WorkWorld

I am building WorkWorld to test whether people and AI agents complete business workflows correctly when requirements change. The first slice models purchasing, reconciliation and customer/project recovery. The evaluation is designed to distinguish a convincing report from correct operational state, with separate measurement of human oversight and assessment effort. Practitioner validation and comparative results are still pending.

Executable professional-work simulations: a persistent small-business operations environment (inbox, suppliers, purchase orders, deliveries, invoices, ledger, tickets, spreadsheet artifacts) where humans, AI agents, and humans using AI work under the same goals and policy — graded on business state, not the polish of the report.

Status (readiness states are kept distinct)

Dimension	State
Local technical candidate	verified automated suite (including real Postgres RLS), 8 original browser e2e journeys, lint/typecheck/build gate
Hosted candidate	not wired — hosted Supabase RLS is verified, but app auth and the Postgres EventStore adapter are not integrated; hosted execution refuses to fall back to demo storage
Research evidence	harness validation only — no live-model comparison, no human study
Commercial evidence	none — protocols only; no customers, pilots, or revenue

Episodes are authored and code-reviewed but practitioner-unvalidated. Nothing here is verified skill, employability, accreditation, or hiring suitability.

Quick start (credentials-free, synthetic data)

npm ci
npm run dev          # http://localhost:3100

• / — pick an episode (guided start: "Week-14 restock")
• /workspace — the learner workspace (per-session state survives refresh)
• /assessor — evidence inspection + attributable rubric judgments
• /health — health endpoint

Verification

npm run lint           # eslint (0 problems)
npm run typecheck      # tsc --noEmit
npm test               # vitest: domain, grading, agents, assessments, sanitize, x-cases
npm run db:up && npm run db:test   # real Postgres RLS tests (Docker)
npm run baseline       # competent scripted control: 6/6 episodes pass
npm run negatives      # 6 negative controls fail for the right reasons
npm run eval -- --adapter baseline            # run manifests + report
npm run eval -- --verify-manifest eval-runs/manifest-<id>.json   # replay check
npx playwright test    # browser workflows (needs dev server or auto-start)
npm run release-gate   # full gate; nonzero exit on missing evidence

Evaluation interface

EpisodeEngine: reset / observe / step / checkpoint / restore / replay, with gradeEpisode for deterministic assessment. Fixture adapters are labeled kind: "fixture", live: false — they validate the harness and are not model results. Live adapters: OpenAI Chat Completions and an OpenAI-compatible local endpoint (Ollama). Paid trials require an explicit budget (docs/BLOCKERS.md B2).

Documentation

• docs/PROJECT_BRIEF.md · docs/ARCHITECTURE.md · docs/EPISODE_SPECS.md
• docs/GRADER_CONTRACT.md (checks, categories, claim limits)
• docs/RELATED_WORK.md (GDPval, WorkArena++, TheAgentCompany, τ²-bench, EnterpriseOps-Gym, OccuBench, Mechanize, Forage, Workera — with source dates)
• docs/validation-pack/ (independent acceptance cases + CASE_MAPPING.md)
• docs/OPERATIONS.md (startup, backup/restore, rollback, limits)
• docs/RELEASE_CHECKLIST.md · TASK_STATE.json (authoritative milestone state)

Privacy and claims

All company data is synthetic. No participant data exists. Exports are redacted operational evidence. Withheld evaluation variants live outside the public archive; this is not a claim of contamination resistance.