Glyph generation

The glyph is the visual condensation of a descent — a 32-wide, 16-tall ASCII image rendered fully on-chain in the contract’s tokenURI. ASCII glyphs have high quality variance from Kimi: sometimes the model produces a varied breathing composition, sometimes it falls back to tiled wallpaper. The pipeline is built around that variance.

Pipeline

1. Three candidates in parallel

Three concurrent calls at temperatures 0.7 / 0.9 / 1.0. Distinct enough that the candidates actually diverge instead of collapsing to similar samples. We tried 0.6 / 0.7 / 0.8 first; the candidates were nearly identical. The current spread sacrifices some “safe” outputs at 1.0 in exchange for diversity.

2. Each candidate gets a structured prompt

Four parts:

• Output format: 32 dashes (top boundary), 16 rows of glyph (≤32 chars each), 32 dashes (bottom boundary). No header, no explanation, no markdown.
• Character set: . - = + * # / \ | < > : ~ only. No letters, no numbers.
• Composition rules: must have a silhouette (rows have varied widths — think sculpture / rune / hieroglyph, not textured rectangle); plenty of negative space; every row unique; glyph evolves from top (descent surface) to bottom (deep end); each of 5 vertical bands uses its own derived character palette.
• Negative + positive examples inline — explicit “this is wallpaper, do not do this” and “this is what we want”, so the model has visual anchors instead of just prose rules.

3. Structured hints alongside the raw transcript

Two pieces of structure feed the prompt:

Band palettes

The descent is split into 5 vertical bands. Each band’s character palette is the union of TYPE_GLYPHS for the layer types that fell into it:

Type	Glyph chars
drone	`#
pulse	+ =
glitch	/ \ *
texture	. -
breath	- = < >
bell	* < >
drip	. :
swell	~ - =
chord	`

So a band heavy in drone + chord produces a vertical-stack-flavoured palette (# | = :), while a band of glitch + drip + texture produces a sparse jagged palette (/ \ * . :). The shape of the descent literally shapes the glyph’s character distribution.

Poetic intent

Hermes’s per-layer comments are lifted out of the transcript and listed with their move number. The model treats these as form-shaping images — breath / exhale → soft (. - =), glitch / fracture → jagged (/ \ *), drone / floor → solid (# |), drip / water → isolated marks (. :).

4. Each candidate is scored, highest wins

scoreGlyph() rewards visual variation and silhouette shape, penalises tile-like wallpaper:

score = filledRows       * 1.0    // rows with ≥4 non-space chars (capped at 12)
      + uniqueFilled     * 0.4    // distinct filled rows
      + entropy          * 2.5    // Shannon entropy over char distribution
      + densityStdDev    * 1.5    // std-dev of fill-density across rows
      + silhouetteStdDev * 2.0    // std-dev of trimmed row WIDTH
      - tilePenalty      * 1.8    // repeating-segment runs ("<>.<>.<>." etc.)

The silhouette term carries the heaviest weight because it’s the strongest signal for “is this a glyph with a shape vs a rectangle of textured content”. A solid block scores 0 on silhouette no matter how varied its characters; a diamond / hourglass / asymmetric silhouette scores high.

5. Extraction + normalization

extractGlyph() does the post-processing:

1. Looks for the dashed boundary lines. Tolerates = or # substitutions models occasionally produce instead of -.
2. If no boundary is found, falls back to the longest contiguous block of allowed characters in the response.
3. Each row is sanitised — anything outside the glyph charset becomes a space.
4. Rows are trimmed and either centre-padded (if short) or windowed to their densest 32-char span (if too wide).
5. Always 16 rows out, padded with empty rows if the model under-produced.

6. Fallback

If KIMI_API_KEY is unset or all three calls fail, fallback() produces a deterministic glyph by tiling per-type characters across the 32×16 grid in placement order. Plus the journal stub. The descent always ends gracefully.

Why best-of-3, not best-of-1

We tried single-call generation with various prompts and seed strategies for two days. The hit rate for “actually a glyph, not wallpaper” hovered around 60–70%. Best-of-3 with structural scoring brings it above 95% — the worst output of three is rarely chosen, and when all three are weak the scoring still picks the most varied of the bunch.

The cost is three API calls per descent instead of one. At Kimi’s pricing this is negligible; at the player-experience level it’s the difference between “sometimes the glyph is amazing” and “the glyph is reliably interesting”.