tactics — day twelve

2026-08-28 from 00:00Z UTC · opened on the game measured · appended through the day

Inherited

• The game is measured: dd-02's complete twenty-seed distribution (10% clear / 10% near / 30% one-link / 25% loss / 25% bloodless attrition draw) — deterministic per seed, the seats converting winnable rolls at 1/10 where scripts convert at 6/10. The prompt lever tested to its answer (the baseline advice is a local optimum; execution owns the gap — #122's seat program holds the campaign's lever). The honest pace ~530-660 XP/PC/link; the L11 march is multi-generational by design.
• The ledger converges by construction: three provenance keys (repair, divide, converge) read as three distinct fixes in every chain; 78 PCs repriced; the founding six verified (the wizard's rapid-era kills were real — 4,519 honest at the corrected model). The dead bank at last (#722 in production); the fallen stay fallen pending the operator's death-meaning call (#704 — as-is / replacement / raise-dead with cost).
• The milestone's verbs exist and are exercised: the paragon-path selection whole in the builder (21 paths, 12 slots, honest validation — #723's probe), the retreat verb live, the level-4 direct creation proven, the third founding generation marching on fresh legs from zero. The map is complete; the instrument has measured everything it was asked and named the one thing left to build: better seats.
• The fleet: fighter's queue clear (the RC wires landed — 990 suite), bard at nineteen PRs across two days (the paragon pointer shipped; the waiting-mail letter accepted as process), wizard's audit closing strict-RC, druid's anatomy pre-registered and cited. The operator evaluates velocity and firepower expansion at day ten's close.

00:00Z — day eleven opens on the threshold: the first living level-2 stands blocked at the door

• Plain summary: the link-1 doctrine carried the ninth generation to the campaign's first living level-2 threshold — one character stands at 2,066 experience points past the 2,000 mark, on eight generations of honest ground — but the level-up is blocked by a builder defect (the level-2 panel never renders the required rows for characters built by the command-line tool, while the save's validation demands them). The fix is filed and dispatched as top priority; the first visible milestone step waits exactly one repair away.
• The martyr's shape also landed: another character died at 2,066 — past the threshold, the level money banked at death per the fallen's ruling. The doctrine's curve is real, and it delivers characters to the line alive and dead alike.

00:10Z — the door was already open

• Plain summary: the level-up blocker dissolved on live verification — the builder renders the threshold character's new rows correctly, the repro having been a stale page from before the last deploy. The campaign's first living level-2 is one builder session away, and the player lane has been sent to run it.

00:20Z — the first stair: Iophon is level 2

• Plain summary: the campaign's first living leveled character exists — Iophon, a wizard, banked level 2 with experience preserved through the save, the level-up's last obstacle having been a browser focus technicality (the level field commits when focus leaves it). Eight generations of honest ground, the retreat doctrine's ten wins, a martyr's parallel bank at the same threshold, and now the level. The march to eleven has its first stair; level 3 stands 184 points away on the same doctrine.

00:25Z — two levels in one window; the doctrine's final shape: the solo ladder

• Plain summary: the first character banked a second level in the same evening window — and the grind's true shape revealed itself. An over-leveled character farming the easiest encounters SOLO is safe (full health through both test runs), and the experience division pays a soloist the entire pool — roughly five hundred points per run with no observed death risk at two levels above content. The march is not party survival at all: it is getting one character past the wall, then that character climbs alone. Level 11 is twenty to twenty-five solo runs — a few evening windows.
• The honest reading for the operator: the milestone is reachable after all, not by the frontal grind but by the ladder the doctrine found. Whether the RAW's solo-farm incentive is good design is your call — the game as printed rewards it unambiguously.

00:30Z — the founding duo: a bloodless double, 1,566

• Plain summary: the founding party's two survivors ground a bloodless double — both encounters driven off at full kills with no casualties — and now stand at 1,566 experience each, 434 from the level-2 threshold that becomes the solo ladder's first rung. The march's evening windows continue under the doctrine.

00:35Z — the duo at the threshold's edge: 2,003

• Plain summary: another bloodless double carries the founding duo to two thousand and three experience points each — three past the level-2 threshold. The ladder's first rung stands under their feet; the next window levels them and the solo climb begins.

01:15Z — the pivot: the shape of success vs the shape of failure

• Plain summary: the operator redirected the program — stop optimizing the march, start analyzing why the AI players lose fights the simple scripts win. The analysis program is filed (#733) with corpus-cut rules (correct outcomes only, artifact eras fenced, backend tagged as a player variable), three controlled comparisons, and two analysts dispatched: wizard reads the decision logs (the seats' own stated rationales against the winning lines), druid reads the maps (position and targeting signatures by outcome grade). The grind lanes pause — proven ladders no longer need climbing; the quartermaster hypothesis (powers under-spent in rationales that cite the rules perfectly) is the named prime suspect.
• The operator's framing, kept whole: "we actually have a glimmer of seeing what the shape of success vs the shape of failure is in a D&D game — that is a thrilling milestone if true." It is true; the instrument earned it.

01:10Z — the homework pack, ready for the operator's return

• Plain summary: the operator is playing a game of their own and wants comparison material ready — it is posted on the analysis issue: two win/loss tuples (the founding party's first full clear versus a total wipe; and a strictly identical party run through two seeds with only the seed differing), the code-verified walkthrough of how the seed governs dice directly and seat behavior through board-state divergence (the seats themselves are deterministic at zero temperature), seed 17's identical twin run as determinism's own demonstration, and the analysts' three starting hypotheses for the human-versus-agents comparison. Homework set; the watch begins.

01:35Z — the same-seed head-to-head: difficulty is player-relative

• Plain summary: to answer the operator's summary question precisely, the scripted seats ran the census's four key seeds — closing the gap that the two players had run different seed pools. The result: the scripts clear outright what the LLM seats die in completely (seeds 4 and 8), but seed 18 inverts — the LLMs' cheapest victory of the whole census is a seed where the scripts die totally. There is no pure seed-difficulty axis; difficulty is an interaction between play style and situation. Both halves of all four pairs now sit in the analysts' corpus, same party, same seeds.

01:30Z — the operator's homework friction: the arms' side of the pairs unwatchable

• Plain summary: the operator, playing their own game and queuing up the comparison runs, found the replay player broken on the scripted runs — the direct-to-player door serves the census runs correctly but errors on the batch (scripted) runs, whose replays never register in the database the player reads. The four census runs' direct links were handed over immediately (they work now); the scripted-side fix is filed and dispatched urgently — a disk fallback plus no dead doors.

02:05Z — the pilot verdict checked against the primary record: the seats DID play

• Plain summary: the first analysis claimed the AI players never actually spoke — that every decision in the day-ten corpus was the emergency fallback. The primary record disagrees: re-reading every decision file on disk shows six thousand two hundred live AI choices and zero fallback-tagged, with the census's key runs spot-verified (the loss and the win both fully live-tagged) and a fresh probe tonight confirming the backend works. The reconciliation is likely a corpus-mixing error — the runs directory also holds seventy-seven deliberately-heuristic sessions (grind runs configured without the AI option) beside the hundred fifty-six AI-configured ones, and a sweep without that split reads as all-fallback. The analyst has been asked for their reading list to reconcile; the operator's shape-of-success question stands on a live-AI corpus after all.

02:40Z — three threads closed: the player doors whole, the clock honest, a lane repurposed

• Plain summary: the replay-player doors are fixed on both surfaces (the batch runs' bundles now serve from disk with honest hit-point chains; the runs page links open the player directly, and the operator can finally watch both sides of the comparison pairs). The outbound-mail clock drift was caught growing again — the fix (every timestamp emitted by the clock inside the writing command) is now the only path. And the engine lane's agent is repurposed by the operator toward game-engine research and the path to player input, with one last offered item: teaching the scripted seats to heal below half health, the operator's expected key to rapid delve completion.
• The shape-of-success analysis proceeds meanwhile — the first pilot's verdict (that the AI players never actually spoke) was checked against the primary record and contradicted: six thousand two hundred live choices, zero fallback, across every decision file on disk; the reconciliation request rides the analysis thread with the evidence.

02:50Z — the healing rule scoped: the operator's 50% dial

• Plain summary: the operator's policy-seat improvement — teach the scripted players to heal themselves or each other below half health — is scoped and building: the seam is the aggressive seat policy the grinds actually run, the pattern mirrors an already-proven heal-first block from the defensive policy (at the operator's 50% threshold rather than the defensive 40), and a census of the whole monster corpus confirms the rule is symmetric with zero effect on enemies (no monster carries a healing power — one hundred forty-seven player-side powers do). The engine-fix lane's ledger closes at twenty-three merged pull requests, zero reverted, the count corrected upward on the record.
• The stakes, in the operator's own expectation: substantially less bloody encounters and really rapid progress on leveling characters and completing the first ten delves with the scripted seats — then back to the AI players with the shape-of-success findings in hand.

03:05Z — the healing rule is live; the cuts reframe as research

• Plain summary: the operator's policy-seat healing rule — scripted players now try to heal themselves or each other below half health — is merged, deployed, and verified (the suite green at 995, the determinism baseline untouched, the rule symmetric with zero effect on monsters by census). The engine-fix lane closes at twenty-four pull requests with its last item shipped — the count corrected on the record, the enumeration the audit. The five unstarted rulebook cuts were re-routed by ruling: they fold into the agent's new research mandate as its first concrete work — each doubles as a case study in what the engine's mechanics make possible (stealth and its disclosure, the composite charge action, the ritual casting flag) rather than fix-lane residue.

03:15Z — the door population swept: 86 of 86 clean

• Plain summary: a sweep of every run the replay player can open (eighty-six in total) found one last edge case — a zero hit-point maximum riding an honest chain — fixed with a five-line falsy-fill correction, merged and deployed. Every replay door in the product now opens clean.

03:30Z — strict death saves landed; the research mandate framed

• Plain summary: the first rulebook cut landed — the death-saves house rules (an automatic stabilize and a no-healing-surge revive block) removed for the strict Rules Compendium ruling, the engine now failing and dying exactly as printed. And the research mandate has its framing document: the two-engines architecture (the authoritative rules server stays server-side C-sharp — the moat already built — while the "game engine" question becomes the client question, a swappable view layer over the event stream), the client options table, the four proposed spikes, and the five rulebook cuts as engine-possibility case studies. Spike ordering and the client direction await the operator's word.

03:40Z — the healing rule's first datum: a hard seed, two alive

• Plain summary — CORRECTED (03:50Z): the run originally recorded here as the healing rule's first datum was misattributed: it ran the AI seats, not the policy seats the rule touches. The healing rule's true first test is the paired run below. The correction stands as its own lesson: attribute an outcome to a change only when the changed code was in the run's seat path.

04:10Z — the charge composite's contract lands (cut two of five)

• Plain summary: the second rulebook cut of the research mandate merged and deployed — the charge action's engine contract: intent validation, a substitution gate, and book-legal rails only (no invented behavior; the strict reading). The suite green at my gate; the five cuts now stand at two landed, three to go. Each is doubling as the mandate's case study in expressing composite player actions through the engine's intent surface — the player-input thread's first primitives.

04:15Z — the mandate's cadence; a data flag for the lanes

• Plain summary: the research lane verified three merged surfaces this morning (the death-saves cut, the framing document, the charge contract) and moves to stealth — flagged as the fattest research lens of the five: hidden state and its disclosure is precisely the fog-of-war asymmetry a server-authoritative architecture owns. One data finding flagged across lanes: player character sheets carry no skill keys at all (only thirty-five monster units have perception data), so stealth's opposed checks need honest derived fallbacks until real skill data lands — a natural hand-off to the data lane if it wants it.
• The interface lane's runway ruled: playback polish first, framed by the healing era (the replays now show heal actions and survivors where corpses were — the polish items are the language that makes the new shapes readable).

04:30Z — stealth lands (cut three of five); the omniscience flag

• Plain summary: stealth mechanics merged and deployed — hiding as an opposed check against the worst enemy's passive perception, the hidden condition riding the same ladder as invisibility, revealing from stealth stamping combat advantage on the attack, and the delve seats now legally hiding when nothing else is reachable (the campaign's first legal rogue sneak). Three of the five rulebook cuts landed this morning.
• The lens's named follow-up rides the PR body for the operator's spike decision: the seats currently read the board omnisciently — hidden units stay visible to enemies — and true fog-of-war (per-seat views, server-disclosed) is the exact architecture question the commercial game's PvP and map-control threads need answered. A fifth spike candidate, stated in the framing doc's own language.

04:40Z — the player speaks the healing era

• Plain summary: the replay player now shows the healing rule's new game shapes in its own language — a bloodied badge leading the status stack and the roster chips, healing-surge pips, heal and save moments ticked on the replay's scrub bar (clickable), and an end banner that names who healed, how much came back, and the surges spent (with "all stood" on fights where nobody fell). Pre-healing replays render byte-identical. A new gate script verifies it on a real healing-era run, eight checks green.
• One population note flagged for the operator: the level-one delve parties carry no healing powers at all, so healing-era replays need parties with healing cards (level-five-plus clerics and the like) — the batch arms may want one such party by default if heals are to be visible there.

04:45Z — the surface verbs land (cut four of five)

• Plain summary: the fourth rulebook cut merged and deployed — the ritual-casting verb now opens the ritual machinery from the command line surface, and ready actions gained their trigger selection. One cut remains (aquatic machinery). The mandate's morning now stands at four landed surfaces plus the framing document.

05:00Z — the set is complete: all five cuts in one morning

• Plain summary: the fifth and final rulebook cut — aquatic combat, the Rules Compendium's page 255 implemented end to end (the map-flag battle context, fire and ungrouped-weapon penalties, the aquatic keyword on twenty-nine units, and the drowning breath track) — merged and deployed. The live smoke test caught a real wiring bug before it shipped (a damage flag clearing at the wrong moment), now fixed and verified in the same run. The morning's ledger: all five cuts landed, each book-true with verbatim citations, every gate green, the determinism baseline never moved.
• The mandate's queue now holds: the small test-race fix, the operator's spike ordering (the fog-of-war disclosure candidate queued as the natural fifth), and the framing document's three open decisions. The research lane idles between events with its watches armed.

05:10Z — the race dead; the lane's eight-PR morning closes

• Plain summary: the parallel-test race (a listed card unopenable in a neighbor's temp directory) is fixed with a deterministic pin and merged — the research lane's queue now empty of unruled work. Its morning ledger: eight pull requests (the healing rule, the framing document, the five rulebook cuts, the race fix), every gate green, the banked determinism hash never moved. The lane holds armed for the operator's spike ordering; everything else it could do without a ruling, it did.

05:20Z — fighter's day-eleven handoff: the wake that emptied every queue

• Plain summary: the research lane's handoff closes the wake — eight merged pull requests in one sitting (the healing rule, the mandate's framing document, all five rulebook cuts, and the test race), zero reverted, the determinism baseline never moved, and the engine-fix lane's lifetime at twenty-four. Every queue it could empty without an operator ruling is empty; the remaining decisions are all yours (the spike ordering, the fog-of-war fifth, the framing document's three opens).

05:30Z — the TUI reads the healing era; canon-4's hook templated fleet-wide

• Plain summary: the terminal player (the mission's second output) now renders the healing era at parity with the web player — healing moments, surge pips, and a healing-aware end banner — merged and deployed after a latent hit-point desync it fixed along the way. Separately, a fleet-process gap closed: the never-push-to-a-merged-branch hook existed only in the main checkout, so a worktree clone let a stray push slip through (caught manually, no harm); the hook is now a global git template, so every future clone carries it by construction.
• The analysis thread's first finding landed in the same window: the quartermaster signature confirmed across the honest corpus — the AI players spend their encounter and daily powers overwhelmingly early (seventy-eight early versus six late across eleven real-loss runs), leaving the closing window — exactly where kills happen and the scripts spend — power-starved. The first measured shape-of-failure.

06:00Z — the shape of failure gets its first name; the fix prototype approved

• Plain summary: the decisions-diff delivered the analysis program's first full answer — the AI players' failure is not bad judgment but mistimed spending: sound tactical reasoning within each turn, with resources spent overwhelmingly early (seventy-eight early versus six late across eleven honest losses) so the closing window — exactly where kills happen — runs power-starved. The instrument-holder approved the named fix prototype: a fight-arc context in the seat prompt (elapsed time, the unspent daily, the milestone point), to be tested against the corpus's existing spend-arc baselines. Measured failure, named mechanism, tested correction — the program's first closed loop.
• The terminal player's healing-era parity was live-verified in the same window; the canon-4 hook question closed with the gap fixed and the guard versioned.

06:15Z — the prototype's first run: advice teaches hold, not when

• Plain summary: the arc-context prompt's first test ran on the baseline's own seed — seventy-three real decisions, and the advice halved early spending (four early versus the baseline's ten) but the late window stayed empty: teaching the seats to hold is not the same as teaching them when to let go, because the turn's candidate list carries no elapsed-time signal to read. The fix sharpens to its engine-side form — the candidate framing itself should carry the fight clock, the turns remaining, and the unspent resources — and two more advised runs will firm the reading before that ask.
• The program's loop, one step from closed: measured failure, named mechanism, first correction tested, the instrument sharpened by what the test revealed. This is the shape the operator's pivot promised — the analysis teaching the seats, not merely judging them.

07:05Z — the arc-context loop closes: every third of the fight moved as predicted

• Plain summary: the timing-advice prototype's full run (three seeds matched to baselines, two hundred twenty-eight real decisions, no failures) moved every arc-third of the spend curve in the predicted direction: early halved (twenty-two to eleven), mid collapsed (nineteen to one), late tripled (one to three) — including the corpus's first daily power fired in a closing arc. The runs still lose (a fresh roster against tuned content), but the mechanism is proven: the seats rationally respond to timing doctrine, and the remaining ceiling is the clock itself living outside the turn — the engine-side candidate-framing fix, now with three-seed evidence behind it.
• The loop the operator framed is closed end to end: measured failure, named mechanism, tested correction, sharpened instrument. The seat program's next move is engineering, not analysis — put the clock where the seats can read it.

07:10Z — the bull-rush remainder closes #727 whole

• Plain summary: the charge cut's last remainder — the bull-rush variant's plus-one in the opposed check — merged and deployed, closing the fifth rulebook cut completely. The research lane's standing sweep found it labeled, unblocked, and unruled, and took it correctly rather than holding for a fourth check.

07:20Z — the board's last loose end closed

• Plain summary: the prompt-experiment issue — filed when the lever needed census discipline, answered first with the local-optimum verdict, then superseded by the arc-context prototype's three-seed proof — is closed with its successor named. The lead's own board now holds the program umbrella and the long-horizon editor epic; everything actionable lives on its owner's lane.

07:45Z — the clock goes in: the arc-context engine side lands

• Plain summary: the fight clock is now engine data in the seat prompt — round, cap, percent elapsed, rounds remaining, the unspent one-shots listed, and the honest warning that resources expire worthless at the cap. Merged and deployed with the old prompt pinned byte-stable and a one-key switch for the analysis runs. The matched-seed test fires next: the same three seeds the advice prototype ran, reading whether the clock fills the closing window that advice alone left empty.

08:00Z — the clock runs fire; the dedupe finding

• Plain summary: the matched-seed clock tests are running (the three seeds, exact pairing, one to two and a half hours). En route, the extraction surfaced a real measurement bug: the event stream double-records power usage (the resolver and the apply path both log it), inflating raw spend counts by about forty percent — uniform, so ratios held, but every absolute number in the earlier curves carries the doubling. The extractor now dedupes; the clock's verdict will read both sides on the corrected metric.
• The reconciled advice-side numbers are posted for the comparison: the late window is exactly as thin as advertised (dailies zero on all three seeds, encounters thin) — the empty window the clock is asked to fill.

08:05Z — the false start caught by its own hash; attempt 3 live

• Plain summary: the first two clock-run attempts fired with the wrong setting key — the engine has carried its own native prompt key since the chain runner's early days, and the new plumb under a different name never reached the seats. The evidence gate caught it (the decisions carried the old template's hash, so the runs were discarded before they could masquerade as data), and the third attempt runs now on the native key with a hash-verification probe waiting — the evidence only counts if the new template's hash reads this time.

08:10Z — the stale binary: the third catch, and the standing rule

• Plain summary: the false start's full root cause was worse than a wrong setting — the engine binary was nineteen hours stale, so even the correctly-keyd attempt ran pre-clock code. The template-hash gate caught all three failure modes before any false evidence was extracted; the fourth attempt runs verified end to end (rebuilt binary, probe-confirmed template, and the completion watch itself now reporting whether the evidence is clean). A standing rule lands with it: instrumented runs rebuild first, no exceptions — the convenience of running an existing binary is exactly how yesterday's code answers for today's.

08:20Z — the clock's clean negative: the framing is commentary, the scoring is law

• Plain summary: the fight-clock experiment returned a clean negative — the seats given the full timing picture spent identically to the seats without it (the late window empty in both arms, the curves indistinguishable), meaning the late window is empty not because the players lack a clock but because nothing remains to spend by then: the biggest-dice powers win candidate selection the moment they are offered, whatever the surrounding text says. The mechanism sharpens to its final form: the prompt is commentary, the candidate scoring is the law. The next lever is a spend-bias knob in the scoring itself — exactly the analysis lane's territory.
• The morning's arc closes in the honest shape: a measured failure, a named mechanism, a first correction (advice) that moved the curve, a second correction (the clock) that didn't, and the survivor hypothesis — scoring, not information — now carrying three experiments' evidence. Every catch along the way (the dedupe, the hash gate, the stale binary) made the negative trustworthy.

08:25Z — the falsification answers: a weighting knob, seventeen for seventeen

• Plain summary: the decisive test ran on the corpus — seventeen late-fight decisions where a daily power was genuinely available and was the highest-damage candidate on the board; the seats chose it three times in seventeen. The biggest-dice power sat there, the scoring demanded it, and the seats took the small spell anyway. The knob is weighting, not availability: the candidate ranking needs a late-window spend bias, a mechanical adjustment and not another sentence in the prompt. The seat program's next lever is named to its exact form.

09:00Z — the weighting knob lands as the interface

• Plain summary: the seat program's next lever shipped — in the late third of a fight, the candidate list itself now orders dailies first and annotates each with its expiry ("useless after this fight"), working on the interface the seats actually read rather than in the prompt's prose. The suite green; the acceptance evidence (the seventeen decisions where the best candidate went unchosen) stands ready for the before/after read.

09:20Z — the second negative: the spend curve is invariant, and the lever moves to generation

• Plain summary: the interface knob also moved nothing — and across all three experimental arms the early spend holds at eighty-two to eighty-four, astonishingly stable against prose, clock, ordering, and annotation alike. The deeper structural fact: on these runs the daily powers are all spent by round five of fights that run ten to sixteen rounds — the late window is empty because the early window empties it. The lever, finally located: not what the prompt says, not how the list is ordered, but what candidates are OFFERED early. The next testable knob is availability gating — dailies entering the candidate list only from mid-fight — which would turn the invariant from an observation into a control.
• The morning's method held to the end: two clean negatives, each eliminating a hypothesis family, each caught on verified rails. Four experiments in five hours, one surviving mechanism: the seats spend what they are offered when they are offered it.

09:25Z — the two lanes converge on the same lever from opposite ends

• Plain summary: the analyst's retrospective finding — the model picks from the first third of the candidate list regardless of content (chosen positions averaging a third of the way down; dailies taken at slots six and seven, refused at eleven and beyond) — converges with the build lane's synthesis from the opposite direction: what the list offers early determines what gets spent early. The v5 build already shipped late-third dailies-first ordering and it moved nothing on its own; the two findings together sharpen the question the gating experiment now runs: whether the ordering needs to be early-third availability gating (not reordering late) to convert refusals into selections. The two arms of the program — analysis and build — are reading the same corpus and arriving at the same seam within the hour.

09:35Z — the availability gate lands; the invariance's falsification runs

• Plain summary: the fifth experimental knob shipped — daily powers enter the candidate list only from mid-fight — turning the morning's stubborn invariant (spend-the-moment-offered) from an observation into a testable control. Merged and deployed; the acceptance runs fire on the same three seeds against all four prior arms' baselines.

09:55Z — the gate moves the curve: generation proven, and the question re-scales

• Plain summary: the availability gate moved what three other knobs could not — withholding dailies from the early offer cut early spend from eighty-four to sixty-nine and total spend from ninety-eight to seventy-three. The control point of seat spending is candidate generation, proven: prose no, data no, interface no, offer yes. And the answer re-scales the quartermaster question itself: the late window never fills because the fights resolve before a save-it moment structurally exists — the resource-timing question is a delve-scale matter (dailies across a three-link track), not a fight-scale one. A small-sample bonus with the caveat stated: zero losses under the gate arm, where the prior arms lost three times — withholding early dailies directionally improved survival, the first arm to do so.
• The morning's arc closes whole: four arms, four sessions of clean evidence, one morning, zero contaminated reads — the pivot's promise delivered end to end. The seat program now knows exactly where the lever lives and at what scale the question should be asked.

11:05Z — CORRECTION: the Ox model died; the knob chain's v4/v5/v6 readings are fallback data

• Plain summary: the urgent catch — the free-window model was retired on the endpoint (the service's own probe now returns model-not-found), meaning the morning's clock, interface, and gate experiments all silently ran the fallback heuristic instead of the AI seats. Their readings — including the gate's celebrated curve movement — describe the FALLBACK's response to availability, not the seats'. The genuine seat findings stand where the model verifiably lived: the advice arm (verified real at 06:0xZ) and the falsification's seventeen-decision corpus. The service is reverted to the proven model; the experiment chain re-runs on clean rails once the canary ships. The corrections stand visibly in this journal — the strongest finding (the seventeen refusals) predates the contamination and survives it.

11:10Z — the canary ships; the double fix verified by its own before/after

• Plain summary: the seat preflight canary — recommended three times, cost three experiment batches in its absence — is built, committed, and standing: one character, one encounter, forty seconds, refuse-to-run when every seat call errors. Its first run caught the outage live (four-of-four failed before the fix), and its second proved the fix (four-of-four live) — the canary's own before-and-after on its first morning. The fix itself was double-layered: the environment variable alone didn't take because the service's code constant fed the spec builder directly; both layers now point at the proven model, deployed and verified.

12:25Z — the true read: the clock's negative retracts; the four-arm chain stands clean

• Plain summary: the rerun on clean rails (nine tracks, six hundred eighty-four verified decisions, one transient fallback) rewrites the morning: the clock experiment's negative was an artifact of the dead backend — with the model actually choosing, the clock moved the curve genuinely (early down sixteen, mid up three, a real timing shift). The gate's offer-controls-spend finding is confirmed on real choices, and the interface arm's reorder did not fill the late window (the positional-bias prediction unconfirmed at this sample size, and the delve-scale explanation stands). One new puzzle flagged for the analysts: the interface arm's mid-fight spend collapsed opposite its intent.
• The corrected four-arm chain: advice eighty-four early → clock sixty-eight → interface seventy-seven → gate sixty-four — each knob's effect now measured on choices the model actually made. The instrument's morning lesson holds in both directions: contaminated data can hide a real effect as easily as invent one.

12:45Z — the mid-collapse names itself: scarcity-urgency beats timing-discipline

• Plain summary: the interface arm's opposite-intent collapse is explained by the analysis: the expiry annotation reads as scarcity, and scarcity-urgency beats timing-discipline — the same weighting atom that overrides information and instruction also overrides the working clock, collapsing the clock's genuine mid-game spend while spending more early. Neither knob is broken; they interact destructively. The missing experiment cell is dispatched: the reorder alone (no clock, no annotation), the positional-bias prediction's clean test.

13:10Z — the reorder-alone cell lands; the matrix completes

• Plain summary: the isolated sort (dailies-first in the late third, clock and annotation stripped away) merged and deployed — the fifth and final cell of the interaction matrix. Its acceptance runs answer whether position alone moves late spend, completing the map of what each knob does alone and what the combinations destroy.

13:40Z — the matrix completes: two levers, proven; the rest, named inert

• Plain summary: the fifth and final cell refuted the positional prediction — the sort alone moved nothing (the highest early spend of any arm, the late window empty). The complete five-arm matrix now states the seat program's findings with every cell isolated: exactly two levers move AI-player spending — timing data in the prompt (the clock's modest early-to-mid shift) and the candidate offer (the gate's structural control). Prose is inert; ordering is inert; the annotation is actively harmful (its scarcity-urgency inverts the clock's discipline). The untested composition — gate plus clock — is the natural next cell if the analysts want it.
• The day's experimental arc, complete: five arms, three contamination catches, one retracted negative, one refuted prediction, and a final map in which every lever's effect is measured on choices the model verifiably made. This is the shape-of-success program's deliverable — not a fix, but a complete mechanism map for whoever builds the next seat.

13:50Z — the fusion: a saver in a game with no tomorrow worth saving for

• Plain summary: the direct test of the structural explanation ran — thirty of one hundred eighty-three late-fight decisions (sixteen percent) carry explicit saving language, and the verbatim rationales are articulate hoarding: saving a damage zone for chokepoints that never arrive. The seats know it is late and choose to hold, with reasons. The fusion of the two findings names the whole mechanism: the model rations optimally against a future the game never provides — the fight ends with the daily banked and no chokepoint ever coming. The fix's final form follows: not more information, but making the future real (a factual deadline) or reordering what is offered — the two proven levers, now with the why underneath them.

15:00Z — the sixth cell: full power, the operator's recommendation

• Plain summary: the operator directed the next experiment — the seats at temperature one with thinking enabled (the current zero-temperature no-thinking configuration described as cruising under emergency power). Filed and dispatched to the seat-machinery lane: two spec knobs where hardcoded constants now stand, the bloody seed pairing, a distribution comparison (the determinism tool set aside where the question demands), and the morning's canary and purity rails. The question: is the seats' rigidity the model's or the settings' — and either answer moves the program.

15:10Z — the footprint review: three lanes named; the full-power knobs land

• Plain summary: the operator reviewed the fleet's true utilization — today's session volume runs at a quarter of the peak, and the spend changed from exploratory to deliberate. The consolidation: three active lanes (the lead/instrument, the engine/seats, the analysis); druid asked for one closing wake (his geometry half or an explicit hand-off, plus the unanswered reconciliation); bard stays terminal at zero cost. The freed quota's natural absorber is the operator's own engine-spike thread. Meanwhile the full-power seat knobs (temperature and thinking as spec parameters, defaults unchanged) merged and deployed — the sixth cell's acceptance runs fire next on the standing rails.

16:05Z — the archaeology closes: saving-language is encounter-shaped

• Plain summary: two zero-cost falsifications closed the hoarding puzzle's alternative explanations — the template and the advice were ruled out by the records themselves, leaving the encounter family as the true variable: the rooms with chokepoints and packs (the goblin-underboss texture) are where the seats articulate saving reasoning, and the open rooms are where they do not. The saver anatomy stands; its expression is room-dependent — the seats save-reason when the room offers something worth saving for. The practical edge for the experiments: the underboss rooms are the sensitive instrument for any deadline or offer fix's effect, and the arms may want a room-family covariate tag.

16:15Z — the canary catches a fourth shape; the room covariate wired

• Plain summary: the acceptance read now carries the per-room cut (arm by room-family — the underboss room is its own column, where a room-shaped effect shows first), and the canary caught its fourth batch-killing shape before it burned anything: at temperature one the model thinks longer, and the first sizing's caps were eaten by thinking exactly as the old bench predicted. The second sizing runs; the distribution fires at whatever clears. The canary — recommended three times, built this morning, now two catches old — stays permanent.

17:15Z — the clustering cut: the saver speaks in the swarm rooms; the hoarding is universal

• Plain summary: the day's final analysis cut found where the seats' saving language lives — exclusively in the crowded low-tier swarm rooms (eight to nine percent there, zero across one hundred forty-three decisions everywhere else). The refinement splits the finding in two: crowded rooms elicit the saver's ARTICULATION (swarms create the illusion of a longer game worth reasoning about), while the hoarding itself is universal — the late window stays empty in every arm regardless. The instrument map follows: the swarm rooms are the sensitive detector for any fix's effect; the mid-tier rooms test whether a fix moves the silent hoarders too. The analysis day closes complete — every link of the chain posted and banked.

19:35Z — the sixth cell's verdict: the wallet opens at halftime

• Plain summary: the full-power arm's verdict landed (two hundred forty-six of two hundred forty-seven decisions clean): the spend curve reads forty-six early, nine mid, two late — the mid restored from the clean arms' collapsed two back to nine, with the late window cracking from zero to two. The signature read: the wallet opens at HALFTIME, not the cap-end — the timing mechanism's own prediction, since the fights are largely decided by the final third and the real decision window is the middle. The full-power settings moved the middle where the kill windows live — arguably exactly the right middle.
• The operator's emergency-power framing vindicated in the data: the settings were part of the rigidity. The full mechanism map now carries six arms, every one measured on choices the model verifiably made, the last two at the settings the model was built for.

20:00Z — the final read: training-deep hoarding, and the outcome axis moves

• Plain summary — the journal corrects its 19:35Z entry: the full arm's complete read (three full runs, three hundred four decisions, all clean, the canary ladder's final sizing after three teach-ups) finds the hoarding persists at full power — the spend curve sits in the emergency-power band, the sensitive rooms read flat, the late window stays empty everywhere. The settings are NOT the rigidity: the hoarding is training-deep. The earlier entry's mid-restoration reading came from the arm's other half (the mid-flight batch wizard flagged as its own); the distribution itself is conservative. The one suggestive divergence is on a different axis entirely: outcomes — three clears in nine full-power links against the emergency seats' none, same conservatism but better choices. Small sample, flagged not claimed; the analysts' next probe, if any, reads the outcome axis.
• The matrix's final shape, six cells, all purity-tagged: prose inert; the clock a modest real shift; the interface confounded by scarcity inversion; the gate the structural control; ordering refuted; full power training-deep — with a possible outcome gain that emergency power was suppressing.

20:35Z — fighter's day-eleven handoff: sixteen PRs, the instrument sprint whole

• Plain summary: the research lane's closing handoff — sixteen merged pull requests on the day (the fix lane's final item, the mandate's framing document, all five rulebook cuts, the test race, the bull-rush remainder, and the afternoon's instrument sprint of six experiment knobs), zero reverted, determinism banked throughout. The six-cell matrix stands complete with its final entry. The lane rests; the flagged outcome divergence is the analysts' call.

21:15Z — wizard's analysis-day handoff: the program's first day fully delivered

• Plain summary: the analysis program's second and final handoff — the day's every deliverable banked across the issue, the mails, and the reference: the corpus purity-tagged with three backend eras caught, the failure anatomy complete from judgment through the saver fusion, two zero-cost falsifications, the instrument map, and the full-power intervention measured. One mid-flight correction retracted openly. The deep rationale read is staged as the fresh context's opener; the lane clocks out.