# Authorial Decisions Contract

Use this contract when sentence-level cleanup is not enough because the draft's selection, order, implication, agency, or closure still feels generic. Keep it behind the ordinary writing conversation; do not make the user learn a second product.

## Boundary

This is an authoring aid, not an AI detector or detector-evasion recipe. Do not score a draft, estimate who wrote it, target statistical prevalence, or force a text away from patterns associated with a model.

Research on long-form fiction can suggest questions worth asking, but it does not establish universal rules for emails, essays, posts, applications, or an individual writer. A clear theme, linear sequence, bodily metaphor, tidy ending, or direct explanation may be exactly right in context.

Use an authorial-structure pass only when:

- the user explicitly asks for deeper restructuring or re-authoring;
- the user says a polished or lightly humanized draft still does not feel like theirs;
- an AI-first draft contains a central decision that cannot be repaired at sentence level; or
- the requested long-form work depends on selection, disclosure, causality, or closure.

Ordinary cleaning remains a minimum-effective edit. Do not silently turn a short email or clear factual draft into a structural exercise.

## What belongs to the writer

Inspect these as decisions, not style tells:

- **Selection:** what earns inclusion, what is omitted, and which example or consequence carries the point.
- **Disclosure:** what is stated directly, what the reader can infer, and whether the draft explains its implication twice.
- **Sequence:** why ideas or events arrive in this order and whether changing the order changes the argument.
- **Causality and agency:** who acts, what causes what, and whether the ending defaults to personal insight or acceptance without support.
- **Counterpressure:** which uncertainty, objection, competing motive, or unresolved fact genuinely matters.
- **Specific world contact:** whether a supplied source, place, object, work, or event is named accurately instead of replaced with vague allusion. Never manufacture specificity.
- **Closure:** what is resolved, what remains open, and whether the ending has earned its certainty.
- **Reader relationship:** what the writer assumes, explains, withholds, or asks of this particular reader.

For fiction, these decisions may also include character agency, subplot function, revelation timing, temporal order, moral ambiguity, and the relation between emotion, body, and setting. Apply them only when the genre and brief support them.

## Decision triage

Classify a questionable structural choice before changing it.

### Supported

The current brief, draft, source, or confirmed writing pattern supports the choice. Keep it even if a research paper or detector associates it with AI writing.

### Reworkable

The underlying point is supported, but its order, emphasis, repetition, or landing weakens the piece. Rework it within the meaning lock. Do not add a new opinion, event, source, motive, or conflict.

### Authorial gap

The draft needs a judgement that the supplied material does not contain: what matters most, which tension is real, what should remain unresolved, or what the writer actually concludes. If the gap is central, ask one focused question. If it is peripheral, stay source-close or remove the disposable passage.

### Format-required

The apparent regularity is required by the medium, genre, house style, accessibility need, or user instruction. Keep it.

## Working method

1. Lock claims, facts, polarity, uncertainty, causal boundaries, and requested outcome.
2. Make a silent decision map covering selection, disclosure, sequence, agency, counterpressure, specificity, closure, and reader relationship.
3. Preserve supported and format-required choices.
4. Repair reworkable choices with the smallest structural movement that solves the actual problem.
5. Ask one ordinary-language question only when a central authorial gap blocks an honest result.
6. Compare a source-close candidate with an authorial candidate. The authorial candidate may reorganise supported material but may not invent the writer's judgement.
7. Run the same semantic, factual, autobiographical, and leakage checks on both candidates.
8. Return one paste-ready result. When the structure changed materially, add a compact note explaining the reason.

Do not create artificial humanity by adding randomness, unnecessary subplots, abrupt time jumps, named brands, direct reader address, moral ambiguity, rough grammar, or unresolved endings. Human authorship comes from the user's choices, not from performing the inverse of a classifier.

## Personal-pattern evidence

Authorial decisions need suitable evidence. Learn selection, disclosure, sequencing, counterpressure, and closure only from complete pieces with reliable boundaries in the relevant context, or from an explicit confirmed preference.

Do not infer these traits from:

- two or three short onboarding answers;
- dictated transcripts with reconstructed boundaries;
- lightly edited AI output;
- one long sample treated as repeated evidence; or
- population-level findings about human or AI writing.

Keep unsupported authorial features Unknown. A user may explicitly confirm one as Preferred.

## Research influence and scope

The distinction between surface cleanup and underlying authorial choice was sharpened after reviewing Russell et al., [*StoryScope: Investigating idiosyncrasies in AI fiction*](https://arxiv.org/abs/2604.03136) (COLM 2026). Its results concern long-form fiction generated from a specific corpus and model set. Write Like Me carries forward only the bounded product lesson that surface editing cannot substitute for the writer's decisions; it does not import the paper's classifier, feature frequencies, or authorship claims.