May 10, 2026
Gideon Build Log: Scenes Became The Working Room
May 10 was the day Gideon stopped feeling like a collection of AI rooms and started feeling like a story post-production workspace.
The important model became simple:
Episodes are the project tree.
Scenes are the working containers.
Writing Session and Forge are modes launched from the scene.
That changed the whole product shape.
Instead of entering Writing Session or Forge cold, the writer opens the scene first. Then the scene decides what kind of work needs to happen next.
The closest analogy is not a generic chat app.
It is closer to DaVinci Resolve for story.
Bins hold the project.
Scenes hold the work.
Validation and editing are modes.
The page remains the anchor.
The Left Navigation Got Quieter
The left menu was simplified to three core rooms:
Episodes.
World Bible.
Log Book.
Writing Session and Forge came out of the top-level navigation.
That matters because those are not places the writer should enter without context. They are actions taken from a scene.
The new workflow is:
Open Episodes.
Select a scene.
Read the scene card.
Open the page if needed.
Review the tier trail.
Discuss, Edit, or Send to Forge.
The scene is the container.
The mode is the tool.
The language changed too.
“Scene Packs” became “Episodes.”
The correct hierarchy is:
Episodes
Season
Episode
Act
Scene
That lines up with how the project already lives in Google Drive.
Google Drive holds the archive.
Gideon becomes the working bay.
The Scene Container Became The Product
The selected scene now has a clearer working shape:
Scene Card.
Scene Script / Page.
Tier Stamps.
The scene card stays primary. It gives orientation before the writer drops into the full screenplay page.
The script moved into a drawer so it is available without dominating the whole screen.
The tier trail now reads newest first, so the latest validation state is easy to find.
For EP203_A01S01, this made the scene feel like a real production surface.
Not a file preview.
Not a chat memory.
A working container.
Forge Became Quality Control
Forge was narrowed.
The Episode Source Board came out.
Forge no longer needs to browse episodes. Its job is more focused now:
Receive staged material.
Run validation.
Produce audit findings.
Name repair targets.
Suggest changes.
Send the result back to Writing Session.
That is enough.
The Forge handoff packet was also cleaned up.
Earlier tests could send duplicate scene text into Forge. That made validation noisy.
Now the packet is meant to carry one clean scene text plus context:
scene ID
episode
repo path
current scene text
scene card
tier trail
Captain notes when available
Forge judges the work.
It does not become the writing room.
The First Real Pass: EP203_A01S01
The real test scene was:
EP203_A01S01.
End of the Tour.
episodes/s02/e203/scene-cards/a01s01.md
The goal was to move the scene through T4 and prepare it for T5.
This is where the product model got tested under actual story pressure.
The T4 pass returned:
Holds With Strain.
Canon Risk: Low.
That was the right kind of result.
The scene worked, but Forge named the pressure points:
Jeb’s humiliation needed more playable behavior.
Passenger misread risked feeling too convenient.
Flight 19 needed stronger moral cost.
Carlos and Martin needed sharper complementary competence.
Linda should not be counted as active in A1S1.
Options Needed An Off-Ramp
The option system worked.
But it also revealed a risk.
Choose option.
Run another Forge pass.
Get new options.
Choose again.
Loop forever.
That is not a writing workflow. That is a validation trap.
So the choice panel moved higher on the Forge page.
Continue With Choice became Build Implementation Brief.
A visible selected-choice receipt was added.
Send To Writing Session moved beside the choice action.
The goal is simple:
Use Forge to decide a direction.
Then get out of Forge and back into implementation.
Notes Are Not The Scene
The biggest creative product insight of the day was about accountability.
The writer does not always want Gideon to apply the edit.
Sometimes the right workflow is:
Gideon validates.
Gideon names targets.
The writer applies the changes manually.
Gideon validates the revised page.
That protects authorship.
A patch brief is not the scene.
A log note is not the scene.
A validation output is not the scene.
The scene changes only when revised page text is intentionally approved.
That distinction became one of the most important product laws of the day.
For the T4 repair, the writer manually worked through changes around:
passenger unease
Carlos preserving signal
Martin covering the bleed
Jeb absorbing the moral cost
Linda being removed from active A1S1 presence
That is how Gideon should help.
Not by replacing the writer’s hand.
By sharpening the work the hand needs to do.
T5 Became A Dry Run Before The Human Rewrite
The user clarified the real T5 process.
At T5, they usually sit with the scene and do a full human voice rewrite.
Current version on one screen.
New version on the other.
Even if nothing changes, the rewrite matters because it returns the page to the writer’s body and voice.
So T5 was reframed for this test:
T5 dry run before the author’s human voice rewrite.
Not final lock.
Not seal.
Not replacement authorship.
The T5 dry run asks:
Is the T4-patched page ready for the human voice pass?
Where is the line drag?
Where is the screenplay too novelistic?
What should the writer protect during the rewrite?
The Save Bug That Had To Be Found
The T5 dry run found a serious product bug.
Forge was not receiving only the screenplay.
It was receiving the Writing Session wrapper too:
Edit.
Short Note.
Updated Active Scene Draft.
That meant the app had saved a diagnostic packet as if it were the scene page.
That is exactly the kind of bug a story workspace cannot tolerate.
Validation has to know what it is reading.
A scene page is not an edit packet.
A note is not a scene.
A validation output is not a source draft.
So the save path was tightened.
Scene Memory Needed Cleaner Semantics
The Active Scene drawer also got clearer.
At first, Current Scene Draft and Latest Saved Output showed the same thing.
Then Latest Saved Output disappeared entirely.
The right model became:
Current Scene Draft is the page.
Latest Saved Output is a brief receipt of what changed last time.
Scene Card is the structured scene memory.
Scene Edit History is the pass trail.
That is the shape.
The drawer now needs to help the writer answer four questions quickly:
What is the current page?
What changed last time?
What does the scene card believe?
What has the validation trail already decided?
That is scene memory.
Not a pile of outputs.
Linda Became A Retrieval Hygiene Test
The scene card had an important contradiction.
The new T4 decision said Linda was not active in A1S1.
But older scene-card language still implied Gus and Linda were monitoring together.
That is not just a copy issue.
That is retrieval hygiene.
If the scene card says Linda is active, future passes may route Linda into the scene when she does not belong there.
So the card was cleaned:
Gus remains radio pressure.
Linda is reserved for A1S2.
Linda is not active in A1S1.
The stale Linda tag was removed.
Small metadata fix.
Big continuity protection.
The lesson is bigger than Linda.
Scene cards are not decorative summaries.
They are routing surfaces.
If they drift, Gideon starts validating against a scene that does not exist.
Current Product State
By the end of May 10, Gideon had a clearer production flow:
Episodes holds the project tree.
A selected scene becomes the working container.
Writing Session discusses, plans, and applies.
Forge validates and names repair targets.
World Bible holds canon.
Log Book preserves decisions.
The tested EP203 scene is post-T4 repair and ready for a clean T5 dry run after confirming Forge receives only the clean scene text.
The important product truth stayed steady:
Forge validates.
Writing Session decides.
Episodes remains the scene home.
The writer owns the page.
Remaining Watch Points
The next refinements are clear:
Confirm the Active Scene drawer shows the full current draft and a compact latest-save receipt.
Rerun T5 only after confirming Forge receives clean scene text.
Make Approve Scene preview exactly what will be saved.
Prevent Scratch Only output from becoming scene text without explicit intent.
Make Scene Card updates first-class instead of bundling them into scene text packets.
Add a stronger distinction between Log Note, Validation Packet, Edit Brief, Scene Version, and Scene Card Update.
Consider renaming Latest Saved Output to Last Approved Change or Latest Save Receipt.
Add a Ready For Human Rewrite state between T4 and the final T5/T6 path.
Product Truth
May 10 made the scene room real.
Not perfect.
Not finished.
But real.
Gideon is becoming less like a chatbot wrapped around files and more like a production bay for story work.
The page stays the page.
The notes stay notes.
The validation stays judgment.
The writer makes the call.
That is the work.
