A breakdown that shows its seams

Point aether at a screenplay and it does what a first assistant director does with a highlighter and a long night: it reads every scene and pulls out the production — the sets, the characters, the props, the vehicles, the wardrobe. Scene by scene, entity by entity.

The tempting way to build that feature is to make it look effortless. Clean list, no hedging, everything crisp and certain. We think that’s exactly wrong.

The script doesn’t always know

Screenplays are underspecified on purpose. “A dark car pulls up” — is that a specific vehicle you’ll design, or set dressing? “INT. WAREHOUSE” appears twice with slightly different sluglines — one place, or two? A radio is mentioned in one scene and heard in another — same prop, or the location just has radios?

A confident tool papers over all of this and hands you a tidy list that’s subtly wrong. You find out later, in the expensive place, when the art department budgeted two warehouses.

So we show the seams

aether’s breakdown says what it actually knows, and flags what it doesn’t:

• inferred — spoken of, never staged. The chapel everyone references but no scene enters.
• merged from two headings — two sluglines resolved to one place, with both originals kept so you can split them back.
• may be one radio, may be three — genuine ambiguity, surfaced instead of guessed.

Every extracted element carries its provenance and a confidence, and — this is the important part — a grounding quote. You can always see the exact line the machine read to make its claim. Nothing is asserted from nowhere.

Honesty as a trust primitive

Here’s the thing about a tool that reads your script for you: you have to trust it, and trust is not built by being confident. It’s built by being right about your own uncertainty. A system that tells you “I’m 60% sure these two headings are the same warehouse, here are the two lines, you decide” is far more useful than one that silently picks and moves on.

It also fits how the rest of aether works. The breakdown is a set of proposals, not facts. A human reviews them, confirms, merges, or rejects — and only then do they cross into the decided record. The machine did the reading. You did the deciding. Both of you can see exactly why.

That’s the whole philosophy in one screen: the AI proposes, honestly, with its seams showing. You decide. And the reasons are on the record.