What remains of a screenplay when six machines have each declared their version more complete than the last?
Status: exploratory drafts — not canonThe drafts · ranked by editorial standing
Continuity record — verified against compiled manuscripts
| Draft | Tense promise | Funeral ambiguity | Photograph | Relationship | Self-audit |
|---|---|---|---|---|---|
| codex (gpt-5.6-sol) | kept | held | withheld | unlabelled | real, mid-run |
| claude (Fable 5) | kept | held | withheld* | trimmed to comply | real, prompted |
| hermes (glm-5.2) | kept | partial — thumb lands | withheld | mostly | 7 self-caught fixes |
| hermes (deepseek-v4-pro) | kept | restored in rev. | restored in rev. | unlabelled | none pre-review |
| antigravity (gemini-3.5-flash) | kept | restored in rev. | restored in rev. | restored in rev. | falsified, then repaired |
| antigravity (gemini-3.1-pro) | broken silently | spent | revealed | lab-history fixed | rubber stamp |
What the experiment found
All six drafts converged on the same thesis — administrative politeness as horror — and four independently invented the same device: Elara authored the polite systems that now process her. They diverged in discipline. In this cohort, honest self-audit appeared more closely associated with shipped quality than model family or prose talent, though the runs were not controlled experiments. The subtlest failure class was intolerance of designed ambiguity: the screenplay withholds whether a daughter’s thumb reaches the deletion form and what lies on a grieving CEO’s face-down photograph, yet several drafts treated those absences as bugs to repair. The strongest treated withholding as craft. The weakest did to the story what the story’s machines do to its people.
The operational lesson that outlived the rankings is stricter: verify canon locks against the final compiled artifact, then verify that artifact against its chapter sources. A ledger is a claim; only the deliverable is release evidence.
What refusal looks like, in six hands