M2 capstone — the difficulty curve reproduces: Keep on the Shadowfell

the project's first designed experiment · zones: 3 × 50 seeds (06:43Z) · areas: all 19 × 30 (15:41Z, full power) · level-1 five-PC party, both sides Aggressive

The question behind this benchmark was the project's biggest: does the honest engine reproduce the difficulty the books designed? Keep on the Shadowfell is the classic published adventure — a village tutorial outside, a keep that punishes under-leveled parties inside. We assembled each zone's full encounter and ran it 50 times. The answer on first contact: yes, and the ordering is monotone in XP. At 450 XP the level-1 party wins comfortably; at 2,650–3,850 XP — eight times the threat — it wipes, and the wipe rate orders with XP exactly as the book intends: the village is the tutorial, the keep demands leveling.

Confirmed, not just reproduced. The declared re-run trigger fired the same day — a 3-of-5 hero re-stat plus an action-delivery engine change. All 150 re-run encounters match the first pass to the outcome vector: 37/0/13 · 0/24/26 · 0/32/18 — 74% / 48% / 64%, identical. The monotone-in-XP curve is doubly pinned.

74%

player wins in the village (37/50) — the tutorial is winnable at level 1

0% → 64%

wipe rate, monotone in XP: village → deep → upper

8.6×

XP threat, village 450 → upper 3,850 — one fixed level-1 party against all of it

Outcomes by zone — the curve

player winswipes (enemy wins)turn-capped

Outcome shares per zone (50 seeds each, both sides Aggressive), ordered by XP budget. Zero-width segments are absences, not rounding: the village records no wipes at all; neither keep zone records a single level-1 player win.

zone	XP	monsters	player wins	wipes	capped	rounds (mean)
village	450^*	5	37 (74.0%)	0 (0%)	13 (26%)	10.5
shadowfell-keep-deep	2,650	11	0 (0%)	24 (48.0%)	26 (52%)	8.1
shadowfell-keep-upper	3,850	16	0 (0%)	32 (64.0%)	18 (36%)	9.3

^* as-summed at run time: the manifest's 450 covered the four statted kobolds; the Orcus Underpriest (level 5) also fielded in the village carried null XP in the manifest — its unit data says 200. The pre-build audit caught the gap; the manifest was corrected the same day (2b06191: underpriest 200, village budget 650). The encounter that fought never changed — only the accounting did. The spread with corrected budgets: 650 → 2,650 → 3,850, ordering untouched.

The full 19-area curve — the fifth pin

The zone benchmark's caveat was "zones aggregate their areas." The OCR and roster passes landed, and the complete pre-registered test ran: all 19 areas, n=30 each, full power (66/66 monster refs resolve). The shape is the book — early delve areas near-100% for the level-1 party, a mid-band grind, and the deep keep at 0–10%, death for the unleveled, exactly as published.

ρ = +0.873

Spearman XP-vs-difficulty across the 15 XP-carrying areas — H1 confirmed strongly (registered threshold: > 0.4)

H2 ✓

every XP ≤ 400 area ≥ 36.7% win; every XP ≥ 800 area ≤ 30% — band checks pass, no span anomalies

10/105

pairwise inversions — all between identical-XP areas (325 vs 325, 175 vs 175): tie noise, not curve violations

Player win rate vs XP budget, all 19 areas (30 seeds each; the four zero-XP skeleton-family areas run but sit outside the stats). Win rates re-derived from the per-area results.json before this chart — all 19 match the README exactly; XP values are the assembly's own budgets.

Two honest flips, both printed. The first analysis pass ranked win-rate and printed ρ = −0.873 — the registration defines difficulty as 1 − win; the sign was self-caught and corrected before publication. And the preliminary run's H3 failure (a span anomaly) did not survive the drift fix plus full data — the flip is reported in both directions. Only one span remains noteworthy: area 12 wins 36.7% where its identical-XP neighbor 17 wins 60% — flagged for the content lane, not hidden from the curve.

Damage — how each zone feels

party damage / encmonster damage / enc

Damage dealt per encounter by faction (mean ± sd over 50 seeds). The feel of each zone in one glance: the village is a rout (166 v 37), keep-deep is mutual attrition (109 v 105), keep-upper leans on the party (138 v 115) — with single-seed party damage peaking at 259, enough to chew through a full level-1 party's hit points.

Two method finds the benchmark surfaced

1. The 50-turn cap was hiding the entire result. The keep zones field 16–21 units; a fixed 50-turn limit resolved almost nothing — the first pass came back 100% capped, a wall of "no verdict". Scaling the cap to unit count (10u→120, 16u→192, 21u→252) let the fights finish, and the curve appeared. A benchmark that measures nothing but its own ceiling isn't a benchmark.

2. Encounter mode must read the assembled map, not the manifest rows. Faction joins in `--encounter` initially fired against the raw manifest lists; once they resolve from the assembled map, both sides field who actually stands on the board. Small harness bug, total difference to every result.

What this does not claim — and what unlocks it

Zones aggregate their areas. The book intends per-area encounters (9–7 areas per delve); the manifests carry area lists but not per-area rosters yet. Zone-level is the coarsest honest cut — the per-area pass (OCR/render of the roster pages) will move every number and is the real curve test. The party is fixed at level 1 across all zones: this is each zone's un-leveled difficulty. The book-true comparison — a party leveled per delve — needs the level-2/3 five-PC sets, which is content-lane work the PC builder (#46) now feeds. Terrain stays off at this phase (standard-fan placement, variant overrides informational).