hard-required set across four pillars:
Identity — id (archival number), namespace, and registry (the real ledger path, which must resolve; an id exists iff its ledger says so). Object — type and title (a language-keyed object, not a bare string). Participants — at least one, each structured as {name, kind, role, mode}; a flat creator string is not sufficient. Temporal — date, when the citation was recorded.
A field with no value is omitted, never faked; a field whose value is pending is stated as pending.
When an AI system materially contributes to a cited object — research, drafting, translation, analysis, processing, or editing — it is credited as a creator with kind: ai, its model name, and its model version. AI participation is recorded in the citation's metadata, never inferred, never hidden, and never inflated: an AI that performed OCR is a processor, not an author.
When a cited work has a Chinese title, the citation renders it English → characters → pinyin, with pinyin placed directly beneath the characters in print formats. The same ordering principle — translation first, original script, romanization — extends to other languages as they enter the corpus.
The reader's language leads, the original script preserves the source, and the romanization keeps the original pronounceable:
A True Record of the Assembled Immortals of the Western Hills (西山群仙會真記, Xīshān Qúnxiān Huìzhēn Jì)
Six entries are drafted but not adopted. They are named here so the shelf can receive them and so a reader can see what the manual is reaching toward. Their rules are not published on this page, and they do not build into an edition.
Part 1, Publishing Process, is declared but holds no entries yet.
A work prepared according to this manual may say so, and point here. The citation below is itself written to the citation schema — including the parts that are not yet settled, which the schema requires be stated as pending rather than filled in.
Object. Adaptor House. AMOS: Adaptor House Manual of Style. Manual of style, public surface. Adaptor House Works, 3 August 2026. https://adaptorhouse.com/works/amos/
Id. Not minted. This page is an instance, not a decision, and instances mint into their entity's own id space — never into the AMOS ledger, which holds decisions only. AMOS has no declared entity space yet, so there is no correct ledger to mint from. Asserting a number anyway is the failure the boundary exists to catch.
Participation of record. Matthew L. Tyler (human; editor; mode: directed). Claude, Opus 5 (ai; provider: Anthropic; mode: drafted).
Provenance. Required, not optional: a provenance chain is hard-required whenever a participant with kind: ai carries mode originated or drafted. This one does.
read — Claude, Opus 5 — inputs: the adopted entries, lexicon, citation schema, editions rule, and namespace registry of the AMOS repository → output: the publishable set.drafted — Claude, Opus 5 — inputs: step 1 output, plus the Adaptor House works conventions → output: this page.verified — pending. The chain is open, and says so.Verification. Human review pending.
The standard's source of truth is the AMOS repository, which holds the entries, the lexicon, the citation schema, the archival-number grammar, the namespace registry, and the decisions record. This page is a rendering of that source, not a second copy of it — when the two disagree, the repository wins.