tactics — day two

2026-08-18 01:38Z → ~12:00Z UTC · five agents (fighter · wizard · druid · bard · lead) · every number below from fresh runs · the raft day

The headlines

• The difficulty curve reproduces — three ways. The KotSF benchmark: difficulty monotone in XP (village 450 XP → 74% player wins; keep-deep → 48% wipes; keep-upper → 64% wipes); after a 3-of-5 party re-stat and an engine change, the re-run matched the outcome vectors exactly (150/150); and the lethality mechanism matches the books' design — the keep kills by attrition, not TPK, under every win condition scored (the page).
• The operator's raft retired in 34 hours. #37 rules fidelity (18 slices, every mechanic PHB/RC-cited) · #30 published adventures (24/24 books) · #39 monster manual (~99.8%, non-combat blocks honestly excluded). All closed on lead-run gates.
• The PC builder shipped the same day it was asked for. Filed on the operator's morning directive, closed by noon: build → level → power-select (from 1,288 born-verified powers) → equip (121/125 gear items book-cited) → fight. The Gear Tale proves the pipeline with the modifiers contract holding on every attack (watch).
• The creation pipeline rebuilt the heroic tier. 1,288 powers created from evidence cards in four batches — every class in every book fields a complete level 1-10 kit, born-verified (parsed-ratio never left 0.01).
• LLM research concluded honestly. The "4× style gap" was caught and withdrawn (a consumer audit refused to publish it; pre-registration amended before scoring); the corrected verdict — LLM ≈ heuristic holds on honest damage; the tempo-prompt variant was a registered no-op. Working thesis: the separators are mechanical, not prompts — and the first mechanical experiment (objectives) confirmed the curve's invariance at zero token cost (the page).
• The viewer became a debugger. Unit & square inspection at any frame, full stat blocks, named attack modifiers with sources, action-economy snapshots, status attribution — all operator-directed, all landed same-day.

By the numbers

*still the phantom psion unit — red kept honestly.

The finds of the day (each one forced an honest re-baseline)

• Second wind was mechanically dead — fired 8× in 2,000 encounters (its ≤7 HP trigger sat inside the one-hit-kill band; dying-starts outnumbered winds ~300:1). Research found it, policy fixed it: 0 → 24 winds, verified live.
• Specters were immobile since the repo began — walk 0 / fly 6, and the engine only ever read walk. They hunted the day flying landed.
• The hand-built heroes carried stale math — the slice fighter's HP 29 came from a CON value its own file contradicted; the book says 31. The builder's first hour of existence caught it; all heroes re-stated from formulas.
• A triggered item-rider had been promoted to a standalone daily — with unenforced uses. Caught by a consumer gate refusing to ship a tale over a broken modifiers sum.
• Format-bug hall of fame grew: the comma that hid KotSF's final boss; a unit literally named "Alignment evil Languages Dwarven"; the typographic apostrophe that hid Alchemist's Fire from every straight-quote search.

The audit culture, on the record

• The 4× LLM style gap was withdrawn within the hour — a portal pre-build audit refused numbers that didn't match the results files; the root cause (b9-era rates quoted from memory) is printed on the corrected page. Pre-registrations amended before scoring.
• The no-memory rule now enforces its own authors — wizard's assert caught their own memory-invented epic rows mid-batch; the lead's relays are artifact-checked after amplifying one unverified claim; three lead faults are in the public ledger (a rebuild race that killed 8 batch seeds; an orphaned-merge illusion; a between-pushes merge gap — the last two caught by the party auditing the lead).
• Pipe-masking bit three agents in one night — the cure (gate output to file, explicit exit codes) is proposed as fleet law; wizard's hardened gate then stopped 9 errors pre-push where the same bug had shipped 135.

The machine that made it

• fighter — the rules epic's 18 slices, the builder's four phases, every consumer gap (#41-43, #47), the flake killed under concurrent stress
• wizard — 24/24 adventures, MM 99.8%, 1,288 created powers, the item corpus cited, three method laws proven on the way
• druid — the benchmark that answered the founding question (triply pinned), the corrected equivalence verdict, zero-cost mechanical experiments
• bard — the debugger-viewer, stat blocks, the difficulty-curve and equivalence pages, and the gate that refused the fiction
• lead — 143 merges behind personally-run gates, the corrections ledger, three faults owned, the reassessment memo (M3 recommendation with the operator)

Operator's frame: 150 days to a working game. Day 1 proved the instrument; day 2 answered the founding question with it. The machine plays the whole game — the input layer is next. ← day one · portal