yysworld
YYBW-010·Decided·2026-04-13

Compute Strategy: Simulate Many Cheaply, Render Few Richly

depends on YYBW-007, YYBW-009

  • ID: YYBW-010
  • Status: Decided
  • Date: 2026-04-13
  • Scope: YY Branching World / YY's World evolution
  • Depends on: YYBW-007, YYBW-009
  • Supersedes / clarifies: Clarifies use of local models, premium models, and media cost control
  • Museum lineage: SK-006

Context

The project repeatedly confronted cost. The core insight was that pictures and clever text are the expensive parts, while structured state transitions are cheap. This led to a layered compute strategy: local models or rules for deterministic tasks, stronger hosted models for premium surface moments, and selective media generation.

Decision

Use a hybrid compute model:

  • deterministic code + local models for routine structured updates
  • premium hosted models for selected narrative moments
  • selective, milestone-based media rather than rich generation everywhere

Why

This matches the economics of the product:

  • state is cheap
  • narrative polish is expensive
  • images/video should mark significance, not fill space

Alternatives considered

  1. All premium model, everywhere — rejected as too expensive.
  2. All local models, including premium prose — rejected for quality risk.
  3. Full image generation for every node — rejected immediately on cost grounds.

Reversals / scars preserved

  • The conversation explicitly recognized that kids want pictures and clever text, but those are the highest-cost surfaces.
  • SNES/SMS format became a direct answer to this cost pressure.
  • Local-vs-remote is not ideology here; it is economic discipline.

Consequences

  • Need structured schemas for local model output.
  • Need triage logic for which branches/events get premium passes.
  • Need restraint in media design: mostly sprites, composables, or lightweight visuals.

Invariants preserved

Compression and Discipline are the main guards; Explainability improves when structured state is primary; Survivability improves because prose is view-layer, not truth.

Freshness boundary

Revisit when cost/traffic data becomes real.