Nanako0129/sepia

Before Instructions only → After Local scripts

fb06be76abea → 815a78594cc5

- 24 | Incident postmortems / RCA | `references/professional-pass.md` + `references/domains/postmortems.md` |
+ 24 | Incident postmortems / RCA | `references/professional-pass.md` + `references/domains/postmortems.md` + `references/discourse-pass.md` §1–3 |
- 32 **Model identity.** Determine two identities before operating, each as family plus version, or unknown: the *author* model (from the user or from metadata) and the *executor* model (from your own system context — a direct statement of the model you run on outranks attribution strings such as commit trailers or signatures). A *version* is the exact release a prose-layer table is tagged with (Fable 5.1, GPT-5.6); when the vendor scopes a statement to several releases and the table is tagged with exactly those (Gemini 3 and 3.1), any of them matches; a later release that shares the numbering but that the vendor does not name (Gemini 3.5, 3.8) is outside the tag and reads the table as a prior. A generation name such as GPT-5 or Claude 5 is a family, not a version. Resolve each role on its own; the two roles are never compared. On write there is no author role. For a role with a known family, load from `references/model-fingerprints.md`: on the fiction route, that family's narrative layer as priors whenever the role's model produced or is producing the story (the author on review, the executor on write, both on refactor and recreate); on every route, that family's prose layer at the style-pass step — *operative* when the release matches the table's tag, a *prior* to check against the draft otherwise. The author's layers act on the text you were given, the executor's on the text you produce. An unknown role, or a family with no table for a layer, loads nothing for it and reports `none`. Never infer a model from the prose — six-way attribution is a trained classifier at 68.4% macro-F1 on 304 narrative features, and reading is not that classifier. Report both identities and each role's prose-layer status in every review.
+ 32 **Model identity.** Determine two identities before operating, each as family plus version, or unknown: the *author* model (from the user or from metadata) and the *executor* model (from your own system context — a direct statement of the model you run on outranks attribution strings such as commit trailers or signatures). A *version* is the exact release a prose-layer table is tagged with (Fable 5.1, GPT-5.6); when the vendor scopes a statement to several releases and the table is tagged with exactly those (Gemini 3 and 3.1), any of them matches; a later release that shares the numbering but that the vendor does not name (Gemini 3.5, 3.8) is outside the tag and reads the table as a prior. A generation name such as GPT-5 or Claude 5 is a family, not a version. Resolve each role on its own; the two roles are never compared. On write there is no author role. For a role with a known family, load from `references/model-fingerprints.md`: on the fiction route, that family's narrative layer as priors whenever the role's model produced or is producing the story (the author on review, the executor on write, both on refactor and recreate); on every route, that family's prose layer at the style-pass step — *operative* when the release matches the table's tag, a *prior* to check against the draft otherwise. The author's layers act on the text you were given, the executor's on the text you produce; on the fiction route, how far a prose layer reaches is set by the Prose layer bullet at the top of `references/model-fingerprints.md`. An unknown role, or a family with no table for a layer, loads nothing for it and reports `none`. Never infer a model from the prose — six-way attribution is a trained classifier at 68.4% macro-F1 on 304 narrative features, and reading is not that classifier. Report both identities and each role's prose-layer status in every review.
- 47 | **recreate** | Full rewrite. Extract the facts, claims, and intent from the original into a bare list; verify nothing invented; write fresh under the domain rules. On a professional route the rewrite adds no doubt, limitation, scope restriction, verdict, recommendation, or admission that neither the source nor the user's request states, even where a domain rule, a pass, or an opted-in voice asks for one. Repeating a figure the source reports is not an addition; a judgment drawn from it is. A missing one is asked for, or on an unattended run recorded on the `Deferred:` line. Sepia inference; ledger `LIAO-2026` measured models adding unsupported qualifications, mostly statements that the results do not show something, when asked to revise a text to hold up under review or to be accurate, or to write from an evidence sheet. Use when defects are structural and the text is short enough that surgery costs more than rebuilding. |
+ 47 | **recreate** | Full rewrite. Extract the facts, claims, and intent from the original into a bare list; verify nothing invented; write fresh under the domain rules. On a professional route the rewrite adds no doubt, limitation, scope restriction, verdict, recommendation, or admission that neither the source nor the user's request states, even where a domain rule, a pass, or an opted-in voice asks for one. Repeating a figure the source reports is not an addition; a judgment drawn from it is. A missing one is asked for, or on an unattended run recorded on a `Deferred:` line closing the output. Sepia inference; ledger `LIAO-2026` measured models adding unsupported qualifications, mostly statements that the results do not show something, when asked to revise a text to hold up under review or to be accurate, or to write from an evidence sheet. Use when defects are structural and the text is short enough that surgery costs more than rebuilding. |
- 68 - **Deletion beats addition** (74% replace / 18% delete / 8% insert). Additions that survive are real specificity, words a broken or split sentence needs to parse (repair is not growth), and the restorations of `references/style-pass.md` §4, allowed only where the same edit removed filler; that paragraph is where the list lives. No register drift: a rewrite must not come out more promotional than its source.
+ 68 - **Deletion beats addition** (74% replace / 18% delete / 8% insert). Additions that survive are real specificity (a fact from the source or the user, never a verdict or recommendation the source lacks), words a broken or split sentence needs to parse (repair is not growth), and the restorations of `references/style-pass.md` §4, on the conditions that paragraph sets; it is where the list and the conditions live. No register drift: a rewrite must not come out more promotional than its source.
- 70 - **Dialogue quotes, quoted material, and protected ranges are load-bearing** — do not regularize them. A caller may declare protected ranges at call time (`file:line` or `file:start-end`, as the caller counts lines): ranges are resolved against the target as received, before any edit, and the resolved text stays protected however later edits shift line numbers. Inside one, do not edit, reflow, or merge with a neighbouring line. A defect found there is still reported, on the `Protected:` line, never fixed. Quoted material is protected without being declared.
- 71 - **Call-time scope and unattended mode.** A caller may name the files to edit; then read and edit only those, and widen nothing. A caller may say the run is unattended; then never stop to ask. A defect that would need the caller's decision (the fiction triage in Workflow B, a specific the text is missing under *Never invent specifics*, a stance the source lacks that a professional recreate does not add (Operations, recreate), a reported feeling, perception, decision, or understanding held back under *Respect the author's voice*) is recorded on the `Deferred:` line and left as is. Silence and a skipped defect are different facts; the report keeps them apart.
+ 70 - **Quoted material and protected ranges are load-bearing.** Quoted material is protected without being declared, whatever its source and however unclear that source is, with one exception: dialogue and documents inside invented narrative (fiction) are draft text the passes may revise, unless the caller declares them protected. Protected quoted material changes only by a correction a domain file allows from source material the user supplies (the quotation rows of `references/domains/journalism.md`), and on recreate it passes through the fact list and the rewrite verbatim. In any quoted text, never regularize dialect or deliberate nonstandard spelling. A caller may declare protected ranges at call time (`file:line` or `file:start-end`, as the caller counts lines): ranges are resolved against the target as received, before any edit, and the resolved text stays protected however later edits shift line numbers. Inside one, do not edit, reflow, or merge with a neighbouring line. A defect found in a declared range or in protected quoted material is still reported, on the `Protected:` line, never fixed except by that domain correction.
+ 71 - **Call-time scope and unattended mode.** A caller may name the files to edit; then read and edit only those, and widen nothing: venue artifacts a pass or domain file asks you to sample are read only when the caller supplies them besides the target; otherwise none are read, and a professional route falls back to its domain baseline. A caller may say the run is unattended; then never stop to ask. With no target (the text for review, refactor, or recreate; the brief for write), ask for one, or on an unattended run say none was given and stop. On an unattended run, a defect that would need the caller's decision (the fiction triage in Workflow B, a specific the text is missing under *Never invent specifics*, a stance the source lacks that a professional refactor or recreate does not add (*Deletion beats addition*; Operations, recreate), a reported feeling, perception, decision, or understanding held back under *Respect the author's voice*) is recorded on the `Deferred:` line and left as is. Silence and a skipped defect are different facts; the report keeps them apart.

Compare on GitHub

Before Not checked yet → After Instructions only

First observation → fb06be76abea

No earlier version was observed for comparison.

Compare on GitHub

Before Instructions only → After Instructions only

fb06be76abea → 815a78594cc5

Compare on GitHub

Before Not checked yet → After Instructions only

First observation → fb06be76abea

No earlier version was observed for comparison.

Compare on GitHub

Before Instructions only → After Instructions only

fb06be76abea → 815a78594cc5

- 13 If no target was supplied, ask for it. If the user wants another operation, direct them to `sepia-write`, `sepia-review`, or `sepia-refactor` instead of switching.
+ 13 If no target was supplied, follow `SKILL.md` *Call-time scope and unattended mode*. If the user wants another operation, direct them to `sepia-write`, `sepia-review`, or `sepia-refactor` instead of switching.

Compare on GitHub

Before Not checked yet → After Instructions only

First observation → fb06be76abea

No earlier version was observed for comparison.

Compare on GitHub

Before Instructions only → After Instructions only

fb06be76abea → 815a78594cc5

- 13 If no target was supplied, ask for it. If the user wants another operation, direct them to `sepia-write`, `sepia-review`, or `sepia-recreate` instead of switching.
+ 13 If no target was supplied, follow `SKILL.md` *Call-time scope and unattended mode*. If the user wants another operation, direct them to `sepia-write`, `sepia-review`, or `sepia-recreate` instead of switching.

Compare on GitHub

Before Not checked yet → After Instructions only

First observation → fb06be76abea

No earlier version was observed for comparison.

Compare on GitHub

Before Instructions only → After Instructions only

fb06be76abea → 815a78594cc5

- 13 If no target was supplied, ask for it. If the user wants another operation, direct them to `sepia-write`, `sepia-refactor`, or `sepia-recreate` instead of switching.
+ 13 If no target was supplied, follow `SKILL.md` *Call-time scope and unattended mode*. If the user wants another operation, direct them to `sepia-write`, `sepia-refactor`, or `sepia-recreate` instead of switching.

Compare on GitHub

Before Not checked yet → After Instructions only

First observation → fb06be76abea

No earlier version was observed for comparison.

Compare on GitHub

Before Instructions only → After Instructions only

fb06be76abea → 815a78594cc5

- 13 If no target was supplied, ask for it. If the user wants another operation, direct them to `sepia-review`, `sepia-refactor`, or `sepia-recreate` instead of switching.
+ 13 If no target was supplied, follow `SKILL.md` *Call-time scope and unattended mode*. If the user wants another operation, direct them to `sepia-review`, `sepia-refactor`, or `sepia-recreate` instead of switching.

Compare on GitHub

Before Not checked yet → After Instructions only

First observation → fb06be76abea

No earlier version was observed for comparison.

Compare on GitHub