Dave audio and voice refinement
Full Adventure conversations retain the current soundtrack at 35% of the user's music volume. The mix fades down over roughly 0.35 seconds and up over 0.7 seconds at the normal update rate. Simulation and ordinary board effects remain paused. Dave's short unsolicited remarks keep normal gameplay/music running. The mix is independent of saved user volume and focus muting. Normal menus still pause music; Dave-owned pauses also cover native saves during a gift or other interaction.
The character prompt now explicitly asks for occasional deliberate grammar or spelling slips, fragments and corrections, with language-specific guidance for English, Chinese and Spanish. Its comic choices are tied to the current exchange: ridiculous household logic, injured salesman pride, literal misunderstandings and grudging admissions. It forbids em/en dashes in dialogue and repetitive stock jokes. Names, prices, effects and urgent tactical guidance must stay clear. It does not change the offer catalog, native balance, billing or campaign memory.
Context size
Before this voice update, the last day's 15 recorded Sonnet 4.6 attempts had 3,945–4,521 input tokens (mean 4,345), with mean output 59 tokens. These are recorded usage values, not a tokenizer estimate. Only model-level aggregates were read.
The revised character section adds 1,685 characters. Actual service assembly, captured with an inert provider and no DB writes or paid calls:
| Part | Fresh chat | Pool with 36 plants and recent dialogue |
|---|---|---|
| Fixed system instructions | 12,721 chars | 12,721 chars |
| Story, rules and memory context | 5,473 chars | 8,052 chars |
| Native lawn snapshot | 498 chars | 2,918 chars |
| Total serialized messages | 19,106 chars | 26,329 chars |
| Rough English estimate (chars / 4) | 4,777 tokens | 6,583 tokens |
These last estimates are not Sonnet token counts and vary by language/tokenizer. The current chapter's direction/knowledge contributes 503/409 characters in these samples; all five chapters are not inserted on every call. Context also includes up to 12 recent messages, bounded remembered player statements, notes, witnessed milestones and tin receipts. The full stored notebook is not sent verbatim. Generation remains capped at 800 output tokens and speech at 900 characters. Cached/free opening and reopening Dave do not generate new calls.
Verification
Native WASM build, 47 browser-shell tests, 50 NPC/server/billing tests, full platform build and typecheck passed. The real browser audio probe samples the engine's mixed Web Audio output after its callback, with isolated account/provider fixtures. It confirms audible quiet music while conversation ticks stay frozen, normal volume/play on return, continued music across a gift/save, and silence in the ordinary native pause menu. Native volume-slider checks also passed: zero stays silent across opening/closing Dave, and a newly chosen volume still gets the quieter conversation mix. The stronger voice is an instruction to the model; these fixture tests do not constitute a live Sonnet writing-quality evaluation.
Published
Platform bee6d65f0 deployed successfully as Railway deployment cb125659-410e-40e3-9d43-5393db8b7a18. Game 5eac9df, release c9451cb1e4, is uploaded to S3 and admitted alongside prior supported builds. The public HTML matches the staged file. Production guest audio testing confirmed quiet, audible music through the free opening and reopening, normal-volume advancing gameplay on return, zero paid requests and zero page errors. Regional hosts were unchanged. Local evidence: output/dave-ui-rebuild/audit-dave-music-production.log and audit-dave-music.log in the main workspace.