# Tokyo — Astra / Ultra execution prompt

Use this with the attached delivery report, original plan, reproduction archive and reviewed PR. Select Astra and Ultra in the client; do not infer an API reasoning enum from the UI label.

Continue Planetwide Tokyo from the exact current reviewed source. First read the delivery report and live repository instructions, then reconcile the PR head and merged main. Preserve completed work. Do not restart known failed sweeps or wait for pending Hamburg/Tulsa work when merged behavior suffices.

The target is 100,000 decimal bytes before general-purpose gzip, Brotli or Zstd. Count every required Tokyo-specific byte: geometry, names, descriptors, dictionaries, indices, corrections, fitted parameters, geographic LODs and source constants. Track genuinely generic runtime/generator code separately, along with mixed code, offline tools, transport size, decoded RAM and generation cost. Do not move Tokyo facts into shared code to exempt them.

Keep the frozen JP-13 selection of 158,235 ways and its topology/grade semantics. Preserve the reviewed geography and source coverage. Keep 100 landmark descriptors, 25 landmark names and 25 street names, a six-visible-name cap and names-off discovery. Preserve actual metres as the physical scale. Generated architecture and unknown heights remain labeled inference.

Use one integrator for shared runtime files and up to four useful independent lanes for representation experiments, fidelity auditing, asset/label research or evidence analysis. Give each lane frozen input hashes, a concrete question, disjoint outputs, a byte/fidelity acceptance test and a stopping condition. Return compact results and artifacts, not long transcripts. Do not delegate blocked work solely to fill slots.

Begin with the measured frontier: current runtime city data 2,861,052 bytes; best new offline road packet 2,436,242 bytes; best feature packet 97,600 bytes. Their sum with the unchanged 725-byte profile is 2,534,567 bytes, not an integrated game package. Existing byte-varint resolution tuning cannot reach 100 KB. Require a new causal mechanism before rerunning experiments.

For any recipe, serialize the complete recipe plus membership, exceptions and residuals, and compare with a literal fallback. Verify an independent decoder against every original runtime ID, coordinate occurrence, name, flag, layer and coordinate frame. Preserve source-node provenance separately; coordinate equality must not create junction identity. Measure meaningful collision, clearance, geometry and runtime tradeoffs before integration.

Keep changes local to existing readers, city profiles and descriptor consumers. Introduce an adapter only for a demonstrated incompatible interface. Avoid speculative abstractions and shared-file churn. Refresh merged main at deliberate integration checkpoints and record exactly which Hamburg/Tulsa capabilities are included or pending.

Reserve a final Free Space period for three to five bounded ideas suggested by the actual results. Preserve the best verified baseline first. Keep valuable negative results. Stop a candidate when its mandatory streams already defeat the target; do not confuse that with a mathematical impossibility proof.

Persist through ordinary reversible work, recovery and packaging using the authorization already present. Ask only when a consequential unresolved choice requires the user. Save coherent artifact checkpoints during long work. After required gates pass, repeat testing only for a concrete remaining risk. Do not silently relax a gate.

Deliver the current review branch, real playable build, source/codec reproduction bundle, complete byte ledger, measured resolution/recognition tradeoffs, actual gameplay evidence and a candid acceptance state. If no candidate meets both bytes and fidelity, finish the faithful build and state the unmet target clearly. Do not substitute a smaller geographic scope or a plausible invented city.

References checked September 6, 2026: [Astra model guidance](https://developers.openai.com/api/docs/guides/latest-model?model=gpt-6-astra), [ChatGPT subagents and Ultra](https://learn.chatgpt.com/docs/agent-configuration/subagents). Preserve this task-specific prompt rather than layering duplicate generic instructions onto it.