The leakage catalogued in the problem page sorts cleanly once a piece of narrative is split into three layers. The split is the classic one from narratology — Genette’s histoire / récit / narration, Bal’s fabula / story / text — and it earns its place here because the project’s own artifacts already own the layers separately without saying so.
The layers#
- Fabula — story-world facts. What happens to whom, when, and where. Species, institutions, quantities, schedules, landmarks, customs, populations, named persons. In this project the fabula is owned by the wiki and canon memory; a story or rendering consumes it and is not supposed to produce it.
- Discourse — ordering and stance. How the facts are arranged for the reader: sequence (flashback, frame, in-media-res), duration and frequency (scene, summary, iterative), and the narrating situation (person, tense, retrospective distance, embedded documents). Owned by the piece’s spec — for a test scene, the constraints block of the scene bundle’s
_index.md. - Text — diction. The actual words: register, rhythm, image, sentence shape. Owned by the voice specs. This is the layer the scene harness was built to compare.
The layers leak downward#
The audit’s central finding is that the layers are not independent, and the dependence runs in one direction: a diction choice implies a narrating stance, and a narrating stance implies facts.
- Wolfe’s register requires a situated, retrospective, fallible first person — a layer-2 commitment — and a situated narrator must be somewhere, so a second character enters a scene specified to hold one. A layer-3 choice manufactured a layer-1 violation with nothing careless anywhere in the chain.
- Le Guin’s “the hour-hand of the morning” is diction that mints a fact: an idiom exists, and people at home use it.
- Herbert’s epigraph is a layer-2 frame — the scene as recovered document — that mints layer-1 facts as a side effect: K.D. exists, kept private observations, and someone collected them.
- KSR’s register runs on quantities. Where canon supplies none (the falls’ discharge, the emitter schedule), the voice invents them, because prose in that register with no numbers in it is not in that register.
The general statement: a voice cannot be rendered without asserting things. A harness can pin the fabula and the discourse as hard as it likes; the diction layer will still generate pressure on both, and the pressure will escape somewhere. This is why the technique for this bundle is a sweep-and-triage rather than a prohibition.
Leaks are measurement: demand profiles#
Read as data instead of as errors, each voice’s leaks are a wishlist — the facts that voice needs the world to supply. The Dawn audit yields a profile per voice:
| Voice | Demands from the world model |
|---|---|
| KSR | Quantities: flows, rates, times, schedules — data/world.yaml-shaped facts |
| Martine | Institutions: org charts, internal documents, professional habits |
| Le Guin | Flora, folkways, inherited practice, in-world idiom |
| Tolkien | Named flora, taught custom, deep time in the landscape |
| Herbert | Species, ecology–politics couplings, documents and epigraphs |
| Wolfe | Calendars, observances, and a place to stand for a situated narrator |
Choosing the book’s voice therefore predicts where the wiki comes under load. A Martine book needs institutional depth the wiki currently lacks; a KSR book will pull parameters constantly and invent any that are missing.
The discourse gap#
The Dawn constraints pinned layer 2 only at its ordering edge — no cuts, no flashbacks, no resolution — and left the narrating situation entirely free. Person, tense, retrospection, frames, and iterative time-depth all varied between renderings without breaking a single written rule, which muddies exactly the comparison the harness exists to make. Genette’s categories are the checklist of what a scene spec must either pin or explicitly free: order, duration, frequency, person, tense, narrative distance, frame. The Dawn prompt already does this deliberately for one layer-1 variable (“her role, history, and interior state are not fixed”), so the precedent is to license leaks explicitly, not to pretend a constraint exists where none was written.
References#
The three-layer split is the orthodox backbone of narrative theory, not an invention of this bundle; what is ours is the synthesis on top of it — the ownership mapping to wiki/spec/voice, the demand profiles, and the downward-leak law as stated here. Delany comes closest to stating the leak outright, and only for SF.
The split itself:
- Russian Formalism — the original cut between events and their arrangement (fabula / syuzhet): Viktor Shklovsky’s “Art as Device” (1917) and his Tristram Shandy essay (1921); Boris Tomashevsky’s “Thematics” (1925) is the systematic statement.
- Gérard Genette, Narrative Discourse (1972; tr. 1980) — histoire / récit / narration, and the discourse categories this page uses as the layer-2 checklist: order, duration, frequency, mood, voice.
- Mieke Bal, Narratology: Introduction to the Theory of Narrative (1985; 4th ed. 2017) — the explicit three-layer model (fabula / story / text) whose numbering this page follows.
- Seymour Chatman, Story and Discourse (1978) — the influential two-level compression (story / discourse), common in film narratology; it merges our layers 2 and 3, which is exactly the distinction the voice harness cannot afford to lose.
- Shlomith Rimmon-Kenan, Narrative Fiction: Contemporary Poetics (1983) — story / text / narration; the standard compact survey.
The layers as working craft, from the writer’s side:
- E. M. Forster, Aspects of the Novel (1927) — “The king died and then the queen died” versus “…and then the queen died of grief”: the story/plot cut in craft-lecture form, decades before the theory hardened.
- Wayne C. Booth, The Rhetoric of Fiction (1961) — the narrating situation (implied author, unreliable narrator) as a deliberate craft decision; the layer-2 dimension Wolfe’s rendering exercised.
- Ursula K. Le Guin, Steering the Craft (1998; rev. 2015) — layer 2 as teachable exercises: person, tense, narrative distance. One of this project’s candidate voices, teaching the stance layer directly.
- James Wood, How Fiction Works (2008) — free indirect style: diction carrying stance, the 3→2 leak observed from the other side.
The leak, stated for SF:
- Samuel R. Delany, “About 5,750 Words” (in The Jewel-Hinged Jaw, 1977) — reading SF word by word as continuous world-revision, a second sun minted by a single adjective. The closest existing statement of “diction mints facts,” and the argument for why the leak runs strongest in exactly this genre.