Studio: design system, reports, live views, Genesis as an autonomous scientist, identity file (features 007-020)

[LucasWaki]

What changed

This branch carries the Studio work of 8 and 9 Sep on top of the benchmark foundations (features 007 to 011), and closes with feature 019 and the first item of feature 020.

Studio (features 012 to 018)

• One design system: a token sheet over Radix Colors, IBM Plex and Lucide vendored, no inline styles, no external resources. A CSS test fails on any hex colour, !important, inline style or external URL.
• Measures computed once on the server from stored results and events: pass rate with Wilson intervals, pass^k, violations, false completion, paired deltas, complete cost. Charts drawn as SVG styled by class.
• Reports as the front door: run and round reports that read verdict first, caveats written from data only, a historical Bare baseline reused when a finished run matches on model, effort and task hashes, a Terms section, public audience by default.
• Live views without duplication, a sticky inspector instead of the scrolling right frame, the weekly budget chip.
• The product graph panel shows any bench version as products with their stored business actions, and the live Product Graph Monarch Enterprise uses, read from Feature Discovery over GET only. Nothing in the bench can write to Monarch.

Genesis (feature 019)

• Cards as inputs: a link, a run id or a sentence becomes a card with a question. A watcher works queued cards one at a time under US$ 2.00 per card and US$ 6.00 per day, counted from settled receipts, and never launches a run. It ignores runs and sources older than its first start.
• A daily index of the Monarch checkout at the declared build, a change record filed in the library, and five read-only code tools. Graphify is used when installed.
• Bounded memory: LAB.md and MONARCH.md at 2,500 characters each, notes per card, an FTS5 record, seven-day probation, thirty-day decay, pinned entries that never decay, a nightly consolidation turn under US$ 0.50.
• A Trello-like board, a fixed library topic list with Other last, and a daily scheduler.

Feature 020, first item: the identity file

• SOUL.md holds Genesis's voice, priorities and what it never does. A starter text is written on first start. Only a person edits it, from the Memory tab. It is the first block of every prompt and no Genesis tool reaches it.

Also

• Windows long-path handling in evidence writes and a short pytest temp root.
• wb studio takes its port from the environment.
• Research log, plan, improvement ledger and handoff docs of 8 and 9 Sep.

Why

The plan of 9 Sep (docs/AI-LABS-IMPLEMENTATION-PLAN-2026-09-09.md) mapped the Studio into seven phases. The Genesis design follows the research note docs/AI-LABS-GENESIS-CODE-AWARE-MEMORY-2026-09-09.md; Lucas's choices for what Genesis grows next are in its Part 3.

For the reviewer

• Python suite: 1832 passed, 3 skipped, 1 failed. The failure is tests/test_run_config.py::test_shipped_create_run_pilot_hash_is_unchanged, which already failed at main and waits on Lucas's sign-off to re-pin the hash (pre-registration rule).
• Browser suite: node tests/browser/suite.cjs, 27 checks pass. Screenshots under tests/browser/snapshots/ are for review, not pixel diffing.
• The watcher spent nothing in the wild, but at its first start it backfilled cards for every past run and source and started four turns that failed the ledger check. Both causes are fixed in this branch.
• Carlos's two doc commits on main (handoff of 8 Sep, deferred item on the public front door) are merged in; the one conflict, in deferred.md, kept both sides.
• Kept out by .gitignore: the Monarch Enterprise checkout snapshot under artifacts/enterprise-local, Genesis test sessions under artifacts/genesis-check, and local scratch folders.
• Nothing here changes the methodology in PLAN.md §1. No task prompt, data or approval rule was edited.

🤖 Generated with Claude Code