Home / Ops / More in this area

The Wheel - Star Architecture + Per-Constellation Orientation Spec (v1)

Updated Jul 12, 2026 · Affirmology_WheelArt_OrientationSpec_v1.md

Summary. Jeff's live review, 2026-07-12. This is the build-critical doc for the wheel's visuals. It answers "how do the stars work" and records the orientation fix per constellation. Feeds the code build.

The Wheel - Star Architecture + Per-Constellation Orientation Spec (v1)

Jeff's live review, 2026-07-12. This is the build-critical doc for the wheel's visuals. It answers "how do the stars work" and records the orientation fix per constellation. Feeds the code build.

THE ARCHITECTURE ANSWER (the key decision, and it saves credits)

The interactive constellation is TWO LAYERS. This is already the plan's two-layer pipeline; here it is made concrete for the stars.

LAYER 2 = the STARS, LINES, LABELS, FILLS: 100% APP CODE. Never Higgsfield. - Each star is ONE code component (SVG/Canvas/Skia) with a FILL PARAMETER 0..1, not a stack of separate images. Empty = thin gold outline + faint inner shimmer (variant 3). Filling = the vessel fills with light across 7 fractions (parameter animates). Full = solid equal glow. One component, every state, animatable, precise. - Star POSITIONS are set by us (real catalog positions where we want fidelity, or hand-placed anchors), identical every render, never invented by an image model. This is what makes the sky astronomically honest. - LINES between stars = code, drawn between anchor points; brightness follows the connected star's fill (the line rule). LABELS = code text near each anchor; on hover/tap the label AND its star outline light up so it is unmistakable which audio the star is. FILL-ON-PROGRESS = code. - STAR SIZE = code, uniform within a constellation, and SCALED BY COUNT: few stars (Libra, Aries) render slightly larger; many stars (Scorpio, Aquarius) render smaller so the sky never crowds. (This auto-fixes Jeff's "Libra stars are too big" note; sizing was an artifact of the image, and images no longer own the stars.)

LAYER 1 = the FIGURE + SKY: Higgsfield art, and it is its OWN movable layer. - Higgsfield supplies the painted silhouette figure + nebula/atmosphere ONLY. Any stars it paints are decorative mood at most and are IGNORED for gameplay (the code stars sit on top at true positions). - Because the figure is its own layer, the APP can ROTATE, FLIP, and SCALE it for free (a transform) to sit correctly behind the fixed code stars. This means most orientation fixes are CODE TRANSFORMS, not regenerations. We only re-render art when the POSE CLASS is wrong (e.g. a standing figure that must recline), not for a tilt.

Consequence for credits: we do NOT regenerate 12 constellations to chase star accuracy. Stars are code. Figures we already have; the app rotates/flips them to fit. Near-zero new Higgsfield spend to land the whole wheel.

THE V2 CORRECTION (Jeff's flag, logged)

Several v2 images (Sag, Capricorn, Aquarius, Pisces, and the earlier reclining/left variants) still show INVENTED star arrangements. That was the model rearranging stars to flatter the shape, which is the thing we forbid for our astrology-literate audience. Those v2 images are NOT star references. They are only useful as FIGURE-pose references. Stars come from code; ignore every image's stars.

PER-CONSTELLATION VERDICTS (Jeff, 7/12)

Format: verdict | fix | fix-type (CODE transform vs REGEN art vs KEEP).

"DOES IT HARM ANYTHING TO ROTATE THE STARS SLIGHTLY?" (Jeff's question)

Small answer: a SMALL rotation of a whole constellation is HONEST, because the real sky presents every constellation at every rotation through the night and year (only relative positions are fixed, never the tilt). So rotating the entire star pattern a few degrees to sit better in the window is legitimate. What we never do is move stars RELATIVE to each other (that is the forbidden rearrangement). So: rotate the whole pattern freely; keep the internal geometry exact. In practice we will usually rotate the FIGURE to the stars (cheaper, and the stars stay put as the person's progress anchors), but rotating the whole star cluster a few degrees is also allowed when it reads better.

WHAT GOES TO CODE (add to the pronunciation-era build brief)

  1. Stars/lines/labels/fills = code, one star component with a 0..1 fill param, positions from a per-constellation anchor list (starAnchors in the wheel manifest), uniform count-scaled sizing.
  2. Figure + sky = R2 image per constellation, rendered on its own layer with app-controlled rotation/flip/scale so orientation is tuned in code (and A/B-able) without new art.
  3. Hover/tap lights the label + the star outline together.
  4. Blank-sky transition varies its region/offset so rotation never feels identical.
  5. Reduced-motion stills everything.