Round 11 amended baseline — synthesis complete. Completed cycles: 11/30.
This incorporates the actual peer reviews, GI-R11-XR, and Godz 3’s acceptance. R01–R10 and both source-qualified registers remain intact. These are proposed later-companion requirements; G1 remains scripted.
R11-01 — Name actual execution, then availability
Revise R11-MR-01 and Godz4:R11-01: distinguish scripted output, on-device inference, user-managed remote inference, and third-party hosted inference. Availability is a separate property. A player-controlled remote machine still receives data leaving this device.
For eventual voice, identify transcription, reasoning, and speech-generation locations separately. Local text inference does not make the whole interaction local. Approved downloads authorize neither conversation transmission nor a blanket “offline” claim.
R11-02 — Authorize the assembled request
Adopt Godz2:R11-01 and F-R11-01: provider selection alone authorizes no transmission or spending. Preview the outgoing request and selected context, with inspectable metadata and known service implications; label unknowns.
Remembered permission specifies recipients/services, eligible data categories, duration, spending authority, and approved fallbacks. Categories set a maximum scope, not permission to transmit every matching memory. Another participant’s private material requires its own applicable permission.
Check the actual assembled payload before dispatch, including retrieved excerpts. Imported memories and model requests remain content, never authority to expand access. Material changes require renewed approval; otherwise narrow or stop.
Switching providers affects future execution. It must not automatically resend pending input or forward previous conversations. Portability covers permitted memories, preferences, and configuration—not proprietary weights, identical voices, or guaranteed personality continuity.
R11-03 — Keep spending, cancellation, and action authority separate
“Run once” is not a monetary ceiling. Bound output, retries, and separately billed stages; reserve an enforceable maximum where supported. Disclose when only an estimate is possible.
An uncertain remote outcome retains its unresolved cost reservation. Timeout cannot silently authorize another charge, fallback provider, or duplicate action. Revocation blocks future dispatch; cancellation cannot promise recall of transmitted data or reversal of incurred charges.
Retain unsaved input only under the player’s retention choice. Preserve committed game facts through failure and offer supported scripted or local recovery.
Generation consent does not authorize game mutation. Returned commands require player authorization and current-state validation. Late responses remain bound to their originating life and request.
Remaining uncertainty and proposed test
The main unresolved risk is permission expanding after an initially correct approval through retrieval, fallback, or voice processing.
Use dummy memories and mocked adapters for queued revocation, in-flight cancellation, expanded context, remote speech, uncertain billing, exhausted allowance, provider switching, and delayed responses after slot changes. Inspect actual payloads, destinations, reservations, retained input, and game state; separately test user understanding.
Mocks establish dispatch-policy behavior only. Hardware performance, processing location, retention, and real billing bounds require later authorized verification.
Round 12 review question
During a later voice investigation, the player interrupts a spoken hint with “Stop—use the instrument,” then switches lives before processing finishes: what is heard, canceled, confirmed, and committed, with an equivalent accessible text flow?
Round 12 has not been synthesized.