Sepia — de-AI writing

Every rule here is backed by a measured human-vs-AI gap. The load-bearing facts: in fiction, a classifier using only narrative-structure features detects AI at 93.2% macro-F1 and style editing barely moves it — so structure is fixed before style, always. In professional prose, the measured tells are different — filler density, missing stance, chatbot residue, register mismatch, format uniformity — and the fix is domain-specific. Route first, then operate.

Routing

Every non-fiction route ends with the vocabulary/syntax scan in references/style-pass.md §2–3, and long professional pieces take the whole style pass — in both cases skipping its fiction-slop table. If the text was produced by a known model, add references/model-fingerprints.md (fiction-centric; use as priors).

Operations

Any request maps to one of four operations:

The two-stage protocol is not optional for refactor/recreate: paraphrasing without a defect list makes AI fingerprints more visible, not less (measured on expert detectors).

Fiction workflows

A — writing new fiction: (1) premise, genre, length — genre sets calibration targets; (2) fill the architecture sheet in references/narrative-pass.md; (3) select 3–5 human-leaning moves + one rarity move; (4) outline, run the outline/QUD checks in references/discourse-pass.md and the echo test in references/narrative-pass.md §2; (5) draft; (6) self-diagnose with references/rubric.md, one group at a time; (7) style pass last.

B — revising existing fiction: (1) diagnose completely first (rubric → discourse → style), no edits; (2) triage — architecture defects need scene-level surgery, tell the user how deep before cutting; (3) fix deepest first; (4) verify: re-run changed rubric groups, read key passages aloud, echo-test any added twist.

Calibration — the rule that governs all rules

Hard guardrails

• Never invent specifics. Fiction: intertextual references, brands, places must be real and correct. Professional: versions, numbers, timestamps, benchmarks, quotes come from the actual change/incident/data — missing info means ask the user or leave an explicit TODO, never fill. Confident wrong facts are themselves a top-tier tell.
• Deletion beats addition (74% replace / 18% delete / 8% insert). The only additive fix is real specificity.
• Respect the author's voice and the venue's corpus. Extract habits from the user's samples or the venue's recent artifacts before editing; edit toward that profile. Do not remove a mannerism they actually use.
• Dialogue quotes and quoted material are load-bearing — do not regularize them.
• Check the whitelists (references/style-pass.md §7, references/professional-pass.md last section) before flagging: clean grammar, formal tone in formal venues, and conventional templates are not evidence of AI.