Ernie.SG Research programme and work in progress
REPResenting Taste: Composable Design Decisions Across Alternatives and Revisions
Investigating representations of creative decisions that are replayable, editable, parallelisable, and remixable.
Problem
Much of creative production is a grind of adaptation. Content changes. Collaborators bring different versions. Work moves between formats and audiences. Each change can mean repeating adjustments, reconstructing earlier choices, and repairing their knock-on effects.
Workflows built around sequential edits make these operations cumbersome and repetitive. The judgments behind the work remain difficult to isolate, reuse or revise.
Approach
Develop software that represents content, creative decisions and their dependencies as explicit, editable structures. A decision records what it applies to, its scope, the conditions under which it remains valid, and its relationships to other decisions. The model is extensible with new decision types and operations.
Struct supplies the document layer: structured content, assets, relationships, and reading renditions. The decision layer is application-owned and proposed; it does not make Struct responsible for creative policy or revisions. It is not a universal Struct schema or a competing canonical editorial store.
- Document Content and assets
- Decisions Scope and dependencies
- Operations Replay, edit, fork, remix
- Renditions Digital and offline editions
A typed API would apply operations with preconditions, validity checks, and conflict diagnostics. People and agents can then share a state, develop variants independently, combine accepted decisions, and check what remains applicable after change.
| Capability | Software method |
|---|---|
| Replay | Record operations against a known state and check validity. |
| Edit | Inspect and change content, decisions, scope, and dependencies. |
| Parallel | Fork shared state for independent variants and contributions. |
| Remix | Compose selected decisions and expose interactions and conflicts. |
Proposed evaluation
Matched tasks will compare this workflow with a strong native workflow using the same content briefs, model access, and renderer. The primary measure is total active editing time: capture, composition, adaptation, checking, and repair. The proposed initial target is at least 20% less active work, with retained decisions, completion, and blind output quality as secondary measures. A five-task pilot precedes a counterbalanced editor study.
Why Now
Replay, branching, and remixing existed before large language models. Autonomous creative work now needs explicit shared state, accepted decisions, and dependencies so people and agents can inspect, combine, and check changes. REPResenting Taste investigates those foundations.
Outcome
An AI-native magazine, as one application of the representation and its operations, that can be created and revised collaboratively, then published across digital and offline formats.
Milestones
Running: Struct’s PDF-import and responsive-reading demo.
Planned: a shared canonical representation for content, creative decisions, and dependencies across forms; a five-task pilot; a complete issue with Margin; held-out editor and reader studies; then a REPResenting Taste research-paper submission.