The Testing Phase¶

When you start a Lvl 3 or Lvl 4 variant, it enters a testing phase under the next major's -0-N suffix. You stay in testing until the variant proves itself, then graduate to the new major version.

What the testing phase looks like¶

flowchart LR
    A[V1 approved] --> B[V2-0-1<br/>first build]
    B --> C{Looks good?}
    C -->|no| D[Iterate]
    D --> E[V2-0-2]
    E --> F{Looks good?}
    F -->|no| G[More iteration]
    G --> H[V2-0-N]
    H --> I{Approved?}
    I -->|yes| J[V2 in approved/]
    F -->|yes| I
    C -->|yes| I

Each iteration bumps the -0-N suffix. The previous file moves to backups/. The new file lands in testing/. No upper bound — you can be at V2-0-12 if that's what it takes.

Why testing matters¶

Lvl 3-4 variants change things you can't reliably predict the outcome of:

New environment lighting might wash out the avatar
New camera angle might not flatter the subject
Split scenes might pace weirdly
New prehooks might feel disconnected from the body

You need to see the output before approving. Lvl 1-2 variants are predictable enough to skip this; Lvl 3-4 aren't.

What you do during testing¶

Every iteration follows the same loop:

Run the Generation Runner on the current V{N+1}-0-N.nbflow
Review in PatchWork — look at every scene's candidates
Identify what's working and what isn't
Make targeted changes to fix the issues
Bump the version (-0-N → -0-N+1)
Repeat until clean

Common testing-phase issues¶

What you'll typically encounter and how to handle each:

Issue	Fix
AI tells in 1-2 scenes (extra fingers etc.)	Let the Generation Runner's auto-rerun handle it (3 attempts default; bump to 10 for stubborn ones)
Lighting wrong in multiple scenes	Edit the image prompts to be more explicit about lighting; bump V0-N; rerun
Composition slightly off from intent	Use the prompt-tuning skill on the stubborn scenes
Pose / motion feels stiff	Edit the Veo prompt's motion description (macro motion, see Video Prompt Rules)
One scene structurally won't work	Reconsider the storyboard — maybe split the scene, change the framing, or drop it
Avatar's face drifting	Lock the avatar reference; if already wired, check the per-node link refs are clean

Iterating efficiently¶

Don't bump V0-N for every tiny change. Batch related changes:

Bad pattern: Fix Scene 01 → bump → fix Scene 02 → bump → fix Scene 03 → bump...
Good pattern: Identify all issues across all scenes → edit all the prompts → bump once → rerun

The Generation Runner reuses cached output for unchanged nodes, so a single rerun after batching is cheap.

Cost during testing¶

Testing iterations burn budget. Strategies to keep cost low:

Use mode=images for prompt validation: image-only passes catch most issues cheaply. Only commit to full video gen on iterations you think are close
Re-roll seeds before editing prompts: cheaper than editing-and-regenerating; sometimes a fresh seed is what you needed
Test on one account first: don't testing-phase across 5 accounts in parallel. Get the test account clean, then fan out
Cap aggressive regen attempts: 3 attempts default is fine for early iterations; bump only for stubborn scenes

See Chapter 13 — Workflow Optimization for the full cost playbook.

When to stay in testing vs. graduate¶

Graduate (drop the -0-N suffix, move to approved/) when:

3 of 4 candidates per scene are acceptable
No AI tells the user is willing to accept
The user explicitly approves the variant

Stay in testing when:

Multiple scenes have issues that haven't been resolved
A scene needs prompt-tuning that hasn't been done yet
The user hasn't reviewed yet

You can be at -0-10 if the variant is genuinely hard. That's fine. The version registry's history list tells the story.

Multi-account testing¶

If you're going to fan out the variant to multiple accounts, the standard approach:

Test on one account (typically the canonical test account)
Iterate until clean on that account
Graduate to V{N+1} in approved/
Then fan out to the other accounts (this is its own bump, but the structure is locked)

Fan-out is part of testing too — see Chapter 7 — Fan-out Protocol. The fan-out itself is a meaningful change and bumps the version.

Documenting changes in the version registry¶

Every -0-N bump appends a history entry to the version registry. Write specific notes, not vague ones:

Good: "V2-0-2: Scene 03 lighting prompt edited to add soft fill from camera-front, Scene 05 reframed wider"
Bad: "V2-0-2: tweaks"

When you (or someone else) comes back to this workflow in 3 months, the history list is what tells you what was tried.

When you're ready¶

→ Next: Iterating with Prompt Tuning — using the prompt-tuning skill during testing for stubborn scenes that won't land with normal regens.