{"data":{"kind":"file","path":"README.md","version_id":"lytstbvncusuhtdi0iai0bsy","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":3916,"modified_at":"2026-11-17T13:26:52.128000","content_hash":"6b07d621ee5aeec1953a662bbbfa9c7849e46b1c9f55da562afca861a7b9cd36"},"entries":[],"content":"> **v0.6.1 Prime Hub release candidate:** Public source snapshot with the Tab-Journal preflight hotfix, Verifiers v1 environment loader, and known fairness caveat. Read [PRIME_HUB_RELEASE_v061.md](PRIME_HUB_RELEASE_v061.md) before publishing.\n\n> **v0.6.0 (expanded public A/B):** A/B runner supports four distinct public task families with pinned frozen v0.4.7 graders: `interop-workbench`, `tab-journal`, `secure-inbox`, `resilient-board` (24 deterministic public task instances, seeds 0–5). Includes per-family public smoke diagnostics, multi-family dry-run, preflight without Codex, SHA-256-pinned fixture manifest, validity-aware paired summaries, and v0.5.5 timeout guard. See [BROWSER_AWARE_V060_EXPANDED.md](BROWSER_AWARE_V060_EXPANDED.md). **Exploratory public research, not a blind evaluation or RL training implementation.**\n\n> **v0.5.0 (experimental):** Public Code-Only vs Browser-Aware Codex runner for `interop-workbench`. See [BROWSER_AWARE_V050.md](BROWSER_AWARE_V050.md). Frozen v0.4.7 verifier hashes are enforced. This is **not** an RL training implementation.\n\n# ModernWeb-RL v0.4.0 — Public Research Environment\n\nA multi-turn coding-agent environment for modern web repair. **10 public task designs, 60 public instances**, deterministic browser-verifier reward, fault-injecting API tests, and a three-engine CI readiness gate. Built independently from GoogleChrome/modern-web-guidance.\n\nThis is the **public, solver-safe distribution**: it does NOT include authored reference patches, private holdout designs, hidden grading scripts, or private calibration workflows. To keep genuine held-out task-design generalization measurable, evaluate on additional private tasks with a physically isolated trusted grader. See `RESEARCH_PROTOCOL_V04.md`.\n\n## Public development commands\n\n```bash\nuv sync --extra test\nuv run pytest -q tests/test_fixtures.py tests/test_research_integrity.py\nnode --test tests/v03_api.test.cjs\nuv run python -m modernweb_rl.local stage interop-workbench 0 ./workspaces/interop\nuv run python -m tools.agent_baselines --probe\n```\n\n## Real HTTP-origin preflight (three engines)\n\n```bash\nuv run playwright install --with-deps chromium firefox webkit\nnpm install --no-audit --no-fund --prefix /tmp/mwr-browser playwright-core@1.57.0\nNODE_PATH=/tmp/mwr-browser/node_modules MWR_BROWSERS=chromium,firefox,webkit \\\n  node tools/e2e_preflight.cjs reports/preflight.json\n```\n\n`guard=passed` is required for verified HTTP-origin scoring. The source package's `MWR_PAGE_MODE=memory` is explicitly a limited local simulation; it is **not** equivalent to three-browser E2E. The CI job `e2e-origin-preflight` refuses to pass on a missing engine or blocked loopback navigation.\n\n## Coding-agent baselines\n\nThe supplied runner invokes **installed** Codex and Claude Code CLIs (never mocks a model). To run it you need model access, a pinned model ID and a secure disposable VM:\n\n```bash\nuv run python -m tools.agent_baselines --suite public --agents codex,claude \\\n  --codex-model YOUR_CODEX_MODEL_ID --claude-model YOUR_CLAUDE_MODEL_ID \\\n  --rollouts 3 --browsers chromium,firefox,webkit\n```\n\nThe authoring container does not have CLI/model credentials; no real-agent performance measurements have been produced. Running these CLIs may incur usage charges. Never publish access credentials.\n\n## Blind holdout policy\n\nHidden tasks are exported only at assessment time using the owner's separate private tool. The agent receives a ZIP containing only editable source, fixture JSON and task prompt, then returns a separate submissions ZIP. It cannot access grader source or authored references on its machine. **Do not put hidden tasks or reference code into this public source tree.**\n\nSee `BASELINE_EVAL.md`, `RESEARCH_PROTOCOL_V04.md` and `V04_VALIDATION_REPORT.md` for reproduction, minimum evidence and limitations. Neither Environment Hub acceptance nor bounty payment is guaranteed.\n","encoding":"utf-8","truncated":false,"total_bytes":3916},"status":null}