May 19, 2026
Gideon Build Log: Trust Has To Be Visible Before The Run
Today started with the kind of small mistake that can break trust in a writing system.
EP201_A01S02 was clearly an EP201 scene. The repo said so. The scene label said so. The card title said so.
But part of the metadata was still pointing at EP203.
The headline:
Gideon moved from smart output to explicit trust infrastructure.
The bigger product shift:
Canon validation now has visible receipts before the writer runs the pass.
Repo cards
179/179 scene cards route cleanly.
Saved records
28/28 workspace records match the episode library.
T3
World, relic, myth, source lore. No character or season-planning default.
T4
Character and relationship consequence. No relic or season-planning default.
T5
Season planning returns for before/after continuity safety.
The Product Lesson
The strongest lesson from the session was not about prose.
It was about trust.
Trust cannot live only inside the model prompt. It has to live in the product contract, the source router, the UI, the tests, and the recording surface.
That means Gideon needs to show what it is about to use before it asks the writer to trust a Forge result.
The new pattern is simple:
check the routes
preview the sources
confirm the tier boundary
then run the pass
That is creative infrastructure with receipts.
What Each Pass Is Allowed To See
1. Trust Panel
The first recording target was the Trust Panel.
This is the cleanest possible proof clip because it does not ask the viewer to believe a model response. It shows the workspace checking itself.
The final receipt:
– 179/179 repo cards clean
– 28/28 workspace records clean
– T3 boundary visible
– T4 boundary visible
– T5 continuity safety visible
Before the writer runs Forge, Gideon can show whether the workspace routes are clean.
Trust Panel shows clean repo-card routing, saved workspace records, and tier source boundaries before any Forge pass runs.
2. Source Preview
The second recording target was Source Preview.
This proof used EP201_A01S02, the exact scene that exposed the trust problem at the beginning of the session.
The preview was staged as:
T3 / Gulp It Deep
Deep source depth
Magic, relics, and world mechanics only
The final source receipt showed:
– Core Canon
– Deep Myth
– Gideon System
– Relics
– Source Lore
– Source Gaps: 0
– No save
T3 now proves it is looking for world mechanics, not character files or season planning by default.
Source Preview shows the T3 source packet for EP201_A01S02 before a Forge run, with zero source gaps and no save action.
3. Forge Handoff Cleanup
The recordings also exposed one more real problem.
When EP201_A01S02 was sent to Forge, the Forge Result panel still showed the last session’s output.
That is a trust break.
A new Forge handoff should feel like a clean workbench. It should not carry yesterday’s repair draft into today’s scene.
The fix now clears:
previous Forge result
previous repair draft
selected choice
choice panel
source preview
gate badge
copy buttons
validate button
cached local repair draft
Then a Playwright regression test was added to keep it fixed.
What Got Better
Gideon became stricter about scene identity. EP201 material should not route through EP203 metadata.
It became clearer about Forge tiers. T3 is now world mechanics, relics, deep myth, source lore, and core canon. T4 is character and relationship consequence. T5 gets season planning again for before/after continuity.
It became more visible. Trust Panel and Source Preview now live in Utilities as pre-run proof tools.
It became more testable. Playwright is installed, browser tests are active, and the stale Forge output issue has a regression test.
Most importantly, it became more honest. The system now has a way to show when the workspace is clean instead of asking the writer to assume it is.
The Trust Stack
The Honest Rough Edge
The session did not start from inspiration.
It started from distrust.
That was useful.
A creative tool for a real show cannot be treated like a toy. If it routes a clearly labeled scene to the wrong episode, or if it shows old Forge output when a new scene is staged, the writer should be irritated.
That irritation turned into better product architecture.
The camera did not just record the product. It helped test the product.
Product Truth
Gideon is strongest when it keeps the writer in judgment and makes the system work visible.
Today proved a sharper version of that idea.
It can check routing. It can preview sources. It can enforce tier boundaries. It can clear stale output. It can test the UI in a browser.
The direction is plain:
A writing system earns trust by showing its receipts before it asks for the scene.
