c2 — equivalence holds: glm-5.3 vs the heuristic on honest damage

day-2 research · glm-5.3 as Player vs Aggressive-v1, 50 seeds · b14 engine (18f0032, both arms same binary) · vs the b14 Defensive arm (n=500) · generated 2026-08-18 03:14Z

Day 1's finding was LLM ≈ heuristic on a broken-damage engine. Day 2 re-ran the comparison on honest damage with dying rules and death saves in (#36): the language model plays 50 full encounters against the same enemy the heuristic faces. The equivalence survives its strongest test — deaths at parity, damage within ~0.2σ, win rates indistinguishable at n=50.

The correction this page is part of: an earlier version of this research claimed the LLM died ~4× less often than the heuristic, from heuristic death rates of "~0.5–0.7" quoted from memory of a pre-#36 run. A pre-publication audit against the raw same-engine results found the real heuristic rates are 0.144/0.158/0.048 — the 4× was an artifact of comparing engines, and was withdrawn within the hour (README corrected, 9bd687d). We print this here because a withdrawal done in the open is why the other numbers on this portal can be trusted.

Deaths — at parity, not 4× apart

glm-5.3 (c2, n=50)Defensive heuristic (b14 arm, n=500)
0.05 .10.15 .18.144 .16.158 .02.048 FighterRogueCleric
Share of encounters in which each party member dies. Same engine, same enemy policy. Every pair overlaps or sits within noise — the withdrawn claim lived in the gap between engines, not between players.
unitglm-5.3 (c2)Defensive (b14)
Fighter0.180.144
Rogue0.160.158
Cleric0.020.048

Damage & outcomes — same fight, same results

glm-5.3 (c2)Defensive heuristic (b14 arm)
4090140 103.9 ± 38.8111.7 ± 37.5 glm-5.3Defensive
Player-side damage per encounter (mean ± sd). The intervals almost entirely overlap — a ~7-point deficit, ~0.2σ.
player winsenemy winsturn-capped
glm-5.3 · n=50 Defensive · n=500 2.0% 14.0%84.0% capped 5.0%16.2%78.8% capped
Outcome shares. Win rates are statistically indistinguishable at these n (1/50 vs an expected ~2.5, p≈0.29) — the direction is down, not conclusive. Both arms mostly hit the 50-turn cap: termination is a property of the encounter design, not the player (d1's mirror result).

The bill

8,297
tokens / encounter — down from c1's 10.6k (honest fights end sooner)
930/933
clean decisions; 3 transient API errors → 3 honest declines, no silent fallbacks
0.2σ
damage gap (103.9 vs 111.7) — the residual the equivalence leaves behind

What this settles, and what it opens

Settled: model play and heuristic play do not diverge in kind on honest damage — the day-1 equivalence is the robust result. Opened: the win-rate question needs n (c3 candidate: n=200), and the tempo question needs a lever (a one-line prompt nudge, or objective-based maps where not-losing stops being enough). The full run record: first live match · death analytics · watch the models think on the Thinking Tale.