A draft can cite a real source and still say more than that source supports. Reviewing it means checking both the reference and the claim.
SceneArc is a media-analysis workspace in active studio development, not a finished public release. Its drafting area, Composer, builds drafts from saved notes about moments in films or television episodes. SceneArc calls these notes observations. A reference connects part of the draft to one of those notes.
During generation, SceneArc checks returned references against the notes supplied to the model. It does not check whether the notes support the writing. That second task remains with the writer.
Choose the notes before generating the draft
The writer starts with filters or notes already attached to an outline. SceneArc retrieves matching notes owned by that writer and limits how many it sends to the model. Notes explicitly added by the writer come first, followed by notes already cited in the draft, then other matching notes.
This order gives deliberate selections priority when the available set is too large. The limit counts notes; it does not measure their combined length or guarantee that they fit the model’s input limit.
SceneArc uses the same final selection in two places: the material sent to the model and the reference check applied to its response. Otherwise, a reference could pass the check simply because it was in the larger collection, even though the model never received that note.
Check the references returned with the draft
Composer organizes a draft into sections and beats. A beat is a short unit of draft text, such as a claim or a piece of narration. Each beat can carry references identified by the saved notes’ numbers, or reference IDs.
When the model returns a draft, SceneArc removes reference IDs that were not supplied for that request and removes repeated IDs within each beat. If a beat has no references left, it flags that beat as needing evidence.
The check leaves the prose unchanged. A beat with a reference can still make an unsupported claim. The hook, thesis, section headings, summaries and closing also need review: they have no reference slots and receive no equivalent missing-reference flag.
The flag tells the writer where a reference is missing. Its absence does not establish that the writing is correct.
Check whether the note supports the claim
For illustration, imagine this scene:
- Saved note: “Mara pauses before signing the contract.”
- Draft claim: “Mara never wanted the deal.”
- Careful interpretation: “Mara’s pause could suggest hesitation, but the note alone does not establish what she wanted.”
Neither the character nor the scene is a SceneArc research record.
If the note was supplied to the model, its reference can pass SceneArc’s check. “Never wanted” still goes beyond what the note records. The revision keeps an interpretation while making its uncertainty clear. A stronger claim would need more evidence from the film.
Try the same check on a draft: open one cited note, identify the words it supports, and mark what the draft adds. Revise the claim or find further evidence for the addition. Where wording, action or context matters, return to the original source as well.
Know what later editing changes
The generation check applies to the reference lists returned with that request. A writer can later attach another note. Saving an edited draft does not repeat the generation check, and the saved references do not distinguish model-supplied references from ones added afterward.
Drafting before selecting sources can help explore an idea. When sources come later, review must include evidence that challenges the draft, as well as material that supports it. Selecting notes first gives the generation check a defined set to work from, but it does not remove that editorial responsibility.
There is a separate limit on the record. SceneArc stores links to the current notes, so editing a note can change what its reference displays. Regenerating replaces the current draft body; the reviewed workflow does not retain its previous version or a snapshot of the exact note contents sent with each request.
Reopening a reference lets a writer inspect a note now. Establishing exactly what the model received earlier requires a separate historical record.
Six questions to ask of a drafting tool
Trace one claim through the tool you are evaluating. These questions test different parts of the workflow. SceneArc’s current answers show why passing one check does not settle the others.
| Question | What the reviewed SceneArc workflow establishes |
|---|---|
| Who chooses the source material? | The writer selects or filters notes; the application prioritizes and limits the final selection. |
| Can you establish what reached the model? | The generation request uses that selection. An exact historical snapshot of its contents is not retained in this workflow. |
| Which references are checked? | References returned in generated beats are checked against that request’s notes. Later saves do not repeat this check. |
| What happens when a reference is missing? | Generated beats left without references are flagged. The other draft fields have no equivalent flag. |
| Who checks whether the source supports the claim? | The writer remains responsible. The reviewed workflow does not record a completed editorial approval. |
| What can you inspect later? | The current draft and links to current notes persist. Earlier generated bodies and generation-time note snapshots are not retained here. |
A useful review should be able to say what was checked, what still needs judgment and what record remains. See the implementation and development status in the SceneArc case study.