PART B: the preview. Jahni, asked what "good" means for ranking a season:
"good is hard to describe, i believe it'll be easier once i'm able to see
the strates candidates." So the ranking policy was blocked on this, not on
more metrics. COMPOSER-NOTES §3.4 promised the preview comes nearly free
once the measurement exists; collected.
Per candidate, rendered from the grid the measurement ALREADY built (no
second sampling of the world): a vertical XZ section and a plan XY slice,
solid dark, air light, walkable cells highlighted, with a chunk scale bar
so a corridor is distinguishable from a cathedral. Plus a self-contained
index.html contact sheet of all 64 -- survivors first, outliers first
within them, recipes, metrics, and every rejected candidate labelled with
WHY (vacuous 3, fragmented 2, law failed 19).
Saved/ComposerPreview/<runid>/index.html 64 candidates, 128 PNGs
Two honesty rules are on the page itself, not buried in a comment: every
number states the window that produced it (margin/step/Z span), and a
2D slice CANNOT show connectivity -- two regions that look joined in one
slice may not be.
Plan-Z is chosen by highest solid/air boundary-transition count, not
hardcoded mid-height; a slab strate would otherwise render as a blank
sheet. Codex flagged the ambiguity in my spec and picked correctly.
PART A: bootstrap jitter for zero-spread fields. ±25% of a field's own
magnitude when corpus spread is ~0, ±15% of measured spread otherwise,
clamps last. Marked in code as a BOOTSTRAP for a starved corpus that
promotion (Tier 5) should retire. 1,422 applications over 168 fields in
the parameter roll; 6,848 over 107 in the structure roll.
⚠️ AND IT DID NOT BROADEN THE DISTRIBUTIONS. Feature scale 0-376 either
way; walkable actually NARROWED (.389 -> .287). Codex's conclusion, and I
agree: the measurement is INSENSITIVE to these fields, not merely
corpus-starved.
The reason is now obvious and worth writing down: the grid samples every
4 voxels, and cave DETAIL -- terraces, scallops, ribbing -- lives at 1-3
voxels. It is below the sampling resolution. A step-4 grid structurally
cannot see feel; it sees size. So "vary the detail" and "score the detail"
need different instruments, and the preview images inherit the same limit.
Box proofs stayed sound under wider jitter: parameter roll 1,204 proved /
1,602,524 voxels / 0 violations; structure roll 659 proved / 877,129 /
0 violations. §6.2 did not reopen.
fable-idea.md F1 ticked (measurement + preview).
Build clean, 25 succeeded / 0 failed, verified by me.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VJiTcTSDNkKBtnNdiBtdaU
Fr0zkaandClaude Opus 5
993512f0bc