Grok workstream: maximum offload (2026-06-12, evening)

Wedge’s directive: offload everything offloadable to Grok (separate budget); Fable keeps only form, story, and judgment. Recorded as tsumugu/PRD-Grok-Workstream.md — a standing brief whose sections paste directly as run prompts.

Decisions

Campaign: TOCFL A1 → A2 → B1. Corrected on the record: Band A was never verified complete — the queue was a jieba frequency proxy. Run R0 fetches the official 華語八千詞, writes sources/tocfl/coverage.md (the band ledger), and the worklist authors A-band gaps before B1.

Grok’s four lanes (§R/§M/§Q in the workstream PRD): cribs for every target character; mechanics — component registry (native 部件 names, variant/position tables, stroke order), sound-series tables, level/freq patch files applied by script; and fact-audit of every authored batch. All output to sources/, never entries/; WORKLIST printed before producing; reports must paste literal command outputs (after a run claimed completion it hadn’t fully delivered — disk is the only witness).

QA split by independence, not by agent: Grok audits facts in Fable’s entries (phonetics vs cribs, bands vs official list, drift vocabulary, series consistency) — flags only, never edits; Fable audits voice (random-10 register sampling). Grok never grades stories (Round 10 written into its brief); nobody audits their own work.

What stays where: Fable — form, story, role judgments, ⚠ CONFLICT adjudication, register QA, contracts. Scripts — zhuyin, senseRef, patch application, crib merge, validation. Wedge — morning skim, push, taste strikes, contract changes.

Also fixed today

Grok run-1 confusion resolved: the 30 cribs were real (old prompt copy, wrong folder — salvaged to cribs-grok/); RUN2 adds WORKLIST-first and disk-verified reporting. Cursor prompt aligned to the TOCFL worklist.

Outstanding

R0 (TOCFL fetch + ledger) → R-runs + Cursor mirrors → author batches; M-runs on idle Grok budget; first §Q audit after the next authored push. Sentence-pass PRD still future (Grok will gather candidates when it opens).

Addendum (same evening) — roster cut to three

Wedge simplified the roster: Grok (facts: cribs, mechanics, audits), Opus (sole author, words and characters), Fable (orchestration, taste QA, contracts, with Wedge). Cursor, Sonnet, and Composer retired. Priced and accepted: the dual-crib cross-check goes single-source (safety now = Opus re-verification + Grok post-batch audit, two layers not three), words author at top-model cost, and rendering/voice become future Opus phases with Composer’s work orders standing as specs.

Reversed within the hour: five-agent stack retained on reflection (the cross-check and the cheap word lane justify their coordination cost). PRDs, CLAUDE.md, and memory restored to the full roster.