The Delve Library — adventures through the engine

The portal's centerpiece: every published adventure in the corpus, run through the engine as written — canonical parties, win rates across simulations, and example delves you can watch fight by fight. The M3 question delves answer: what does an adventure do to a party that fights it as written? A chain runs encounter after encounter with state carried between fights — HP, healing surges, dailies persist; there is no full reset at the dungeon door. The strip below is the format every delve tale will use: one card per encounter, and between them the carried-state trail — the attrition line that says whether the party is holding, spending, or dying on its way down.

Every 100% here was a lower bound while 39% of monster specs carried no abilities — the backfill tranches are re-pricing the corpus book by book, and those current readings live on the difficulty ledger.

Adventures in the corpus

Authored delves — from the editor

The first chain experiment — the composition gap contracts

Four parties (a 5-PC delve control; martial, caster, and mixed trios) ran the same five-link winnable ladder — KotSF areas 1 → 2 → 3 → 14 → approach-and-village, XP 350–850, top-off rest, 50 chains each, 1,000 encounters total, pre-registered before any run.

DELVE control (5-PC) MIXED trio CASTER trio MARTIAL trio

Party HP% carried between fights (mean over 50 chains; the pre-fight value at each link). The control holds ~95 flat — surge economy holds. Mixed and caster decline 93→57 despite top-off rest: the chain spends what single encounters don't. Martial sits at 100 flat for a different reason — it heals to full and still loses 78% of chains on the fights themselves.

Two honest limits, printed with the result. Zero daily powers were spent across all 100 heuristic-played chains — the Aggressive policy never picks dailies, so the compounding mechanism this experiment went looking for (daily exhaustion) cannot manifest under heuristic players; this run measured surge economy only. And the registration's own outcome tiles were badly designed (the wash band overlapped the linear band — the analyst's flaw, not the engine's). The gap between martial and caster contracted to 8 points; LLM players who spend dailies are the next probe.

The duel format — sketch (P-B, awaiting first scrimmage)

Two parties, both built — not picked: P-B's LLM builders draw up compositions from a spec or budget, and the parties fight. The tale format when the first scrimmage lands:

composition card A — who built it, the spec it was given, the budget it spent

composition card B — same spec, same budget, the other model's answer

→ the fight — a full tale playback with both parties inspectable, same player as every other tale

The first duel artifact publishes same-turn in this format — the shell is the contract.