Twist written mid-draft

Design, not implementation

Nothing described here is built. See the model overview.

Novelists split into plotters, who outline before they draft, and discovery writers, who find the story by writing it. It is the same split as spec-first and code-first. The craft advice is the same too: a twist found in the draft has to be planted back through the chapters before it, or it reads as unearned. This example runs that planting through the model.

The system

A fantasy novel of twenty chapters, drafted as far as chapter 14.

Artifact-set	Holds
{premise}	The synopsis: what the book is about and how it ends
{outline}	Chapter-by-chapter beats: who does what, and when the reader learns it
{story bible}	Characters, backstories, world rules, and the timeline
{manuscript}	The chapter drafts

The chapters are one artifact, not twenty. A continuity rule is written once for the manuscript, not once per chapter, for the reason given under Artifact.

Connections:

• {premise} to {outline}
• {outline} to {manuscript}
• {story bible} to {outline}
• {story bible} to {manuscript}

The last connection holds what the outline does not state: eye colour, how far a horse travels in a day, where a character was in a given year.

The premise reads: Ilse, heir to a fallen house, learns who destroyed her family and must decide whether revenge is worth becoming what she hates. The outline names the betrayer as Chancellor Aubrec, a minor court official, revealed in chapter 16. It ends in chapters 18 to 20 with Ilse executing him. Those chapters exist only in the outline, since the draft stops at chapter 14. The story bible says Ilse’s mentor, Maren, spent the year the house fell at sea.

Workflows

Workflow	Input	Owned	Output	Shape
Plotting	{premise}	{outline}	none	one link
Drafting	{outline}	{manuscript}	none	one link
Reverse outlining	{manuscript}	{outline}	none	one link
Worldbuilding	{story bible}	{outline}	none	one link
Continuity pass	{story bible}	{manuscript}	none	one link
Canon update	{manuscript}	{story bible}	none	one link

Every set is revised in place until the book is published, so each is owned by the workflows that write it and none is an output. A printed edition would be an output: it cannot be revised, only superseded by the next edition.

Two edges carry a pair of workflows with opposite roles. On {story bible} to {manuscript}, continuity pass treats the bible as canon: the draft must agree with it. Canon update treats the draft as the record: what a chapter establishes is written back into the bible. On {outline} to {manuscript}, drafting writes chapters from the outline, and reverse outlining, a technique discovery writers use, rebuilds the outline from the chapters. A book written only by plotting needs the first of each pair. A book where discoveries happen in the draft needs both, and this author works both ways.

Every workflow keeps the default leash. A write that only adds proceeds when the criteria pass. A write that removes or reverses part of a set’s standing specification needs approval. The author works alone, so the author is the approver and the person who made the change at once. What the outline and the bible state are criteria, the betrayer’s name included.

The author does not loosen the leash on {outline} or {story bible}. A discovery writer tries scenes and cuts many of them. If reverse outlining and canon update could rewrite the outline and the bible from any draft, each experiment in a chapter would become the book’s plan and its canon unseen. The outline and the bible are where the author keeps what the book has decided, so every change to what they state reaches the author. The cost is a review per reversal, by the person who wrote the scene.

The change

While drafting chapter 14, the author sees that the betrayal lands harder if it comes from Maren. They write the scene in which Ilse finds Maren’s seal on the letter that doomed her family. The diff touches chapter 14 only.

Variant A: the book has an outline

Expected run

1. Lift. The change touches {manuscript}, the source.

2. Find candidates. Drafting and continuity pass own {manuscript}, so both are upstream candidates. Reverse outlining and canon update read the manuscript, so both are downstream candidates.

3. Distill, per workflow. Each candidate reads chapter 14 within its own span and states what the change is for, not whether it is right. Drafting treats the outline as canon and continuity pass treats the bible as canon, so a judgement of rightness would have them call the chapter an error while reverse outlining and canon update call it new canon. All four state one intent: the person who betrayed Ilse’s family is the person she trusts most, Maren. The name stays in the intent. A reading that kept only someone Ilse trusts would let a replay choose a different character, and what the author found was who, not a category.

4. Ask the controllers above.

   • Drafting asks the outline’s controller. The outline names Aubrec, so it is affected. Its criteria:

     • Maren betrayed the house, and Aubrec did not.
     • The reveal comes in chapter 14.
     • At least two earlier scenes admit both readings of Maren: innocent on a first reading, damning on a second.
     • The climax still turns on Ilse’s choice about revenge.

     The outline is drafting’s input, so drafting routes the job to every owner of the outline: plotting, reverse outlining, and worldbuilding. Each reads the root intent and the outline’s criteria within its own span. Reverse outlining reads the manuscript as well, so its intent carries the chapter’s expression. Plotting asks the premise’s controller, which holds: the premise names no betrayer, and the climax still turns on revenge. Worldbuilding reads a bible that is itself about to change.

   • Continuity pass asks the bible’s controller. The bible puts Maren at sea, so it is affected, with the criterion Maren’s backstory places her where she could have sealed the letter. The bible is continuity pass’s input, and canon update alone owns it, so the job is routed there.

   • Reverse outlining and canon update ask nothing above. The manuscript is their input, so they read chapter 14 itself and run from it.

   The first two criteria and the bible’s criterion replace what the outline and the bible state. They are not joined with it. A replacement removes part of the set’s standing specification, so the writes that carry these will meet the leash.

5. Schedule. The run ledger holds the jobs. Reverse outlining and canon update read the source, which is settled by definition, so they start. Plotting’s job has nothing pending and runs. Worldbuilding’s job reads the bible, which canon update is about to write, so it waits. The outline’s controller writes once all three have contributed. Drafting’s reconciliation reads the outline and continuity pass’s reads the bible, so both wait. Neither running job waits on the reconciliations at the manuscript, so the two edges with opposite roles never wait in both directions.

6. Replay. Canon update puts Maren ashore in the year the house fell. The bible said she was at sea, so its controller reports that the write reverses the standing specification, and the leash stops it for approval. Canon update asked nothing above. The check runs at every write to the set, so it catches this one anyway. The stop reaches the author, who approves, and the approval is recorded as a decision.

   Worldbuilding’s job is then ready. Chapter 2 of the old outline had Maren’s ship coming into port, so its contribution to the outline moves that beat. The ledger hands the outline’s controller all three contributions together, each with its provenance, and they join at the controller. Plotting’s contribution adds nothing the others do not, and does not conflict. The joined write names Maren, moves the reveal to chapter 14, plants her in chapters 3 and 9, ends on Ilse’s choice about her, and makes chapter 2 consistent with the bible. It removes Aubrec as the betrayer, the chapter 16 reveal, and the ending in which Ilse executes him. Those are part of the outline’s standing specification, so the leash stops this write too, whichever of the three workflows contributed each part. The author approves. Drafting runs once, after the approved write.

7. Reconcile at the source. Drafting and continuity pass both reconcile at the manuscript, and the criteria from the outline and the bible join at its controller. It keeps the author’s chapter 14, which meets the reveal criterion. It adds setup scenes in chapters 3 and 9, which the setup criterion requires and the change never considered: nothing before chapter 14 supported the reveal, which is why the scene read as a twist rather than a turn. Chapters 15 to 20 are not drafted, so the manuscript has no ending to change. The new ending lives in the outline, and drafting writes those chapters from it later. Its report lists the kept chapter and the two added scenes. The write adds scenes and removes nothing the manuscript stated, so the leash lets it proceed. The reconciled manuscript reaches reverse outlining and canon update as a change of its own, and the outline already plants the setup and ends on Ilse’s choice, so nothing further runs.

Settled state

• {premise} is unchanged.
• {outline} names Maren, reveals her in chapter 14, plants her in chapters 3 and 9, and ends on Ilse’s choice about her, with the author’s approval recorded.
• {story bible} puts Maren ashore in the year the house fell, with the author’s approval recorded.
• {manuscript} carries the author’s chapter 14 and derived setup scenes in chapters 3 and 9. Chapters 15 to 20 are still undrafted.

Status: Holds

The run carries the twist back through the outline and out to the chapters that need it. The schedule in step 5 is where it could stall: on an edge with opposite roles, each side’s input is the other side’s owned set. The ledger settles it. The manuscript is the source and counts as settled by definition, so reverse outlining runs and drafting waits.

No second chapter 14 is derived. The author’s scene is reconciled against criteria derived above it without its prose, and it stands as written, so there is no choice between two versions to record.

The reverse workflows matter. With only plotting, drafting, worldbuilding, and continuity pass, nothing owns {outline} or {story bible} from the manuscript’s side. Step 6 of selection then raises the affected sets as obligations, and the gap in the declarations is visible: the author needs reverse outlining and canon update, and has not declared them. The run no longer ends silently with the outline still naming Aubrec.

The leash stops the run twice, at the bible and at the outline, and that is what this author wants. Both writes reverse what the set stated, and most of each was not written by the author. The author wrote chapter 14. The agents wrote Maren’s new backstory, the moved reveal, and a new ending. Passing the criteria cannot accept those, because the criteria were derived from the author’s own intent. Part of each stop is ceremony: the author is asked to accept a name they already chose by writing the scene. The rest is not. The reconciliation at the manuscript does not stop, because it only adds. If a setup scene had to rewrite something chapter 3 already said about Maren, it would reverse part of the manuscript and stop as well.

The run also leaves one choice to whichever state it reaches. The criteria say Aubrec did not betray the house, and nothing else about him. One replay keeps him as a red herring and another cuts him, and both states meet the criteria. Confluence lists actors as the usual topology, but topology is only what the criteria constrain, so a character no criterion names is a coordinate. An author would not call a character a coordinate. The model’s reply is that a difference that matters is a criterion nobody stated, and stating it starts the next iteration. The reply is consistent. The leash no longer lets the choice pass unseen: either write removes Aubrec as the betrayer, so the outline stops, and the author sees whether it keeps or cuts him before approving. The criteria still do not decide. The author does, and the approval records it.

Variant B: the author writes without an outline

The book has no {outline}. The connections are {premise} to {manuscript} and {story bible} to {manuscript}. Drafting reads {premise} and owns {manuscript}. Continuity pass and canon update are unchanged.

Expected run

1. Lift. The change touches {manuscript}, the source.
2. Find candidates. Drafting and continuity pass own {manuscript}, so they are upstream candidates. Canon update reads it, so it is a downstream candidate.
3. Distill, per workflow. The intent is the same as variant A.
4. Ask the controllers above.
   • Drafting asks the premise’s controller. The premise holds that the climax turns on Ilse’s choice about revenge. It has no place for a betrayer’s name or a setup beat, so for those criteria it answers too coarse, and nothing lies above it. No set above the manuscript is affected, and no strain is recorded on {premise} to {manuscript}. Drafting’s only job is its reconciliation at the manuscript.
   • Continuity pass asks the bible’s controller, which is affected as in variant A, and the job is routed to canon update.
   • Canon update reads the manuscript and asks nothing above.
5. Schedule. Canon update reads the source and starts. Continuity pass’s reconciliation waits for the bible. Drafting’s has no pending input, and the ledger may hold it so the manuscript is reconciled once.
6. Replay. Canon update puts Maren ashore. The write reverses what the bible stated, so the leash stops it for the author, who approves.
7. Reconcile at the source. The manuscript’s controller keeps the author’s chapter 14, adds setup scenes in chapters 3 and 9, and meets the bible’s criterion. Chapters 15 to 20 are not drafted and there is no outline, so there is no ending to change. The write only adds, so it proceeds.

Settled state

• {premise} is unchanged.
• {story bible} puts Maren ashore.
• {manuscript} carries the author’s chapter 14, with setup in chapters 3 and 9. Chapters 15 to 20 are undrafted, and only the premise says how the book ends.

Status: Unresolved

The run finishes the book, and the earlier question of whether strain on a too-coarse input clears or stays open no longer arises: the premise’s controller answers too coarse, and no strain is recorded. Two things remain open.

• Criteria with no home above the source. The reveal and setup criteria pass through the premise and land on the manuscript, the only set that can hold them. There they are derived by the controller that also reconciles the author’s chapter, so they are not independent of the change they check, which is what expression stays with the source exists to prevent. The model now says so, and the reconciliation report names criteria the source’s controller derived itself, but naming them does not make them independent. The leash does not help. It compares a write with the standing specification, and these criteria were built during the run, so there is nothing standing to compare them with. The one stop in this variant is on where Maren was, not on who betrayed the house. Where the setup criterion should live depends on how levels are identified, which is still open.
• Coupling inside one artifact-set. Once the run closes, the setup in chapter 3 is held by nothing but the chapter. A later edit that removes it strains no connection. In variant A the outline carries the beat, and cutting it strains {outline} to {manuscript}. Inside the set, keeping chapter 3 consistent with chapter 14 is the controller’s job, and nothing records the beat for the controller to keep.

What it tests

• Direction in the workflow. Fiction uses both directions on the same edges: canon constrains the draft, and the draft amends canon. Without the reverse workflows, selection raises the affected sets as obligations and names the missing declarations.
• How workflows are selected. An affected input is routed as an intent to all three of the outline’s owners, and the ledger collects their contributions so the outline is written once. It treats the source as settled, so edges with opposite roles never wait on each other.
• Distillation stating what a change is for, not whether it is right. Asked for rightness, the workflows that treat the outline and the bible as canon read the chapter as an error. It also keeps a proper name: an intent abstract enough to drop the name lets a replay swap the author’s discovery for a different one that meets the same criteria.
• Controllers answering for their own sets where the only set above is too coarse. In variant B no strain is recorded on the premise, and the setup criterion’s only home is the source.
• Intent and criteria in a field where quality is argued. Whether a twist is earned is a craft judgement, but two earlier scenes admit both readings can be checked by a reader who has not reached chapter 14.
• Reconciling at the source in creative work. The author’s scene is a prediction, and it stands as written because it meets the criteria and prose is a coordinate. If a style guide constrained the prose, the prose would be topology, and reconciliation could change the author’s own sentences.
• The leash default on contradicting writes, with a solo author as the approver. It stops the bible and outline writes, including canon update’s, which asked nothing. The author keeps it on both sets, so a draft experiment cannot rewrite the plan or the canon unseen. Part of each stop is ceremony, because the author approves a name they already wrote.
• Change reaching backwards in reading order. The setup lands in chapters the author thought were finished. Connections are undirected, and the reach they give is not bound to the order a reader meets the chapters in.
• Coupling inside one artifact-set. The model sees connections between sets. The consistency fiction depends on most, one chapter against another, lies within a set, and variant B shows what is lost when no connection stands in for it.
• Actors as topology. Variant A leaves a character’s survival to the replay, and the leash puts that choice in front of the author at the outline’s stop.