What will they
build without us?

ObservedThe longest Dockyard run completed with memory and prior-knowledge active. The controller eventually paused fresh decisions at the protected daily request reserve.

LearnedSurvival stability improved, but request quota, multi-call planning loops and action-timeout churn had become the next constraints.

ChangedPrepared one-call evidence briefs, a one-hour planning cadence and hard admission controls for the free-request ledger.

ObservedResidents stayed alive and completed work, but planning continued until the OpenRouter daily request ceiling was reached.

LearnedThe free-model quota was being consumed by repeated planning attempts and tool loops faster than useful work was being produced.

ChangedAdded quota-aware attention throttling and designed prepared evidence to reduce ordinary decisions to one model call.

ObservedAfter an audited health-and-food recovery, all 20 residents passed preflight and ran for eight hours with memory and Minecraft knowledge enabled.

LearnedFresh cognition returned, but 3,836 corrective decisions and 954 recovery events showed that mechanical loops still dominated too much of the run.

ChangedCarried the stronger survival escalation forward and focused the next run on planning efficiency and quota use.

ObservedThe first longer memory-enabled run was manually stopped after stale queue labels, repeated background work, sparse knowledge reuse and two Rowan Brunel deaths.

LearnedA 30-minute lease alone did not guarantee meaningful reconsideration; failed survival and background loops could still block fresh planning.

ChangedEscalated repeated no-progress activity and unresolved low-food searches back to resident cognition.

ObservedThe no-cost preflight connected every resident but refused a paid start because four residents were below the minimum health or food threshold.

LearnedThe admission gate correctly prevented an unhealthy persistent-world population from entering another autonomous run.

ChangedRecovered only the affected residents under audit, preserved positions and required a fresh no-reset preflight.

ObservedThe refreshed OpenRouter treatment passed admission, preserved resident positions and produced real work orders and successful movement.

Learned1,089 invalidated actions and repeated recovery showed that path validity and just-in-time checks, not basic connectivity, were now the main problem.

ChangedTightened action verification, recovery feedback, survival priorities and bounded controller fallbacks.

ObservedThe first extended placement attempt exposed a severe death-and-respawn loop alongside bounded-tool planning failures.

LearnedSuccessful connection and placement were not enough; respawn safety, resident fitness and provider tool completion had to be tested together.

ChangedPaused the programme for runtime, provider and survival work before attempting another admitted live start.

ObservedControlled placement immediately produced repeated health loss, death and respawn events before autonomous planning began.

LearnedThe target world could connect 20 residents but could not yet place and retain them safely under the live survival rules.

ChangedRecorded the run as failed and moved the next iteration to explicit placement and respawn diagnosis.

ObservedA fourth short rehearsal connected all residents with healthy server timing and no model calls.

LearnedThe host remained responsive with the full roster, clearing the way for a survival-bearing placement attempt.

ChangedAdvanced to v5 with the same persistent world and controlled placement procedure.

ObservedThe full roster was connected again at the controlled opening time with clear weather and no model calls.

LearnedThe basic target-host topology could reproduce the opening conditions without stressing the Minecraft tick loop.

ChangedRepeated the check once more before moving from infrastructure validation into survival testing.

ObservedA second zero-cost rehearsal repeated the 20-resident connection on Dockyard with healthy tick performance.

LearnedThe initial infrastructure result was reproducible rather than a one-off successful connection.

ChangedContinued refining controlled opening conditions and founding placement evidence.

ObservedThe first placement rehearsal put the full 20-resident roster through the Dockyard controller with zero model calls.

LearnedThe target host could support the complete resident connection and telemetry path below the 50ms tick budget.

ChangedEstablished the target-infrastructure baseline and began repeatable placement iterations.