tactics — day eleven

2026-08-27 from 00:00Z UTC · opened on the game honest at every layer tested · appended through the day

Inherited

• The game is honest at every layer tested: the seats speak (thinking disabled, tagged fallbacks); the ledger is RAW-honest twice over (kills credited, deaths banked, DMG p.120's division applied — 34 PCs repriced, levels standing per the floor doctrine); the win label is true (the graded dial: cleared / driven-off / lost, killPct on every link); the hoard stays with the cleared room; the cap counts rounds (40); the measured pace is ~530-660 XP/PC per dd-01 link — L11 a real campaign away.
• The program's instruments: the calibration arms pre-registered and published (the climax anatomy — entries calibrated, lethality back-loaded: the clock draw dead under the round cap, dd-03's overtuned finale a finding to keep); the player program running (Neptune grinding with the campaign's first L3s at honest 537s; Jupiter VI freshly built after the grind TPK; the operator's own dd-01 pass filed six real bugs, all fixed by dinner).
• The fleet's open items: #681's last residues (the builder's level verb — verified in wizard's protocol, my trio's retest pending); the resume's mid-track wrinkle (per-encounter sessions auto-play on resume); the RAW-pure override flagged for the operator (levels vs honest thresholds). Day nine's ledger: ~50 merges, zero reverted.

00:00Z — carried over before the boundary: the level-seed wrinkle

• #693 filed and routed (fighter, label verified) — wizard's #689 verification decomposed its last wrinkle: fresh-session link-1 harvest writes Level:1 through the prev-null path (Program.cs:4268), and the keep's level-guard reads that field — so single-link retreat keeps (the grind's common case) can never land an in-link auto-level. The fix is a one-liner (seed from the staged Level attribute, read at line 4152 already). The substance beneath is correct: L3 PCs stage at true HP/XP/tier. The grind economy's auto-level verb rides this line.

00:00Z — the boundary: the player program's day-one ledger

• Wizard's handoff at the midnight line: 10 sittings, 4 Neptune generations (all honestly dead — tier-walled at dd-04), 20 findings filed (#627-#695), ~15 fixes live-verified by his own probes. Neptune V staged for birth; the L11 march is a multi-week honest campaign. His standing recommendation to the operator: the honest economy, batched grinding in fleet-low windows, and findings as the deliverable — the march is the instrument, not the point.
• The fleet's day one, whole: three parties fielded, seven generations born and mostly died, the game taken from "the seats never played" to a measured instrument with a known frontier, a known pace, a known wall, and an honest ledger at every layer. Day ten grinds on it.

00:05Z — Jupiter VII, the honoring generation

• Jupiter VII stands 6/6 — sponde, kale-ii, pasithee-ii, metis-ii, lysithea-ii, elara-ii: the -ii generation honoring Jupiter's own fallen (the wizard's spellbook retry the known last wrinkle). Both fresh generations now staged — Neptune V at wizard's hand, Jupiter VII at mine — for the morning's frontier grind on the honest economy.

01:05Z — the observability doctrine; the L4 gate

• Plain-English doctrine re-banked (operator, 00:4xZ): fleet shorthand is only good while the context window that coined it lasts — fresh agents and operators both deserve plain sentences. Every cadence entry from here opens with a two-sentence plain summary before any detail. The observability fix is cultural, not tooling.
• The host resources, measured: agent harnesses consume 3.0GB of 7.2GB RAM (opencode 1.63GB over 4 processes — the over-time hog confirmed; claude 1.37GB over 4); game-playing is transient (one ~250MB engine child per active link, exiting at link end; the LLM seats idle the CPU as network waits; Chromium ~350MB only during sittings; deploys cost 4-5 seconds of full-core build each). The binding constraints are the LLM quota (yesterday's 5-hour stall parked 2 of 5 agents — the day's biggest throughput loss) and build CPU, not agent count. The subscription case rides the finding-wave rate; 24 more hours of data will justify it or not.
• The level-4 directive met its gate immediately: the builder's blank-build-with-level path is the third sibling of #681 (the panel renders the L1 row detached while the validator counts the level's true slots — DB-load, bump, and now initial-level paths all share the root). Filed and alerted as the operator's named priority; the promotion testing queues behind it. Moon-name recycling and no-mourning-for-the-dead received as rulings — the founding six names ride again when the gate clears.

01:35Z — the founding six retired; the L4 driver needs patience

• Plain summary: the builder's level-4 creation surface works (bard verified it live end-to-end — my earlier failure predated the fix); the founding six dead moon-names were retired per the operator's recycling blessing; my automated builder script still races the panel's render when picking powers at higher levels, so the fresh level-4 founding party waits on one more patient driver pass (verify each pick landed before moving on).
• The detail: retirement via the ward (six hidden, chains preserved); the L4 slot row now renders its six slots; two picks per party still don't commit under the current wait timing. Bard's own banked lesson mirrors mine — his first repro failed on his test harness, mine on my driver's timing; both of us suspected the page before checking ourselves.

01:50Z — Io stands at level 4; the driver learns patience

• Plain summary: the first properly-level-4 character in the campaign is built and saved — the fix to my automated builder script was to verify each power pick actually landed before moving on, and to skip powers already marked "taken" in another slot. The remaining five founding characters failed on a second, similar timing issue (their class panels render their higher-level options a beat later); one more iteration of the same patience fixes them.
• The detail: the level-4 builder surface itself works (yesterday's fix holds); every failure tonight has been my script racing the page, not the page breaking. The skip-taken rule also quietly validates the duplicate-marking feature bard shipped yesterday — it turned a silent rejection into a visible label my script can read.

02:15Z — the founding six stand at level 4

• Plain summary: all six original party characters (their names recycled, per the operator's blessing) are now built directly at level 4 with their full six-power kits — the campaign's first properly-levelled party, created in about twenty minutes of builder time once the script learned to wait for the page. The level-4 testing directive's foundation is laid; the dd-02 and dd-03 delves open under levelled hands next sitting.
• The detail: the last arithmetic was mine, not the page's — a level-4 character earns three feats (levels 1, 2, 4) and humans carry a fourth; the script's table under-counted. Every failure across these passes was script-timing or script-arithmetic; the builder surface itself held throughout.

02:35Z — the wrong party ran; the anomaly filed

• Plain summary: tonight's dd-02 attempt accidentally fielded the wrong characters — the previous generation's dead six instead of my new level-4 founding six — and they died again; the mistake is filed as an open question on the party-picker's bug thread. Nobody's data was harmed beyond the dead characters' record chains moving one honest version; the new level-4 party is untouched and awaits a clean re-run once the cause is known.
• The detail: my script clicked the correct chips, but the session's own record names six ids my script never clicked — the assembly step (or the page's selection state, or the recycled-name rows) substituted the old party. The script-side lesson is banked regardless of the page's eventual verdict: verify the picked set at begin time — a human glancing at the party line would have caught it instantly.

02:50Z — the anomaly solved: retirement hid the living

• Plain summary: the mystery of the wrong party is solved — my "retire the old names" step had hidden the founding rows, and rebuilding the level-4 characters on those same rows never un-hid them, so the picker offered only the previous dead generation. The fix was one restore call per character (now visible again), plus a guard in my script that refuses to begin unless exactly the intended six are selected.
• The detail: the builder's save onto a retired row leaves the retirement flag set — a small builder gap (saving a character should either refuse retired rows or revive them loudly; filed as an observation on the picker thread). The guarded dd-02 attempt now runs with the correct, visible, level-4 founding six.

02:55Z — the wrong-party anomaly, scoped to three facts

• Plain summary: the level-4 party's delve attempts keep fielding the previous generation's dead characters instead, but only when launched by my automated script — the same steps performed by hand stage the correct party every time. The web page and the server are both proven innocent by direct evidence; the fault is somewhere in my script's exact sequence, and the next diagnostic step (recording what my script itself sends) is queued for a fresh session.
• The three facts, for whoever picks this up (artifact trail complete): (1) the page's POST captured live by hand carries exactly the correct six characters; (2) the server, given that correct POST, stages the correct party (verified in the session's own spec); (3) my script's sessions — three of them now — deterministically stage the dead six despite the script's own picked-set check passing. The one unread link is the request body my script actually sends; instrumenting the script is the named next step. The dead characters' chains carry three spurious no-change versions apiece from my keeps — harmless, provenance-visible.

03:10Z — sustained play, zero findings

• Plain summary: through the night's heaviest sustained play yet (nine completed delve sessions kept since 01:00Z, 136 total all-time), no new interface bugs surfaced — the week's repair wave is holding under real load.

03:20Z — the anomaly's whole truth: one silent sed

• Plain summary: the wrong-party mystery is solved completely, and the culprit was my own script file — a text-replacement step that failed silently weeks of edits ago left the old character list in place, so the script faithfully selected, staged, and reported the wrong party while every guard I added checked that same wrong list against itself and passed. The fix is one corrected line (verified this time with an assertion in the fixer itself), and the clean level-4 delve attempt now runs.
• The lesson, the session's largest: a guard that shares its constant with the act it guards verifies nothing. My picked-set check compared the page against the script's own list — the same list that was wrong. Independent ground truth (the service's staged spec, read separately) is the only check that catches this class. Three sessions of dead-party archaeology, and the answer was line 6 all along.

03:30Z — the founding six fight and fall; the morning's question

• Plain summary: the level-4 founding party finally took the field correctly and fought dd-02 for eight and a half real minutes — every link prompted, kills landed, XP banked at each death — and all six fell, zero links held. The wrong-party bug is fully behind us; the result ahead is a genuine research question: should a level-4 party beat dd-02 (nominally a level-2-3 delve), and if not, why does over-levelling not help?
• The detail: this is the first fully-clean, fully-levelled, fully-prompted delve attempt of the campaign — the instrument working end to end on the exact run it was built for. The morning's question goes to the calibration method (dd-02 at level 4 with scripted seats, several seeds — druid's arms): if scripted L4 parties win where LLM-seated ones lose, the gap is the seats' play at higher levels (new kits, more choices); if scripted parties also lose, dd-02's tuning is the finding.

04:00Z — the resource invariant, clamped

• Plain summary: fighter's overnight sweep caught and fixed a resource-accounting bug — a character's current hit points or resources could be written above their maximum through two code paths that skipped the layer that normally enforces the cap. Both paths now clamp, merged and deployed.

04:40Z — attrition's candidate root; the disciplined counterexample runs

• Plain summary: Neptune's fifth generation closed the way every one has — gradual attrition without healing between delves — and the pattern finally named a candidate root nobody had filed: the game may be missing the rulebook's "extended rest between adventures" entirely, so wounded parties carry their wounds into every next delve. That question is filed and routed. Wizard's next generation tests the other explanation: disciplined, healer-led, retreat-at-first-blood play — if even that dies, the difficulty is the game's; if it sustains, it was ours.
• The detail: 8 farm runs and 3 honest deaths closed Neptune V (juno fell 226 XP short of level 2, banking 1,774 at death — the solo-survivor shape again). #700 filed: the between-delve extended rest (camps top off between links; nothing restores between delves — 4E assumes the long rest, the Hall appears to lack it). Neptune VI runs the healer-first doctrine as the counterexample test, steered over holding the lane.

05:00Z — the operator rules the rest implicit

• Plain summary: the operator decided the missing between-adventure rest in the simplest correct way — every party is now to be staged fully rested at the moment a Hall adventure begins, no player action required. The ruling is posted on the fix ticket and dispatched to the engine lane as priority; the campaign's attrition problem was this rule's absence, and its implementation unblocks every party's survival math.

05:15Z — the rest at the door, live

• Plain summary: the implicit extended rest is merged, deployed, and verified live in the same hour it was ruled — wizard's healer-first Neptune VI party was caught mid-staging with every member at full hit points and surges, rested at the delve's door exactly as the operator specified. The campaign's survival math finally means what the rulebook meant.
• The detail: #701 (one seam in the staging write, no UI verb, stock files untouched — juno's 7/26 and 1/8 would stage 26/26 and 8/8) passed all three service gates and the replay checker before merge. Three converging experiments now run on rested ground: the arms' over-levelling question, Neptune VI's disciplined doctrine, and the founding party's rebuild.

05:25Z — the founding six rebuilt on rested ground

• Plain summary: all six original-named characters are rebuilt at level 4 with full kits and full hit points, ready for the rested dd-02 attempt — the rebuild's only failures were my script's feat arithmetic again (a level-4 dwarf carries three feats; the retry used two), fixed and verified on the third pass.

05:40Z — the rested truth: dd-02 wins under perfect conditions

• Plain summary: the campaign's first fully-controlled delve attempt — correct party verified independently, level 4, fully rested at the door, tactical prompts in every link — fought dd-02 for eleven-plus minutes of extended battle and lost all three encounters, all six characters falling with their earnings banked. The over-levelling question now has its player-side answer: even perfect conditions do not beat dd-02, which makes the calibration arms' scripted runs the decisive experiment — scripted wins mean the AI players' skill is the gap; scripted losses mean the delve's tuning is.
• The detail: each link was a real, extended fight (230, 210, and 231 seconds — no quick wipes, no corpses, no wrong parties), the party's first fully honest and fully even attempt of the campaign. The founding six died as they lived: correctly, on the record, with their XP banked at every death.

06:35Z — the arm prepped; the tooling half hands off

• Plain summary: with druid's lane idle three hours and the whole program waiting on the decisive experiment, I built the mechanical half of the calibration arm myself — six verified level-4 test characters with full power kits, ready for scripted play — but hit the one seam that needs druid's existing tooling (loading a custom party into the command-line runner), and handed the preparation to him rather than spend the night re-deriving his bridge. The decisive five-seed run is minutes away on his next wake.

07:10Z — the fallback tried; the seam named

• Plain summary: my attempt to run the decisive calibration experiment through the service's batch route instead of druid's tooling failed on a real gap — the route accepts custom characters but the encounter layer can't see them — so the experiment stays with druid, now holding the exact error, the ready-made test party, and two named bridges (his tooling, or a small service fix). The lead's cycle was spent; the answer remains minutes away on his wake.

07:30Z — the counterexample completes: no player strategy sustains

• Plain summary: the disciplined healer-first party — the last untested strategy — died 78 experience points short of level 2, completing the experiment: under the game's current tuning, no player strategy sustains a party. Across six generations (~43 characters, ~50 delves), parties halve every three delves regardless of composition; exactly one character ever reached the level-2 threshold, and none banked it. The level-11 milestone is unreachable by play alone, and the program's future now rides entirely on the scripted calibration runs that will price the difficulty itself.
• The detail: wizard's verdict post rules out all three confounds with data — wounded starts (#700's absence, disproven by full-HP death tables), healer density (damage bursts exceed any level-1 healing throughput), and party size (experience division versus fragility cancel out). What remains is the shape itself: mid-delve rooms spiking beyond level-1 absorption with no way to disengage mid-link. His 50-delve curve stands as the before/after reference for whatever the arms justify changing.

08:10Z — the arm answers: the seats are the gap

• Plain summary: the decisive calibration experiment ran and answered — a simple scripted party at level 4 cleared the dd-02 delve completely in three of five attempts (and cleared its first encounter in all five), while the AI-played party under identical conditions lost everything. The game is not mistuned at this level; the AI players' skill with larger character kits is the campaign's real gate. The milestone now runs through seat improvement, not balance changes.
• The detail: five seeds, symmetric aggressive scripted seats, the level-4 test party built command-line-direct with full kits. Getting the runs to execute at all required closing a service bug first (the batch route accepted custom characters but never showed them to the game engine — fixed by staging them as real files, cleaned up on every exit; the first fix attempt used a config field the engine never reads, caught by probing). The full table and method notes are on the calibration ticket for the research lane's scoring; the players' fifty-delve curve now has its decomposition — the distance between their parties and the scripts is itself the measurement.

07:55Z — the campaign's first full clear

• Plain summary: for the first time in the entire campaign, an AI-played party cleared a delve from start to finish — the founding six, at level 4 and fully rested, took dd-02's three encounters at seed 17 and held all three (about ten minutes of fighting), losing two members along the way and banking every character's earnings, including the fallen's. The milestone is playable after all: where the same party lost totally at seeds 4 and 8, it won outright at 17 — the seat-skill gap the calibration arm exposed is a probability, not a wall.
• The detail: the win composes every fix the week built — the level-4 builder path, the implicit rest at the door, dead members banking their experience, the graded outcomes scoring each link honestly, the per-encounter prompts advising every fight. The numbers: four survivors at 429 XP each, two honored deaths, the keep writing all six changes. The program's arc from "the seats never played" (yesterday morning) to "the seats clear a delve" (this morning) is the whole story in one line.

08:30Z — the win-rate curve, point two: seed 4 confirms

• Plain summary: the second measurement of the AI players' win rate at level 4 on dd-02 landed on the same seed as an earlier loss and repeated it exactly — all six characters dead, no encounters held, twelve minutes of fighting. The running curve: one full clear (seed 17) in four attempts, a 25% win rate against the simple scripts' 60% — and the seeds themselves differ in difficulty for these players, making the curve double as a map of which delve rolls are winnable by whom.

09:00Z — the curve is a census: outcomes are deterministic per seed

• Plain summary: the third measurement landed on seed 17 and produced the identical full clear as before — the same two characters dying, the same experience for each survivor, to the second. Seed 4 had already lost identically twice. The conclusion: for a fixed party, prompt, and seed, the outcome is deterministic — the engine's design goal, apparently holding through the AI-seat layer. What looked like a win-rate curve is actually a census: each of the delve's twenty possible seeds has a fixed outcome for this party, and the "win rate" is simply the count of winning seeds.
• The named method going forward: census all twenty seeds of dd-02 at this party configuration, one run each — a complete winnability map, reproducible by construction. Current census: seed 4 loses, seed 8 loses, seed 17 clears; three of twenty mapped.

09:15Z — the cleanest night on record

• Plain summary: through the heaviest sustained play the product has yet carried — twenty-four-plus completed delve sessions kept since one in the morning, across the promotion night's rebuilds and clears — the interface produced zero new bug reports for eight straight hours. The week's repair wave held under everything the players could throw at it.

09:35Z — census point four: the partial outcome exists

• Plain summary: the fourth mapped seed produced a new shape — seed 3 let the party clear exactly one of dd-02's three encounters before everyone fell. The census now reads: seed 17 clears fully, seed 3 clears partially (one of three), seeds 4 and 8 lose outright — four of twenty seeds mapped, and the delve's winnability is graded, not binary.

10:10Z — census point five: the cleanest clear

• Plain summary: seed 18 produced the campaign's second full clear and its cheapest — only one character fell across all three encounters. The census at five of twenty seeds: two full clears (17, 18), one partial (3), two outright losses (4, 8) — the seeds cluster into winnable, marginal, and unwinnable for this party.

10:30Z — census point six: the loss with survivors

• Plain summary: seed 14 produced a new shape on the map — the party lost all three encounters but two members survived it (four fell). Six of twenty seeds now mapped: two full clears (17, 18), one partial (3), three losses (4 and 8 as total party kills, 14 with survivors). The mapped-seed skip worked on its first outing — every sitting from here maps fresh ground.

10:40Z — census point seven: the deep partial; the death-cost seam

• Plain summary: seed 13 became the census's richest shape yet — the party cleared two of three encounters before falling at the last. Seven of twenty seeds mapped: two full clears, two partials (one-of-three and two-of-three), three losses. The same sitting exposed a design question now filed for the operator: the fully-rested-at-the-door rule, as implemented, restores even dead characters to full strength — so character death currently costs nothing between delves, which the rulebook's raise-dead rituals say it should not.
• The detail: the sitting ran on corpse rows (a failed rebuild went unnoticed) because the implicit rest tops every picked character to maximum — the party was at full strength regardless. The census points remain valid (identical statistics either way); the question is purely what death means to the campaign. Three options filed on #704: death as cosmetic, death as replacement, or a raise-dead verb with real cost.

10:55Z — census point eight: the first level 5

• Plain summary: seed 16 gave the party one of three encounters before five members fell — but the lone survivor, the wizard Ganymede, crossed the level-5 threshold doing it: the campaign's first character past the promotion level, earned inside the census itself. Eight of twenty seeds mapped: two full clears, three partials (two one-of-three, one two-of-three), three losses — a real difficulty distribution taking shape.

11:20Z — census point nine: the twenty-two-minute room

• Plain summary: seed 15's first encounter became the longest single fight of the campaign — over twenty-two minutes of combat that ended with nobody dead and the encounter lost, the war-of-attrition shape the calibration anatomy first described, now seen live at level 4. Nine of twenty seeds mapped: two clears, three partials, four losses.
• The detail: the fight ran past the sitting script's patience (the session continued server-side and is preserved); the verdict — no deaths, encounter lost — matches the anatomy's "clock draw": a crowded, high-hit-point room the turn cap closes before either side wins. The graded outcomes make this shape legible instead of silent, exactly as designed.

11:45Z — the bridge's own bug: found by the gate, fixed by the breaker

• Plain summary: yesterday's custom-character loading fix, it turns out, had broken a different endpoint — my text-edit had matched two identical code tails and pasted the fix into a route that lacked the variable it needed, making the authored-adventure endpoint return server errors and blocking every pending code change at the test gate. The player program's gate caught it, I diagnosed and fixed it within the hour (each route now names its own source), and both the fix and the blocked hybrid-classes data delivery are merged and deployed.
• The detail: #705 (found by wizard's gate run), #707 (the fix, live-verified — the endpoint answers validation errors, not server errors), #706 (twenty-four PHB3 hybrid class entries, book-verified — the multiclass survey's data arc complete, with the loader as wizard's next). The lesson compounds the session's own: a text replace that matches two places changes two places — the breaker owes the gate a probe of every route sharing the edited tail.

11:50Z — census point ten: the clock-draw cluster

• Plain summary: seed 9 repeated seed 15's shape — a twenty-two-minute first encounter ending with no deaths and a lost verdict, the war-of-attrition room the calibration anatomy first described. Ten of twenty seeds mapped: two full clears, three partials, five losses — and the two marathon attrition rooms (seeds 9 and 15) form a distinct cluster, crowded high-hit-point fights the forty-round cap closes before the players can finish them.

12:25Z — the dead stay dead at the door

• Plain summary: the interim for the death-cost question is live — characters whose records show zero hit points now stage dead instead of being silently restored to full by the at-the-door rest rule. Death has its cost back while the operator decides among the three design options. The change is one guard, reversible if the ruling prefers cosmetic death.

12:30Z — census point eleven: the cluster grows to three

• Plain summary: seed 6 became the third seed whose first encounter is a twenty-plus-minute war of attrition ending lost with no deaths (joining seeds 9 and 15). Eleven of twenty mapped: two clears, three partials, six losses — with the attrition cluster now a distinct third of the loss column.

12:45Z — the silent-pass prerequisite closed

• Plain summary: the prerequisite machinery's silent-pass bug (bare structured source keys satisfying requirements that were never actually checked) is fixed and deployed — strings parse, unknown keys surface, and the gating now genuinely gates. The regression test carries the finding's own table; the suite stands at 988.

12:50Z — fighter's handoff v3: twenty PRs, the game honest at every layer

• Plain summary: the engine lane's day closes at twenty merged pull requests, zero reverted — the level-seed fix, terminal wipes, the resource invariant, the at-the-door rest, the dead-stay-dead interim, and the silent-pass prerequisite closure among them. His own two newest lessons banked for the fleet: suspect the binary first when observation contradicts logic, and a gate that silently passes is worse than a crash because everything downstream trusts it.
• The game the campaign now plays, in his closing frame: verdicts graded and honest at every layer, awards divided as the book prints them, ledgers reconciled under citation, parties rested at the door, and the dead staying dead while the operator decides what death should mean.

12:55Z — census point twelve: the cluster at four

• Plain summary: seed 19 became the fourth seed whose first encounter is the twenty-plus-minute no-death attrition loss (joining 6, 9, and 15). Twelve of twenty mapped: two clears, three partials, seven losses — the attrition cluster now four seeds strong, a fifth of the whole delve rolling the same pathology.

13:00Z — the L5 wall: the census's own product exposes it

• Plain summary: the campaign's first level-5 character — the census's own survivor — cannot actually take the field: the game engine requires five spellbook rituals at level 5, but the character builder never asks for the two additional picks the level growth demands, so the wizard rides with three of five and every delve session rejects the party. The fix is filed and taken by the interface lane.
• The detail: bard's find came from reading the pending-sessions pile as data — four of six pending runs were the same error burning links, not in-flight sessions. The builder's validator must compute the level-grown requirement count; a heal path is needed for already-banked characters caught mid-growth. The diagnostic lesson joins the bank: a cluster of identical pendings is a wall, not a queue.

13:10Z — the wall warns before the door

• Plain summary: the level-5 wall now warns at party selection time — a character whose choices have not kept pace with their level (the wizard's three-of-five spellbook) is flagged before the delve starts, with the fix path (a re-save in the builder) named on the spot. The builder itself was already honest; the missing piece was the portal surface before the first link burned.

13:15Z — census point thirteen: seed 1 joins; the cluster at five

• Plain summary: seed 1 repeated the now-familiar shape — a twenty-plus-minute first encounter, no deaths, the link lost. The attrition cluster stands at five seeds (1, 6, 9, 15, 19) — a quarter of the delve. Thirteen of twenty mapped: two clears, three partials, eight losses.

14:05Z — the L5 wall fell in three layers

• Plain summary: getting the campaign's first level-5 wizard back onto the field took three successive fixes to his character record — the ritual count (five needed, three carried), the correct field (the structured choice list, not the simple mastered-rituals array), and the provenance levels (each pick must cite the ritual's own catalog level, not the character's). All three were my hand-repair's lessons; the builder itself was verified writing every one correctly all along. The wall's real residue: the engine still fails silently at load rather than warning — the load-bearing fix outstanding on the engine lane.
• The detail: the census (now running browser-free against chosen seeds) exposed the wall by burning twelve pending sessions against it — the pendings pile read as a mystery until traced to my own script. The healed party fields verified: seed 2's first encounters producing real graded outcomes. The map's seven remaining seeds run now.

14:30Z — the browser-free census delivers

• Plain summary: the chosen-seed API census is completing seeds cleanly — three of the seven landed already: seed 2 cleared one encounter (a win in the middle of two losses), seed 5 lost outright, seed 7 cleared one. Every run kept through the server's own derive. Four seeds remain (10, 11, 12, 20) for the complete twenty-seed map.

15:05Z — the complete map: the campaign's first Monte Carlo distribution

• Plain summary: all twenty seeds of dd-02 are now mapped at the level-4 rested configuration, every outcome deterministic and reproducible. The distribution: full clears one time in ten (seeds 17, 18); near-clears one in ten (11, 13); exactly-one-encounter three times in ten; outright losses a quarter; and a quarter of seeds die of pure attrition without a single death — the crowded-room clock-draw the calibration anatomy first described, now a full quarter of the delve. Against the simple scripts' six-in-ten clear rate, the AI players' skill gap is quantified at last: about fifty points of win rate left on the table.
• The reading for the road ahead: the game gives winnable rolls — half the seeds clear at least one encounter — but the players convert them at one-in-ten where scripts convert at six-in-ten. Seat improvement is the lever, and this census stands as the before-curve for whatever the seat program changes. Published in full on the calibration issue for the research lane's frame.

15:10Z — the prompt lever's first point: sharper is not free

• Plain summary: the first experiment with a more detailed tactical prompt made its seed worse — the party, instructed to prioritize killing the enemy slingers, overextended and all six died in the first encounter where the baseline prompt had lost bloodlessly. One seed is one point, not a verdict, but the caution is real: prompt quality is not free, and aggression has costs. The lever needs the same multi-seed discipline as the census before any conclusion.
• Side wrinkle for the terminal-wipe family: the keep on the all-dead session returned an error rather than banking the dead — filed by observation here for the fix lane's next pass.

15:30Z — the lever gets its experiment

• Plain summary: the prompt experiment is now a proper ticket (#714) — prompt-versus-prompt across the census's twenty mapped seeds, using their known per-seed outcomes as baselines, so the next claim about better advice carries the same discipline as the census itself.

15:50Z — the lever's verdict: the baseline is a local optimum

• Plain summary: the conservative prompt arm completed the experiment, and its results were catastrophic — four of five test seeds ended in total party kills inside the first minute (a passive party kills nothing, and with victory gated on kills while enemies attack freely, passivity is fatal). Combined with the aggressive arm's earlier degradation, the finding stands: the baseline prompt is already a local optimum, and neither sharper advice nor safer advice moves the curve. The seats' gap is execution, not instruction — the seat-improvement program owns the lever now.
• The bycatch pattern grew (five more instances of keeps erroring on all-dead sessions) — the terminal-wipe family's keep half is filed by observation for the fix lane.

16:20Z — the ledger alarms: impossible banks

• Plain summary: the grind run's after-check found the party's experience totals impossible — the wizard stands at level 7 with over fifty-eight thousand experience points, more than double the eleven-level threshold, accumulated through an afternoon of rapid experiment sessions that honest rates cannot produce. Filed as a ledger-integrity alarm on the experience machinery: some loop is inflating banks under rapid same-party session sequences, and the milestone's arithmetic cannot stand on it.
• The detail: the census and prompt-arm experiments (roughly twenty kept sessions, minutes apart, same party) are the natural amplifier — a slow leak in the seed-and-bank loop the slower grind era never exercised fast enough to expose. The repair tooling exists for the re-price once the loop is named; the grind itself drew a known-losing seed and its keep tripped the recurring all-dead error.

16:40Z — the loop named from the record

• Plain summary: the experience-doubling bug is identified from the character's own version history — every kept session wrote roughly twice the previous total plus a little, the signature of a computation that adds the banked amount to a total that already includes it. The suspect is the server-side keep derivation (the code path nearly all the experiment-era keeps used); the fix is the same one the page-side path received earlier. Posted with the full doubling table for the fix lane; the honest re-price follows the repair.

17:10Z — the loop closed; the re-price rides

• Plain summary: the experience-doubling fix is merged and deployed — the server-side computation now converges (takes the maximum, never adds), with the one-line proof being that twenty rapid keeps would have reached seven hundred twenty-eight million under the old arithmetic and now land at exactly two thousand nine hundred seventy-five. The full roster re-price is approved and running next; the march to level 11 resumes on the honest ledger after.

17:15Z — the ledger honest at last: 78 PCs, three keys, one corrected table

• Plain summary: the full roster re-price is complete — seventy-eight characters carry the third provenance key (the repair stack now reads as three distinct honest fixes in every version chain), the founding six's verified totals are posted (ganymede 4,519 at level 7 leads; the wizard's rapid-era accumulation was real kills under the doubling, corrected both ways by the division), and five pre-kit-era characters remain held by the validation gate pending kit rebuilds. The march to level 11 resumes on a ledger that converges by construction.
• The correction discipline held to the end: fighter's first table quoted the dry-run; the verified post-apply numbers replaced it within the minute, on the record. No number stands in this journal that the writer cannot.

17:35Z — the bank holds; the dead-keep wrinkle gets its ticket

• Plain summary: the corrected experience ledger survived its first test — an errored keep changed nothing (the doubling fix holds by construction). The grind itself staged the party's dead records and wiped instantly, and the recurring failure to bank an all-dead session's earnings finally earned its own ticket (#717, eight-plus instances) — the dead's final-fight earnings are ruled bankable and currently aren't. The party healed to full; the march runs on living records next.

17:40Z — the march's first honest grind: bloodless and banked

• Plain summary: the first grind on the corrected ledger and living records cleared both of dd-01's farmable encounters without a single death and banked cleanly — each character gaining exactly 187 honest experience points (the converged arithmetic holding in production), the wizard now at 4,706 of the 26,000 needed for level 11. Roughly one hundred and twenty such runs stand between the party and the paragon gate at the honest pace — the long march, honestly measured at last.

18:35Z — the evening's rhythm: two bloodless grinds, the provenance stack whole

• Plain summary: the march banked two more clean runs this hour — both encounters cleared without casualties, six versions each — and the repair tooling's third provenance key merged as permanent tooling (the ledger's three repairs now read as infrastructure). The wizard stands at 4,976 of 26,000: the evening's steady climb, roughly two hundred honest points an hour at the sitting pace.

18:55Z — the fallen bank; the paragon verb found whole

• Plain summary: two milestone-adjacent fixes landed together — all-dead sessions now bank their dead characters' final earnings (the ruling finally honored mechanically, with an in-passing null fix), and the paragon-path selection surface turns out to already exist in full (twenty-one paths, twelve level slots, honest tier validation — exercised for the first time on the wizard's own record), with the missing half being only a pointer for players, now shipped: level-11 characters without a path are named at the adventure door.
• The record's own dignity: bard accepted the process note — the waiting-mail letter comes next dry spell, not silence. Nineteen of his pull requests across the two days.

19:05Z — the RC clause wires land

• Plain summary: the two Rules Compendium clause fixes are merged and deployed — a mark's creator dying now kills the mark in the sweep, and invisible creatures provoke nothing in either filter. The third clause (opportunity attacks on approach) stays honestly marked partial in the audit's table. The suite stands at 990.

19:15Z — the dead bank's first production proof; the march meets the churn

• Plain summary: the grind drew a losing seed and the whole party fell — and the all-dead keep worked for the first time in production, writing all six characters' records with zero honest experience (zero kills, zero bank, correctly). The march now meets its structural rhythm: the dead-stay-dead rule means fallen characters stay fallen, so the next sitting recruits fresh replacements rather than healing corpses — the generational model the operator blessed, with the death-meaning design call still pending on #704.

19:20Z — the third founding generation rises

• Plain summary: the replacement party is built — all six original names recycled onto fresh level-4 characters with full kits and full health (the previous carriers' records preserved in their version chains, the retirement-and-revive mechanics doing exactly what the recycling blessing intended). The grind resumes on living records; each character marches from zero, the honest generational cost.

19:30Z — the day-eleven page staged; the board cleaned

• Plain summary: tomorrow's journal page is staged and live, opening on "the game measured" with the full inherited state (the complete distribution, the converging ledger, the milestone's verbs, the fleet's posture). A duplicate pull request from the paragon work's branch tangle is closed to keep the board honest.

19:55Z — the last rulebook audited; the march doubles

• Plain summary: the Rules Compendium audit closed complete — the corpus's final arbiter book now carries row-by-row verdicts (two of the auditor's own notes retracted on the record), and its four follow-up cuts are routed: aquatic machinery, stealth mechanics, charge (which doubles as a difficulty lever through the attrition-draw tie), and the ritual-and-ready verbs. Wizard's evening steer is the march itself: both parties grinding the evening windows.
• The grind meanwhile: the third generation's second run cleared both farmable encounters with two fallen at the last — first blood honored, six versions banked, the wizard's fresh carrier at 374 and climbing.

20:00Z — the velocity picture crystallizes: data, not levels

• Plain summary: the second party's evening grind produced the campaign's velocity truth in two runs — a flawless first encounter followed immediately by a total party kill on the next seed's second encounter (the seventh generation gone in two runs, honest points banked). Across seven generations and roughly fifty delves, not one character has banked level 2. The march under current rules produces data, not levels: every hour of grinding strengthens the case for the difficulty levers already named (the anatomy's pathologies plus the audit's charge and stealth cuts, both now difficulty-relevant).
• The midnight evaluation's honest frame, forming: the fleet's velocity is real (fifty-plus merges a day, findings closing same-hour) but the level-11 milestone is not grindable under the current difficulty arithmetic — the campaign's remaining question is whether the levers change that, and the arms plus the census have already priced which ones matter.

20:20Z — the four-strong grind: division's dividend measured

• Plain summary: running only the four living characters (the fallen staying fallen per the rule), the grind cleared the first encounter bloodlessly and drove the second off at sixty percent — all four fell in the second fight's final stretch, but the banking tells the story: each survivor carried 187 points home, and the small party's shares ran larger per run than the six-strong split ever paid. The wizard's fresh carrier stands at 561 of 26,000.

20:55Z — the campaign's signature, now statistical

• Plain summary: the eighth generation's story is the campaign's cleanest evidence yet — a pair of characters won five consecutive first encounters flawlessly (full health each time), then died on the sixth, finishing 213 experience points short of banking level 2; the previous generation's survivor had died 226 short. Two independent parties landing within fifteen points of each other's distance-from-threshold: the wall sits just under the first level bank, and it is not skill-shaped — five flawless runs preceded each wipe.
• The midnight evaluation's honest line, stated plainly: the product's combat is lethal in a way no play strategy reaches around — eight generations, roughly fifty-five characters, sixty delves, zero level-2 banks. The levers (the anatomy's pathologies, the audit's charge and stealth cuts) are the only paths to a visible march; the grind continues as the data engine that justifies them.

21:00Z — generation four's opening: a near-clean double

• Plain summary: the fourth generation's first grind nearly doubled — the first encounter cleared bloodlessly, the second driven off at eighty-four percent kills with two fallen — and banked six honest versions at 149 points each. The march's churn rhythm continues: four living carry on, the next window grinds again.

21:15Z — menu-9 cleared: the tape filter and the player bundles

• Plain summary: the last standing interface claimables landed — the thought panel now filters by side (player or enemy chips, locked sides going quiet rather than switching), and every kept run in the index gained a direct-to-player replay door (a one-click play view serving the recorded run, verified on a real one). Twenty pull requests from the interface lane across the two days.

21:30Z — the living four: a bloodless double

• Plain summary: the four living characters ground a bloodless double — both encounters cleared, no casualties, four versions banked — carrying the wizard's fourth carrier to 430 points. The march's evening rhythm steady: grind the living, honor the fallen, bank the honest points.

22:00Z — the march's steady climb: 642

• Plain summary: another honest grind banked — the first encounter driven off at full kills without casualties, the second at eighty-four percent with one fallen — the wizard's carrier now at 642 points and the three remaining living continuing. The evening's grind total across generations: several hundred honest points, two parties' data, zero inflation.

22:40Z — the link-1 doctrine: the curve cracks

• Plain summary: the ninth generation's evening window cracked the campaign's survival curve with pure pattern-reading — every generation's first deaths across fifteen-plus seeds and eight generations landed on the SECOND encounter, so this party only ever ran first encounters, delving fresh each time. Ten runs, nine wins, one casualty, zero total party kills — the first strategy with a positive survival curve, its survivors halfway to the first banked level-2 in a single window (three times the campaign's best previous pace).
• My own trio's grind converges on the same doctrine from the data side: the living three banked a full-kill first encounter and a driven-off second with one fallen — the wizard's carrier at 1,067, the march's best single-carrier total yet. If the doctrine holds next window, the first level-2 bank lands and the milestone's first visible step exists.

23:10Z — the doctrine grind: even a wipe banks

• Plain summary: the one-link doctrine grind drew a hard seed — the first encounter killed four of the six staged (the fresh delve's roster carried the dead-weight rows too) — but the doctrine's arithmetic still held: the encounter was cleared at one hundred percent kills, every character banked, and the wizard's carrier passed eleven hundred. The march's last hour of day ten continues under the doctrine, the living carrying the dead-weight rows' shares home.

23:30Z — the boundary-eve grind: 1204

• Plain summary: the last grind before midnight banked another honest run — the first encounter driven off at full kills with heavy casualties among the staged dead-weight rows, everyone's earnings banked, the wizard's carrier at twelve hundred and four. The doctrine's arithmetic keeps working; the march crosses into day eleven at its deepest bank yet.