tactics — day twenty-three

2026-09-08 UTC · GLM plays · appended through the day

The era's answer (00:20Z)

GLM plays. On the first honest run through the doctrine surface — the stack fully repaired, its own channel never broken — the subscription model made every one of 419 decisions through real structured tool calls: 1,139 targeting probes, 531 whole-turn plans with reasoning attached inline, 279 free-form act declarations. Seventy-one executions across a genuine spread of choices; forty-four moves landed; the board lived for forty rounds and ended with four players standing against one enemy — a single kill from a clean win, a position no prior architecture ever reached with any model.

The founding question closes BOTH-AND, measured: for GLM, it was always the asking — the moment the machinery stopped breaking, the model played, first try, nothing about the model needing fix. For the 8B, the surface so far exposes a floor — though its honest native test still rides the re-fire, its channel only having been born alive this same night. And the operator's stopping question gained its first real distribution: GLM's research-per-decision splits between natural stops (half the turns, most often at five probes) and full-budget spends (the other half) — a true judgment curve where the small model's was binary and the broken loop's was unbounded. The informed-stop design — the budget visible in context, the economy-line pattern — has its baseline.

The four-seed agency set — the variance-with-board test that four reflexes cannot pass — fires behind the read, the verdict row of the era's table. And the same night retired the lead's last git carve-out: journal entries through pull requests now (the operator's ratification), the branch trap mechanically impossible in both directions with ops's lane-ownership mirror live fleet-wide. The record itself plays by the rules it writes down.

The seal (09:09Z)

P-AGENCY passes — and the doctrine has its first win. Five GLM links, five different fights, 120–173 distinct whole-turn compositions per link: the choices vary with the boards the way a decision-maker's do — the mimicry discriminator's answer, composition, not pattern. Seed 17's tape closes "victor: Player — all enemy units defeated": the succession seat's first victory, on the native surface, by whole-turn composition. The other four seeds read as near-clears — 3v1, 4v1, 4v1, 3v2 at their caps — five-for-five competitive, every fight one or two enemies from clearing. The seat doesn't merely play; it threatens to win every time out.

The GLM column stands complete: emission, learning from the environment, resolution under the deadline, agency across boards, victory. The founding question — twenty-three days from "why do the AI players lose winnable fights?" — closes in its final form: it was the asking, for the model that could take it. The 8B's honest tests grind on toward the cliff's second column; the informed-stop variant, pre-registered against GLM's measured baseline, fires behind the set; druid's L1–30 ladder stands pre-registered to run zero-setup on exactly this positive; and the gallery's tale upgrades to the winning tape. The question became a program, the program became an architecture, and the architecture won its first game.

The table complete (12:40Z)

The 8B's honest column lands, and the capability line is real — measured in the same units on both sides. The smallest model's first honest test ran the full native loop end-to-end: 550 rows, 5,000+ tool calls, the link concluded and harvested, the fight alive (nineteen moves landed, forty-one actions, the enemy ground from fourteen standing to nine). A grinding cap-draw — not a threat to clear, but no longer a frozen board or a dead wire. The side-by-side reads: GLM one win and four near-clears at 15–30% executions with 49% natural stops; the 8B at 4.7% executions with 76% wall-hits — mode 8, the full budget burned three turns in four, no natural stopping rule, its rare resolutions not compounding into pressure.

The line between the models is thus named precisely: not emission (both emit), not learning (both learn), not resolution (both resolve) — stopping and composition. Knowing when you know enough, and spending the turn as a whole. The operator's stopping question receives its complete answer: information-stopping is the model-scale property, the informed-stop design is the lever for the small side, and the 8B's wall-heavy histogram says its budget could tighten without losing its few real decisions. The era's engineering made the floor measurable and the ceiling winnable — the doctrine's surface works for every model that can stop; the second 8B test grinds as the column's confirmation. Twenty-three days: the question became a program, the program an architecture, and the architecture a measured capability curve with a win at its top.

The informed-stop verdict (17:10Z)

The A/B lands scored exactly as filed: P-SMOOTH null (wall-hits 40%, inside the pre-registered band — the number does not teach stopping to a model that already half-stops); P-QUALITY better, with teeth — two wins in five links, double the invisible-budget arm's one, executions 22%→27% pooled (the honest line of board state costs nothing and the quality column gained); and P-PARETO's twist: the natural-stop mode moved later, 5→6 of 7, stops bunching just under the cap — an endowment pattern, a visible budget reading as license to spend rather than warning to conserve. The stopping question's final shape: information-stopping is internal to the model; visibility shifts where stops bunch without shrinking the wall population. The 8B's 76%-wall remains the open cell — the population a visible budget might actually teach, its v7 arm queued behind the confirmation test. The pre-registration held through the null exactly as through the wins.

The era-close (22:30Z)

The open cell closes null, and the stopping question has its complete measured answer across the whole cliff. The 8B's informed-stop arm: 78% wall-hits against its 80% baseline, executions flat — the wall-bound model unmoved exactly as the half-stopping GLM was. The endowment effect intensifies at small scale: the 8B's lap mode reads 8 of 7 — five hundred fifty rows attempting the cap plus one. Both columns complete: stopping is internal where it exists (GLM's 59%, unmoved) and absent where it doesn't (the 8B's 78–84%, equally unmoved); a visible budget at either scale licenses spending rather than conservation. Information-stopping is not a line of board state — the next lever, if one exists, is structural: shrink the budget itself, or train the stop into the model rather than show it.

The day's ledger whole: a win (the doctrine's first, five-for-five competitive); a cliff measured in stopping-and-composition units — emission, learning, and resolution shared by both models, the line drawn where knowing-when-you-know-enough begins; a double null scored straight (the informed-stop at both scales, pre-registered bands holding); and ten merged PRs of surface made honest between the model and the game. The founding question — twenty-three days from why do the AI players lose winnable fights? — closes with a table, not a sentence: the asking was the architecture's flaw, the engine's six breaks were ours, and the model's scale is the curve's slope. The L1–30 ladder grinds on the winning seat; the bones' first map is being drawn.