# GUIDE-DEV-20261006-01 — review: speed, token efficiency and AAA detail standards

Owner request (2026-10-06): review the wiki's game development direction, suggest how development can be faster and
more token-efficient, and add every small AAA-level detail (sound, effects, optimisation and so on).

## What was produced

| File | What it is |
|---|---|
| `packages/game-data/data/polish-standards.json` | The AAA detail catalog: 127 proposed details in 11 areas (37 core, 71 slice, 19 polish), plus crowd coalescing values, the audio mix table, the surface table, the platform capability table and five perf scenes |
| `docs/production/POLISH.md` + `apps/wiki/components/widgets/Polish.tsx` | Wiki page; every table is rendered from the data file |
| `docs/process/EFFICIENCY.md` | Measured token facts, proposed token rules, proposed changes for faster development |
| `tools/ctx.mjs` | Context packs (one skill, monster, zone, catalog slice or roadmap task) |
| `tools/task-log.mjs` | Automatic Claude Code session attribution, multi-session totals, turns and average context at `finish` |
| `packages/game-data/tests/polish-standards.test.mjs`, `agent-tools.test.mjs`, `apps/wiki/tests/polish.spec.ts` | Tests |

Pointers and corrections were added to VFX, Audio, Performance, the skill gate, the development guide, How AI builds
Zoen, Testing, Decisions (OPEN-11 to OPEN-14) and CONTINUE_HERE.

## Token measurement (section 1 of the efficiency page)

`token-analysis.mjs` in this folder reads the transcripts through `tools/token-ledger.mjs`. Output of the run on
2026-10-06 (transcripts of 2026-10-03T11:24Z to 2026-10-05T18:57Z, the first turns of this task included):

```text
messages 4096 | first 2026-10-03T11:24:03.525Z | last 2026-10-05T18:57:23.469Z
tokens: in 8,250 out 5,428,579 cacheWrite 15,722,897 cacheRead 1,917,825,624
true cost total $ 557.43 {"claude-sonnet-5-5":74.05,"claude-opus-5-5":479.14,"claude-haiku-4-5-20251001":0.15,"<synthetic>":0,"claude-fable-5-1":4.09}
time {"wall":55.56,"active":26.8}
cost split $ { input: 0, output: 101, cacheWrite: 73, cacheRead: 384 }
context per message (tokens): p25 263,588 p50 465,391 p75 666,824 p90 829,022 max 965,526
share of cost in turns with context >50k: 100% >100k: 98% >150k: 96%
sessions 12 | top 2 by cost: $238 (1,559 turns, 28.4 h, max context 965,526) · $196 (1,443 turns, 12.5 h, max context 964,680)
tasks 229 | with --session 73 | sum of per-task cost $ 1039 vs true total $ 557.43 | ratio 1.86
unsessioned finished tasks 156 | overlapping pairs 117 | tasks involved 85 | their summed cost $ 602
messages outside every task window 127 $26
task duration minutes: p50 6 p90 49 max 544 | over 60 min: 18 of 229
```

Re-run: `ZOEN_LEDGER_CACHE=off node logs/evidence/GUIDE-DEV-20261006-01/token-analysis.mjs` (numbers grow as new
sessions are added).

## Facts checked on the web (2026-10-06)

Unity manual pages for the GPU Resident Drawer, dynamic resolution, STP, audio in Web, Web technical limitations,
pipeline state warm-up, the Audio Random Container, Adaptive Performance and Web native plugins; Sony ASWG-R001
(loudness); WCAG 2.3.1 (flashes). URLs are in `polish-standards.json → sources` and on both wiki pages. The ASWG-R001
PDF could not be opened by the fetch tool; its numbers (−24 LKFS home, −18 LKFS portable, −1 dBTP) come from search
results that quote it.

## Not done

- Nothing in the catalog is implemented or measured: no game runtime exists.
- `CLAUDE.md` and `KICKOFF.md` are unchanged; the token rules wait for OPEN-13.
- `CONTINUE_HERE.md` still carries its round history (slimming it is proposed, not done).
- The full Playwright suite was not re-run; only `tests/polish.spec.ts` on desktop and mobile.

## Captures

`polish-desktop.png`, `polish-mobile.png`, `polish-sfx-*.png`, `efficiency-*.png`, `polish-360.png` (written by the spec).