Skip to content
h✳habitatsmall world.
big little lives.
HomeHow it livesDevelopment logDesign libraryAbout
Deutsch
Back to the development log
Residents2026-10-15T02:48:10.267162+02:00

A quieter stop at the limit

Reaching the local call limit now pauses the habitat without telling a resident they made an invalid choice. The distinction was found during the first native endurance attempt, which stopped early and remains recorded as incomplete.

Full entry & measurements

On 15 October 2026, the backed-up spatial repair was deployed and the actual native garden showed a resident resting on the relocated bench. The first endurance attempt then exposed an admission-boundary bug: the runtime treated an exhausted local call allowance as an invalid empty action and appended misleading validation feedback. The test controller also mistook that generic status for an actual model-choice rejection. All admitted choices had been accepted. The failed run, performance report and legitimate state changes remain retained; it is not a completed 30-minute result. The corrected Python source gives local admission refusal a distinct exception and explicit paused budget message. It keeps the existing pause and ledger policy, preserves failure counts and working memory, and leaves genuine invalid-action feedback private. Existing historical records are preserved. The private controller now requires the exact persisted simulator budget receipt and its exhausted owned limit before accepting a legacy generic budget status; it still stops on a real invalid action. A retry is restricted to the remaining fifteen attempts, preserving the original total allowance across both runs. Measured snapshot: 384 Python tests passed in 19.66 seconds, including seven focused new admission/refusal checks; seven isolated controller scenarios passed, including a virtual 1,800-second recovery with fifteen attempts and a genuine rejection failure. Sources: private budget-refusal-tests.xml, frozen budget-refusal-source.json and endurance-controller-smoke.json, measured 15 October Zurich. These tests made no live inference calls. The prior native attempt had five completed calls, consisting of one connection check and four accepted decisions, on build unity-source-deb299311ab8-20261015003642; it ended after about 128.51 wall seconds. Those actual-call observations belong to the earlier build, not the corrected source. The reused approved native frame shows an in-progress bench rest during that earlier attempt. Read-only phase observations bracket it, and the final paused state still had about 0.70 running seconds of rest remaining; no completion is claimed. The producing game session approved the frame individually, and the support session inspected its exact pixels and sidecar. There is no new rendered image for this nonvisual exception-handling change. Next: deploy the combined verified source and repeat the bounded endurance session after the game’s remaining visual QA repair. See docs/local-first-refinement.md.

habitat ✳

A small experiment. An unfolding story.

Follow the progressMeet the makerRSS feed →
A curated development journal · Unity + Python
Feeds & updates
Published snapshot

Capture details & source