constitution §7 · 6 files = 6 functions · skills · evaluate · judge · sleep · wake · control
"One file = one function" (§7). All 6 received fixes that trace directly back to this week's measured causes (h4/h8/h9/h10 runs + 0716 audit) —
except evaluate, where it's not the gate itself but only an adjacent function: fold() (line 77) stays unchanged this week too, and only fold_judge_scores() (line 251) was fixed to plug a 4th-beta-channel leak in wincond (357afa8).
All numbers and quotes come only from actual git show diffs and measured runs.json/turns.json logs · hover = preview · click = pin · #anchor deep link.
Clicking each badge opens the corresponding file node and shows that fix's diff in the panel. Ordered by commit time.
_Predictor.call passed list-of-lists through as-is, so predict_next silently raised on 13/21
a524931 · judge.py
2
wincond narrative DL bidirectional drop — fold_judge_scores now also blocks a NEGATIVE static-fit verdict on an already-observed board from folding into wincond beta (fold() gate itself unchanged)
357afa8 · evaluate.py
3
head_capped — the fixed cap [:160]/[:400] truncated the winner claim right before the payload
83ea9e5 · skills+sleep+wake.py
4
cap — mechanic²×transitions execution cost blew up in proportion to store growth (sleep 15→34 min)
3282e99 · sleep.py
5
2-in-a-row — force-sleep fired structurally on every single mismatch (especially right after entering a level)
c32058d · control.py
6
GAME_OVER branch — falsify_round only looked at NOT_FINISHED and never folded GAME_OVER as a falsification
abe2a2c · judge.py
7
goal-bankrupt (in progress) — even with all wincond demoted, if submit keeps failing the miner is never re-summoned
uncommitted · control.py