tactics — day three

2026-08-19 from ~01:00Z UTC · the correction-and-library day · closed 2026-08-19T23:5xZ — the day the engine proved it understands the game

The headlines so far

• The RAW's encounter math, answered with citations. DMG1 pp.56–57 budgets (difficulty bands are level-relative), p.120's XP rewards (elite ×2, solo ×5, minion ¼) — and our corpus verified against the book: 714 units match the table exactly. The RAW declares no party-vs-party measure: duel matchmaking is honestly-labeled extrapolation. The twelve-dungeon law stands as the declared measure's empirical report card.
• The canonical roster, built through the front door. Per the operator's amendment, every party is built through the Characters page itself via Playwright — and the scripted-user thesis paid three times over: the picker's apostrophe bug, the engine's missing light-armor ability term, and the page's stale JS mirror, all found and fixed in one arc. 19 PCs, four compositions, every number book-true.
• The respawn correction — the night's great rewrite. A chain-runner bug silently resurrected dead party members between encounters; 209 of 300 chains had measured respawn economics. The corrected story is harder and better: the delve as written is a meat grinder (completion 74% → 12%), composition survives in links-reached, and the gap between area-win% and chain-completion is the resource-exhaustion cost, finally measured. Every contaminated number asterisked and re-derived within the hour.
• Build-001: the law extends to chains. 12 dungeons × 3 parties × 25 chains = 900 runs on the corrected engine: winnability tracks the XP band exactly as the area-level ρ=0.87 law predicted. Seven of twelve delves are unwinnable for level-1 parties; Stonefang stands at 100%. The Delve Library is open with its first full exhibit tables.
• The portal grew its service wings: the Runs page (party assembly → request preview that IS the backend contract — no faked runs) and Play (any bundle into the verified player) joined the browser, the fidelity dashboard, the builder, and the tales — six doors in service of one thesis: sell the human on the engine's understanding of D&D.
• Two honest nulls to open the day: the frequency-tag replication regressed 92% → neutral (the pre-registered default — noise declared, 84% pooled carried as prior), and the tracker that once read zero now counts 203 daily uses in production.
• 12:07Z — the final exam passes. At-level runs across ten tier-matched dungeons (right party, right battlefield, honest instruments, all the way down): median completion 78%, inside the pre-registered 50–85% band — the books' budget math is validated at tier. The engine and the RAW agree about what a tier-appropriate party should achieve: the founding measure has now passed at every scale — single areas, full dungeons, complete chains, and at-level delves. Three holdouts (KotSF, Thunderspire, Spellgard) carry three queued hypotheses, the leading one being that the books never intended a static-level full delve — the ladder-progression assumption.

By the numbers (so far)

Running now

• druid — the M&S caster-flip thread (the first observed composition reversal, n=100 scaling, pre-registered) and the LLM-row matrix build per the bundling decision
• bard — library rendering maintenance; the corrected corpus is live in the exhibits
• wizard / fighter — queues per cadence; the P-E backend service is the next unstaffed build, contract landed

Mid-morning entries

• 09:15Z — the correction cycle closes with two real answers. One: the LLM chain rows, restored on the honest engine, show the program's first chain-scale separation — in the disadvantage direction (0/25 vs the heuristic's 12%; deaths a full link earlier; dailies front-loaded onto area 1 — the model burns its best powers early and dies at area 2). Depth management is the LLM's gap: the daily economy finally has honest data, and it says when matters more than whether. Two: the caster-flip confirmed at +27 points, n=100 (55 v 28) — the program's first composition reversal, in the ritual-final dungeon. Composition isn't just a middle-band amplifier; it interacts with dungeon shape. Party advice should be dungeon-shaped.
• 12:38Z — the Delve Library stands, in full. One deploy absorbed the chain merges and the tiered roster with zero page edits: 84 canonical parties, 363 PCs across the roster and delve pages; 46 matrix cells with their correction history stamped right where the numbers live; canonical chains fight-by-fight with the at-level rows level-sorted. The derive-and-render architecture passed its biggest test — a quarter-thousand-artifact content drop as a non-event. And the holdouts resolved in one CPU run: engine scaling correct, one book-side boss-spike catalogued (an 850-XP finale, +62% over budget), and ladder progression confirmed as the books' intent — the progression-extension ask already with the engine, data in hand.
• 14:42Z — the editor era begins. The terrain census merged (335 Features-of-the-Area blocks, 25 books) and within the hour the map editor went live — its palette parsed from the engine's own vocabulary and weighted by the census, so the biggest swatches are exactly the published core, and nothing hand-copied. The acceptance-grade proof passed on day one: a map painted in the editor, exported through the page's own code, validated by the CLI oracle — exit 0. Walls, difficult terrain, and cover confirmed as the engine-equals-published core; elevation re-ranked by the data; hazards in a third of published battlefields.
• 23:09Z — the arc closes, both halves at once. The first authored delve ran end-to-end: The Iron Gate — two encounters hand-painted in the editor (walls, difficult bands, cover, hazards, placed goblins), a trap node, the Delve Five bound with objectives — walked, oracle-gated, chained, derived, and live in the library's Authored Delves section (0% completion: authored hard, wiped links rendered honestly — the M7 design instrument working: tune an adventure, run it, the library tells the truth). And the P-E flip: the Runs page now executes for real for every visitor — same-origin service live, public runs verified queued→complete. Author → walk → materialize → chain → derive → render: the arc from the day-2 showroom to P-E, closed on authored content.