v1.0.0 — policy as product

Release v1.0.0 — policy as product

Date: 2026-08-14
Baseline prior: v0.3.0 (8feb31a) file replay + concurrent isolation demo
Host: https://pax-orchestration.vercel.app (project pax-orchestration only)

LIVE

• Loadable default policy JSON characterizing v0.3 handoff graph and budgets (ADR-0007)
• Restrictive fixture policy that changes happy path (forbid-research-to-writer)
• Chaos isolation suite n ≥ 40 with easy/medium/hard predicates
• Isolation simulation n=1000, swap_rate=0, writer_only_violations=0 (JSON+HTML)
• Policy UI (/ui/policy) + generated capture
• Load probe p50/p95 for fake POST /v1/tasks (docs/assets/load.json, single isolate honesty)
• Trace pack (JSON / directory / zip) with policy hash, task, seed, trace, result; UI import + round-trip tests
• CASESTUDY ≥1500 words; DEMO-DAY 25 min script
• Free path still fake specialists, $0, CI with empty keys

Still PLANNED (honest)

• Multi-instance durable run storage (KV / Postgres / object store of every run)
• Redis / shared session fabric
• Real LLM specialists as default free path
• Multi-tenant policy SaaS or remote policy registry
• Multi-agent answer quality claims from fake specialists
• Capacity planning from single-isolate load numbers
• Any host other than pax-orchestration / pax-orchestration.vercel.app

Non-goals this tag

• No Redis
• No inventing durable server storage for recycle demos
• No pax-mao project

multi-agent-orchestration v0.3.0 — offline file replay and isolation

multi-agent-orchestration v0.3.0 — offline file replay and isolation proof

LIVE in this release

• GitHub Release object for v0.2.0 already published; this tag is the durable-proof follow-on.
• Replay from file: browser “Load trace JSON” renders the Replay panel with no server
  round-trip after load; python -m mao.replay path.json exits 0 on valid schema-1, 2 on invalid.
• Schema-1 export download (/ui/tasks/{id}/export.json) plus committed fixture and tests.
• ADR-0006: client-side file replay instead of Vercel KV/Postgres this week.
• Concurrent fake tasks keep Writer text and traces isolated under a shared process store.
• Recycle-equivalent (RunStore.clear) still allows offline file replay.
• Optional P2 HTTP Research remains fail-closed (unset → capability_missing; 5xx → degraded).
• DEMO-DAY tightened to a 15-minute run → download → 404 → load file story.

Deliberate limits

• Server retention is still process-local last-128; recycle still 404s server lookup.
• File replay is client durability, not a multi-instance audit database.
• Specialists remain deterministic fakes unless HTTP Research is explicitly selected and configured.

multi-agent-orchestration v0.2.0 — hosted replay and chaos isolation

multi-agent-orchestration v0.2.0 — hosted replay and chaos isolation

LIVE in this release

• Credential-free public console at https://pax-orchestration.vercel.app.
• Versioned trace_schema: 1 envelopes with deterministic logical offsets and seed recording.
• Process-local last-128 run lookup plus a Replay panel and generated ui-replay.png.
• Public typed stop reasons and explicit Writer/system result ownership.
• Four offline chaos goldens: specialist crash, two Critic rejections, max handoffs, and
  attempted Writer impersonation.
• DEMO-DAY.md, expanded case study, authority/stop matrix, 53 tests, strict mypy, Ruff, and
  the free scorecard with no provider credentials.

Deliberate limits

• Retained runs, traces, and metrics disappear on instance recycle or FIFO eviction.
• Specialists are deterministic fakes; this release makes no model-quality or multi-agent
  uplift claim.
• Remote process isolation and durable audit storage remain future boundaries.

multi-agent-orchestration v0.1.0 — bounded specialists, visible policy

multi-agent-orchestration v0.1.0 — bounded specialists, visible policy

The first public release of portfolio project P3. It demonstrates coordination mechanics on
a deterministic, credential-free default path; it does not claim model-quality uplift.

LIVE in this release

• Typed Research, Critic, and Writer handoffs over an in-memory bus.
• Writer-only user-facing final output, global handoff budgets, at most two Research retries,
  and explicit degraded or budget_exhausted terminal states.
• Ordered JSON-safe traces without full task text or fake-path wall-clock timestamps.
• Library, POST /v1/tasks, JSON CLI, and accessible server-rendered task console.
• Optional P2 HTTP Research behind explicit AGENTIC_RAG_URL; fake remains the default.
• Fail-closed handling for missing configuration, timeout, transport, HTTP, and contract
  failures. P2's raw trace is not copied into P3's result.
• Twelve deterministic routing/accounting goldens plus two zero-network boundary cases,
  reported as provider="fake" and billed_usd=0.0.
• Python 3.12 CI gates for ruff, strict mypy, pytest, and the evaluation runner.
• Non-root Docker image and standalone P3 Compose service.

PLANNED or deliberately not claimed

• Hosted LLMs as the default, a hosted public demo, and the P5 multi-service platform.
• Remote-process specialist isolation or a claim that separate models improve answer quality.
• Live P2 availability in CI. The live integration smoke remains opt-in and skips when P2 is
  unset or unavailable; mocked tests prove the boundary contract.
• Quality or factuality gains from the fake specialists. The goldens measure policy,
  ownership, routing, budgets, trace shape, and accounting only.

Reproduce the free path

python -m pip install -e ".[dev]"
ruff check .
mypy src/mao
pytest -q
python -m mao.evals.run
python -m mao.task --task "Compare hybrid vs dense retrieval in one paragraph"
python -m uvicorn mao.main:app --port 8000

The browser console is then available at http://127.0.0.1:8000/.