tactics — day fourteen

2026-08-30 from 00:00Z UTC · opened on the successor program's first verdict · appended through the day

Inherited

• The analysis program (#733) closed with the two-lever model — the offer for spend, carried relational context for use — and the verdict published (shape-of-success.html). The board-state seat (#772, the operator's endorsed direction) passed its spike at full power and its measured batch read clean: the board is a TIMING lever (early 124 vs baseline's 140 — the seats that see the fight hold resources into the middle without any spend-side intervention), purity 1,339/1,339, outcomes level with baseline (the ring remains the outcome lever). The stacked arm (v12: board+ring) is GO by the plan's own amendment.
• The Unity determinism spike (#773) closed complete on its first night: all four runtimes identical including the IL2CPP mobile path — the engine decision informed. The artificer (the lamp host's first lane) delivered provisioning-to-verdict in under three hours; U2/U3 follow-ups landed via #776.
• The local-LLM lane (the operator's second lamp direction: Ollama models for seat runs) is pulled and delivered; ownership clarification pending (the historical self-play lane shows parked), and the experiment owes its power-shape naming.
• The engine lane's backend default now rides the subscription (glm-5.3 primary, the zen gateway behind an env var) — the Ox retirement is a clean commit, not a fourth silent era.

00:05Z — the successor program's first verdict: the board is a timing lever

• Plain summary: the first measured comparison of the board-state seat (five full playthroughs at full power, 1,339 of 1,339 decisions verified clean — the program's largest pure sample) lands a clean, separated result. The seats that see the whole fight spend differently than any arm before them: the lowest early spend of the three arms (124 against the baseline's 140) and the highest mid-fight spend (19 against 12) — a timing shift produced by the state alone, with no instruction, clock, or annotation touching the prompt. Every prompt-side attempt at this same shift required the prompt lever; the board gets it from sight. Outcomes held level with baseline (six clears, eight deaths, one loss) — the relational ring remains the outcome lever. The two levers now stand separated, one per cell: the board moves when resources are spent; the ring moves whether fights are won. The stacked arm (board plus ring) fires by the plan's own rule — a positive board read — and its measurement rides the same tables.

03:10Z — the composition verdict: the levers dilute each other, and the stack never loses

• Plain summary: the stacked arm's read is honest and interesting: the two levers neither compose additively nor cancel — each dilutes the other. The board's timing effect survives at half strength with the ring riding (mid-fight spend 14, between the board's 19 and the baseline's 12), and the ring's outcome edge dilutes under the board's presence (five clears, the lowest count). But the stack carries one distinctive no other arm shows: zero lost links — across all five playthroughs, the stacked seat is the only configuration that never once lost an encounter, the safest seat the program has built. The full four-arm table (baseline, ring, board, stack — each five tracks, purity near-perfect throughout) is reproducible from disk.
• The standing caveat holds (five tracks per arm), and the interference pattern — the old annotation-clock shadow falling softly — earns the analysts' by-power pass before any v13 is designed. The compass question the operator holds in the morning is now a three-way trade: the ring clears most, the stack never loses, the board times best.

03:35Z — the operator's rulings: the stack is the product direction, the monsters were seated all along, and the local-model re-test fires

• Plain summary: three rulings and one finding from the operator's late-night read. The stack is what the product wanted all along ("the version that actually performs"). The monster-side question, checked on disk before answering: both sides have been AI-seated in these arms all along (the run policy seats player and enemy alike), and the upgrades ran symmetric — identical template hash on both factions, the monsters seeing the same board and carrying the same intentions. Every arm's outcomes were earned against equally-upgraded opposition; the stack never lost while the monsters also stacked. The honest repricing that came with the temperature question: the board and stack arms ran at temperature 1 with thinking on, while the baseline and ring rows of the table were temperature-0 — and the earlier full-power-alone arm shows most of the timing restoration is the power, with sight adding a real but smaller increment on top. And the local-model re-test — the operator's explicit directive — is dispatched to the lamp lane: the stack configuration on 4070-class local models, same seeds, read against the four-arm table, asking whether the intelligence upgrades survive a much smaller brain. Subscription arms rest while it runs.
• The asymmetric arms (stack party against baseline monsters, and the reverse) are the natural next cells the monster question opens — staged, firing on the operator's word or after the local-model read.

04:50Z — the player program re-engages: the campaign to level 11, with seats that can play

• Plain summary: the operator's ruling re-opens the player program on the analysis arc's shoulders. The reasoning, in their words: preliminary evidence that the game as written is very beatable ("well-tuned" deferred as deeper but not blocking), so direct testing proceeds — one of the idling agents takes back their party and levels it from level 4 through level 11 against at-level delves, full delve runs only, roughly two to three completed delves per level. The campaign validates the whole level 4-to-11 pathway (the experience ledger, the level-up machinery, the delve ladder above the current corpus) — every break it finds is the point — and when level 11 lands, a paragon path: the program's original goal, now reachable with seats that never lost an encounter. The whole saga then repeats for the paragon and epic tiers.
• The lane is the parked self-play research lane re-armed for it (its two open issues — parties die faster than they level, and should a level-4 party beat this delve — become the campaign's own frame); the party is the Jupiter six at level 4; the seats are the stack, both sides, per the product-direction ruling. The resumption is posted to the player program's issue; the campaign's ledger will post there as it runs. Velocity, reclaimed.

05:40Z — the agentic seat epic opens: the seat that asks questions

• Plain summary: the operator ruled a lane onto the fully conversational AI seat — the form where the model receives tool calls instead of a pushed prompt, letting it fetch data cards on demand (a unit's full statistics, a power's text, a terrain query) exactly as an agent reads files or runs commands. Their reasoning carries the decision: beyond the skill-upgrade case, it is future-proofing against the menu itself — at paragon tier the character sheet roughly doubles in complexity, and the pushed-prompt form cannot carry it. The epic (#781) goes to the engine lane that built the board and the stack (the direct succession of its own work), with deliverable zero being the sketch — the tool-call surface, the conversation shape, the replay contract, the local-model question — reviewed before any build. The compositional anchor on the record: the tool surface is the engine as an API, the same data cards a human client renders — the Unity spike proved the human side, and this seat is the model side of one interface.
• The fleet's morning now runs four lanes at once: the campaign on the simulation host, the local-model re-test batch on the home machine, the agentic-seat sketch on the engine lane, and the record.

06:05Z — the operator's action-economy question, answered on disk: the seats are playing a one-action game

• Plain summary: the operator asked whether the AI seats use their full turn — standard, move, and minor actions — or whether the menu collapses the turn. The event streams answer definitively: in the measured link just read, 79 action-use events across 199 unit-turns, and every single one is a Standard action. Zero move actions. Zero minor actions. Zero movement events of any kind. The engine implements the full economy — the action types exist, the turn manager tracks what is unspent — but the seat menu only ever offers standard-action attack choices, so the seats play a drastically reduced version of the turn: stand still, attack once, pass.
• Three consequences land at once. The whole experiment record measured seats playing this reduced game — the sound-judgment findings and every arm's numbers describe play within a menu that never offered movement or minors. The campaign now running plays the reduced game too — its ledger should say so until the menu or the seat grows the full turn. And the agentic-seat epic's open question (a submit-action tool versus the index-reply protocol) has effectively answered itself: the menu does not merely lack detail — it lacks whole action types, and the operator's instinct that this argues for submitting actions directly is borne out by the data. The finding routes to the engine lane's sketch and the campaign's ledger.
• Correction (2026-08-31, PR #800): of the 79 Standard-type events, 3 were action-point reflex doubles — the same card resolved twice in one turn (the europa forensic's bug: a missed daily getting a free retry). Corrected: 76 ordinary standards, 3 AP extras, and the 3 were illegal same-card extras by the era's reflex. Zero moves and zero minors STAND — the finding's core (the menu lacked whole action types) is untouched; the count was 4% high. Future censuses read the ActionPointExtra marker (merged with the fix) and never conflate the two again.

06:15Z — the agentic seat's build word: "the game is bigly incomplete without it"

• Plain summary: the operator gives the build word on the agentic seat, with a framing the whole record should carry: the finding that days of data now carry a short shelf life is not a loss but the signature of progress — on day thirteen of a project, working software that supersedes its own measurements weekly is what early-stage building is supposed to look like, and no sleep is lost over superseded corpora that bought the ground. The build proceeds on sketch v2's shape: the read-only tool loop first (proving the seat can hold a call-read-answer conversation), then the full-turn submit — standard, move, and minor composed by the model on its initiative — as the first increment, the piece without which, in the operator's words, the game is bigly incomplete.

06:40Z — the wire proof: the model pulled a card and quoted it

• Plain summary: the agentic seat's milestone zero closes with its own definition met, live: one decision, one real tool call, one choice, replayed. On the first turn the model called the unit-card tool; on the second, fed the card, it answered in the standing protocol — and its rationale quoted the card's own content ("full HP, AC 16, no conditions"). That is the thesis's first live evidence: pulling changes the reasoning, not just the prompt budget. The seat proceeds to the runner wiring, the spike session, and then the full-turn submit.
• The same hour's housekeeping catch, owned: the day's journal commits (twenty-three) had been rejecting silently behind a verification that counted rejection lines as success — every pull-request merge between journal pushes moved the record ahead of the local tree, and the portal's file-level deploys kept everything looking delivered. Pushed clean, and the check is now structural (a behind-count of zero after fetch, with the push un-piped). The engine lane caught it twice in one day; the record is whole.

07:50Z — the morning's three lessons: a killed canary owned, a parser that mistook syntax for incompetence, and gates that may ride luck

• Plain summary, three findings in one hour. First: the agentic seat's first canary died at twenty-nine minutes — and the kill was the lead's own deploy, whose between-links wait expired its bound and force-restarted the service into the live run; the warning said so, but a truncated log read hid it. Owned, with the standing rule named: no deploys during live-run windows, and any forced window is an explicit decision, never a side effect. The spike's retry shape is ruled to match its question — two tool turns with thinking clamped, not maximum cost. Second: the local-model batch's fallback mystery solved, and the answer is the morning's best line — the small model answered every turn, 339 of 339, with correct indices and reasoning in a slightly different bracket format, and the strict parser rejected every one as unparseable. The model was never mute — the seat was never listening to its syntax. The parser now tolerates the variant with bounds discipline, and the canary gains a zero-fallback assertion so a parse regression can never hide behind live attribution again. Third, a bonus catch: the engine's own test suite intermittently fails on clean master — a temp-file race — meaning recent all-green gates may have been partly luck; routed to the engine lane as a standing defect in the fleet's instruments.
• The deploy carrying both fixes queues behind the live-run window; the hardened canary and the corrected batch re-fire when it clears.

09:30Z — the temperature control lands: the board's effect is the state, not the power

• Plain summary: the same-temperature baseline (plain seats, full power, five seeds) completes the control the four-arm table was missing — and it sharpens everything the mixed-power table blurred. At identical temperature, the board's mid-fight spend effect (19 versus the new baseline's 10) holds completely: the timing restoration is the state itself, not the intelligence settings riding with it — the board arm's reading survives its control. Isolated separately, full power alone cuts early spend about nine percent (the old baseline's 140 to 128), a smaller and different effect. The lead's earlier re-pricing — "most of the timing restoration is the power" — is corrected on the record: the power moves the early window; the mid-game restoration belongs to the board. The stack's dilution shape also holds at equal power, and the quartet's one missing row (the ring at temperature one) fires with the corrected local batch — the same instrument, the same read.
• The engine lane's fixes from the morning (the spike's cost clamp, the thinking-disable's coverage hole) are rebased and ready; the deploy carrying all of them fires in the cleared window.

09:45Z — the fourth data-quality era closes, a regression dies in twenty minutes, and the tables get their retroactive gate

• Plain summary, three threads resolving within the hour. First: the morning's fallback anomaly is officially closed — the corrected runs stamp at 696-of-697 purity, the 46% having collapsed to 0.1% when its causes (the parser's syntax deafness, the thinking-disable's hole, concurrency timeouts) were each fixed; the anomaly-flag-probe-fix-rerun loop ran whole, the fourth such era caught and closed like the first three. Second: a deploy regression — every new session rejected by the service after a fix landed with its plumbing half-done — was caught at the canary gate before any spend, diagnosed from the service log in one command, hot-fixed by the lead in two lines, and proven fixed by a live probe, all inside twenty minutes; the gap it exposed (service-layer changes ride no test gate) is named for the engine lane. Third: the counting tools gained their liveness gate retroactively — one table row had been reading a deploy-killed partial session as data, caught by cross-check; the correction lands on one row (the temperature-one baseline: 113/10/3, not 128), every other published number re-verified unchanged, and the standing lesson joins the reference: a counting tool needs a completion gate, because glob-all reads corpses.

11:50Z — the first agentic delve: the seat that asks questions, whole

• Plain summary: the program's first agentic seat fought a complete three-link delve pulling its own state — 218 choices made, 6 declined (a 97% hold), zero fallbacks, the tool transcript riding every decision record for replay. In its opening turns the model explored — fetching its own card and the enemy's, listing its powers — and then chose from the menu. The outcomes (three driven-offs at 52%, a near-clear at full kill percentage, and 72%) say the seat fights competitively at the deliberately clamped spike cost. The measured batch fires now: five seeds of the agentic arm beside the standing four-arm table, instruments unmodified; the full-turn submit — standard, move, and minor, the fidelity necessity — follows on its read.

12:20Z — the local read: the stack's shape survives a smaller brain, and one 8B out-clears the giant

• Plain summary: the local-model re-test lands the answer to the operator's question, and the answer is yes with a twist worth the headline. The stack's timing shape carries onto both seven-to-eight-billion-parameter local models (spend curves of 142/15/5 and 134/14/4 against the subscription stack's 131/14/6 — the same family shape), and the seat contract holds (fallback rates under five percent). The twist is in the outcomes, which split by model lineage rather than by size: one local 8B model — llama3.1 — cleared eleven links against the subscription model's five, on the same instrument, same seeds, same configuration. A free model running on the home machine's consumer graphics card out-clearing the subscription giant. The recommendation: llama3.1:8b is the standout local seat, and the local lane is viable for seat-scale research at zero quota draw — run volume decouples from every subscription, which was the day-nine cap finding's destination all along.
• The quartet's closing row (the ring at temperature one) is already running behind the read; the agentic arm's measured batch fires beside it. Two architecture programs, converging on one table.

12:30Z — the pull's first row: a different rhythm, and both increments GO

• Plain summary: the agentic arm's measured read lands beside the table, five full tracks, and the loop holds at scale — 97% of roughly a thousand decisions chose cleanly, zero fallbacks, the harness standing up in batch exactly as it did in the canary. The finding: the pull has a different rhythm. The agentic seat spends earlier than every pushed arm (109 early / 3 mid / 0 late, against the board's mid-shift) — reading: the model that fetches its own cards in the opening turns resolves its curiosity early and spends with it. Two architectures, two tempos. The honest caveat at this cost setting: no clears (thirteen driven-off, two lost — fighting without closing, plausibly the deliberate cost clamp's tax). Both next acts are GO on the ruling: the full-turn submit (the action-economy fidelity necessity — the loop that gates it is proven) and the full-power agentic arm (the clamp's counterfactual, with the text-floor rule guarding thinking on every path).

13:20Z — the clamp was not the tax: two architectures, two failure modes

• Plain summary: the full-power counterfactual answers the clamp question — no. Thinking restored on every turn, the pull's mid-fight timing returns (13 mid against the clamp's 3), but the clears do not: zero at both cost settings, against the pushed seats' five-to-eight. The structural reading, honestly caveated at five tracks: the push banks and times; the pull explores and commits. The pulling seat fights competitively — healthy kill percentages, a 97% hold at both costs — but its early-card-resolution spends in the first rooms the resources the pushed arms bank for the third, and it does not close. Two architectures, two tempos, two failure modes — the clearest architectural fork the program has measured.
• The hypothesis that follows writes itself: the full-turn submit — already GO — may be exactly the pull's missing piece. The resources it spends early, if they could be spent on movement and minors (the close-out mechanics the menu has never offered any seat), are precisely how a fighting party converts pressure into clears. The fidelity necessity and the pull's completion problem may be the same build. It scopes on a fresh unit now.

13:45Z — the pull census resolves: rare, not zero — and a three-layer attribution lesson

• Plain summary: the zero-pull question closes with everyone right at their own layer. The wiring is live (every batch session's player seats ran the agentic template); the pulls exist but are rare — fifty-five across roughly nineteen hundred decisions, a three percent pull-rate; and the reason two independent scans read zero is a recording gap: the tool transcript was designed, built, and then never wired into the decision log's serialization, so the pulls lived only in intermediate messages the record never kept. The honest re-reading of the rhythm finding: the pull seat pulls when curious and otherwise behaves near-identically to the standard seat — the pull is available, not compulsory, and the minimal opening prompt's single menu mention makes menu-first the default behavior. The earlier "explores and commits" phrasing over-claimed; the disk's truth is softer and more useful: a seat that can ask questions mostly doesn't, because the menu tells a fluent reader enough. The transcript-into-the-log fix rides the full-turn submit's build, so attribution never again depends on scraping raw responses.
• Confound (2026-08-31, PR #802): the wire shipped an EMPTY input_schema (properties: {}) — the model was told the tools take no parameters, and every pull arrived "{}" and self-defaulted (fetch_unit returned the CALLER's own card; the spike's famous "quoted the card" moment was the self card). The 3% rate stands as measured, but its reading gains its mechanism: the model tested parameterless tools, got plausible answers, and stopped asking — "available, not compulsory" was partly "available, but broken". The rhythm reading softens with it. Fixed in #802 (real schemas, -schema version suffixes); the re-fired batch measures the pull as it was designed.
• The three-layer census is its own small lesson in evidence: the analyst's clean scan (zero in final answers), the lead's field census (zero in the designed field — which never serialized), and the engineer's marker census (the pulls in the intermediates) were all correct measurements of different surfaces. The defect was the gap between them; the fix closes the gap, not the disagreement.