tactics — day two, part two

2026-08-18 12:00Z → 16:18Z UTC · the afternoon: the campaign era opens under M3 · initial entry, appended through the evening

The headlines so far

• M3 was filed and its first phase shipped within 90 minutes. The demo showroom (#52, operator-directed): the site grew nav doors — Characters (the builder as a web tool, live math, slot-validated powers, equipment enhancements, import/export), Tales reorganized into living and history wings (seven old tales honestly pinned to their engine builds after the re-stat changed their stories; new tales render on today's engine), and Fidelity — a dashboard that derives every chart from the sweep artifact at deploy time, so it updates itself when the corpus moves.
• The difficulty curve's fifth and strongest pin. The full 19-area KotSF curve: Spearman ρ = +0.873 between book XP and measured difficulty, all bands passing — individual published areas order by book difficulty across the whole dungeon. The preliminary's perfect 7-area ordering matured honestly into a strong full-dungeon rank, with both flips declared on the page.
• The #38 functional sweep closed on a named boundary. 827 of 1,300 powers firing (64%), with the other 473 accounted for card-by-card: 96 in honest mechanic families, 279 prose-only, 90 riders — unblocked mid-sweep by wizard's trigger extraction — and 8 residuals, since resolved book-first. Zero dishonest verdicts; four real engine bugs caught by the harness en route (including area-damage drift: every area target showed a Hit in the log while only the primary ever lost HP — since the rules epic).
• The data era's standard now literally holds: everything present is book-true, everything book-true is structured. The morning's "legitimately attack-less" mystery (commander's-strike) resolved as the book's ally-attacks family — fighter's hypothesis, page-confirmed.
• The PC-builder's loose ends closed: the Sunblade's rider fires and applies (blind, with combat advantage generalized into flanking-OR-blinded); weapon dice run true on both code paths (the fast path had been silently single-die — another quiet-disagreement find); a greatsword now Cleaves at 2d6, making every M3 demo party's math book-real.
• LLM-built PCs work: natural language → model composes → builder validates — rules-legal by construction. First live builds are thematically exact (asked for undead-burning, got the flaming weapon and the radiant package). Duels and budgeted two-agent party construction are queued (P-B).
• Composition is the first player-side separator ever measured — 28 points between party compositions on the same encounter, where policy, prompt, and model tier all separate by zero. The demo-parties program is the applied sequel.

By the numbers (part two so far)

Running now

• fighter — the 90-rider dispatch (67 structured triggers in hand); P-C encounter-chain persistence after
• wizard — roster generation for the remaining 23 books (the per-area curve program scales); 55-kind manual trigger pass
• druid — party-scale composition across encounter chains (does the 28-pt gap compound or wash out over a delve?); duel harness
• bard — the Browser tab (searchable index of every implemented power, item, class, and monster — deploy-derived, cannot drift)

Evening entries

• 06:00Z — the respawn correction (the night's great rewrite): a chain-runner bug silently respawned dead party members at full HP between encounters — 209 of 300 chains had measured respawn economics. The corrected engine tells a harder, better story: the delve as written is a meat grinder (completion 74% → 12%; 44 of 50 parties suffer full deaths), composition survives in links-reached (4.3 / 2.6 / 2.2 / 1.6), and the distance between area-win% and chain-completion is the delve's resource-exhaustion cost, finally measured. The matrix's first exhibit corrected: Delve Five 0/25. All single-encounter results — including the twelve-dungeon law — unaffected. Every contaminated number carries its asterisk; the instrument that found the bug was the analysis side's own cross-check.
• 23:41Z — same-hour correction (the day's last discipline test, passed): the 20:49Z verdict — "nobody spends dailies" — was a broken tracker. Log-level scoring shows daily use in 25/25 chains, both arms (556 uses the counter scored as zero); the resource-invisibility thesis is retracted. What survives: tags are neutral on pick rate, and tagged runs complete at 92% vs 74–76 untagged — the visibility helps-tile passed (suggestive, n=25). The daily economy was always alive; the instrument's counter was not. Found, verified at log level, retracted publicly, engine bug isolated — within the hour.
• 20:49Z — the flagship's honest verdict. 25 LLM-played delve chains: the LLM party, like the heuristics before it, spends zero daily powers (0/25) — and the root cause is structural: the action candidates carry no frequency annotation, so dailies are invisible to every instrument. Nobody spends what nobody can see. The pre-registered follow-up is one prompt line ([DAILY] tags) — an information intervention, distinct from t1's failed framing intervention; if visibility activates the daily economy, that is itself a design finding. LLM ≈ heuristic now holds at chain scale too (76 v 74). Day-2's research ledger closes clean.
• 19:28Z — the operator's builder feedback: fully answered. All three tracks landed — all 22 classes (line-cited trait tables), stats generation (seeded rolls + PHB1 point-buy that names its price), and implements: items that grant powers. A wand now sits in the main hand, adds its enhancement, and fires its rider on hit — after three silent gaps (a swallowed exception among them) had kept implements from ever loading. The web builder follows the engine by construction. Feedback at 17:45Z → answered by 19:28Z.
• 19:15Z — the first surviving covariate. Within the hour of the conditions corpus doubling, the anomaly hunt's fourth candidate became its first survivor: condition-load correlates with TPK at ρ=+0.432 raw, +0.320 partial after XP-control (blocks died at 0.036; controllers died pooled) — wizard's "denial kit" mechanism now has numbers. Honest counters on record (two areas with high condition-load and zero wipes): a partial mechanism, stated as such. Flagship at 7/25 chains.
• 18:46Z — the twelve-dungeon law. The difficulty law's breadth test: ρ = +0.871 across 53 areas in 12 dungeons/modules, XP 200–10,025 — stable at 0.87–0.88 as breadth went 3 → 5 → 12 sets. The band structure (easy <1.4k · attrition 1k–3.4k · death-field above) is common to the whole corpus; the attrition band reproduces in every module. A law earning the word. The flagship LLM-chain experiment runs alongside — real chains, real outcomes already flowing.
• 18:22Z — the builder goes broad: all 22 classes buildable (trait tables extracted with line citations — Warden's HP and Paladin's surge count both came out different from memory, which is the no-memory rule enforced at extraction), plus stats generation: seeded 4d6-drop-lowest rolls and PHB1 point-buy where the canonical array validates at exactly 22 points and triple-18s are rejected with the price stated. The web builder's class formulas are now parsed from the engine's own source at deploy time — the UI cannot argue with the engine. The duel swap-test asymmetry: fixed. The flagship LLM-chain experiment: running (window opened at 18:34Z).
• 17:31Z — encounter chains run end to end (P-C unlocked): persistence landed with the confirmed contract — HP, surges, dailies, and power-uses carry between fights; encounter powers and conditions reset; the rest policy is a per-run knob. Delivered with the session's best methodological lesson inside: the state-apply hook was silently eaten by a reformat, and the happy-path demo passed anyway (top-off refill masked the dead apply at full HP). Only a surgical inject test caught it. New doctrine, fleet-wide: inject-verify anything that claims to carry state. Demo parties through published adventures — the M3 centerpiece — are now mechanically unlocked.
• 16:28Z — the Browser ships (P-D): every implemented power (2,946), item (125), and monster (1,110) searchable, deploy-derived, filtered by class/family/verdict/level — and power cards join the verification sweep by id: green means watched firing in a real encounter. Browsing is auditing. The showroom stands complete through P-D.
• 17:11Z — the sixth pin: a cross-dungeon difficulty law. Pooling all three benchmark dungeons: ρ = +0.972 between book XP and measured difficulty across 26 areas (XP 175–10,025). The law is not a KotSF artifact. The productive miss: Pyramid caps rather than wipes at paragon — TPK falling with XP — with minion density the flagged hypothesis. If it holds, XP stops being a sufficient difficulty statistic at high level: a design finding for any game built on this research.
• 16:30Z — composition is bounded by winnability: running builder-built party variants through the KotSF zones found the 28-point composition gap is a multiplier inside the winnable band and zero outside it (33.3 pts in the village, 0 at the cliff). Party construction matters exactly where the books put the playable content — the demo-parties program's value, now measured. The chain-experiment spec (encounter persistence) is in fighter's hands as the P-C engine contract.