The operations platform engineered for a workforce of humans and AI.

The hardest problem in mixed human and AI operations is not intelligence. It is two writers, one file.

• Task locks, dispatcher claimsLIVE
  Every actor claims a task through an atomic, lease-based dispatcher before working it. Enforced by database triggers, not convention. No collisions, a crashed worker's lease expires, nothing deadlocks.
• File locks, resource locksLIVE
  Database-backed locks per file path: acquire, renew, release, FIFO fairness, TTL expiry, holder identity, task attribution, audit history. A competing writer gets a refusal naming the holder and their task.
• Hash-anchored writesLIVE
  A lock only binds writers who ask for it, so every write carries the hash it was based on and a stale write is refused, not absorbed. Every mutation journaled, attributed, backed up. Even an out-of-band manual edit is detected and snapshotted.
• Collaborative SOP editingLIVE
  Documents edit one section at a time with atomic splice semantics. Two humans, two AIs, or one of each can work the same SOP and cannot destroy each other's work. Proven against a twelve-scenario adversarial matrix, all green.

• PostgreSQL is the single source of truth. Every write transactional, on the most trusted open-source database in the world.
• Write queue and dead-letter queue. What cannot land immediately is durably staged, never dropped. Exhausted retries preserve payload, actor, history and error class.
• Engine-independent drain retries every minute and pushes a phone alert the moment anything hits the dead-letter queue.
• Backups in depth. Hourly snapshots, daily tested restores, a backup before every shared-document change. Rollback is routine, not a crisis.
• Error-handling standard, zero exceptions. No code ships without boundary handling, timeouts, structured logs and loud failures, checked by daily automated audit.

Rules, SOPs and category context live in a queryable database, delivered live over authenticated APIs and a connector at the moment of work. No mounts, no VPN, no shared drives. A new user on a phone gets the same complete context as a veteran at a workstation, and a debug view shows exactly which rules applied to any session, on demand. Stale knowledge is the root cause of most AI failure, and APOP attacks it at the source.

Every rule and SOP resolves to one authoritative record, with no copies drifting apart in wikis and inboxes. New guidance that conflicts with what is binding does not overwrite and does not vanish: it surfaces as an owner-approved proposal, and nothing changes without a decision on the record. A human and an AI read the same rules and reach the same answer.

• One release-managed codebase, and each tenant gets its own server, database, filesystem subtree and environment. Crash isolation and data isolation come free.
• No tenant hardcoding, ever. Tenant specificity lives in config, flags and context. That is what keeps APOP a product, not a pile of custom deployments.
• Promotion pipeline. Dev, QA, scripted promote with regression pre-flight, health-gated restart, one-command rollback. A recent production cutover: regression green before and after, seconds of downtime.
• Feature flexibility, four channels. Context programming with no code, per-tenant config and connectors, sponsored default-off flags, and a premium fork for true divergence.
• Predictable cost. Shared codebase, per-tenant isolation, SaaS tiers. Optional features arrive as configuration, not change orders.

The engine claims work through the dispatcher, loads full category context, reasons with a frontier-class model, and routes output through guards that err safe. It never works unclassified tasks, never claims a human's personal tasks, and never self-modifies. It proposes, humans approve. Self-healing was tried, failed instructively, and is permanently banned. All AI work passes two-stage QA.

Start conservative, graduate the trusted. Autonomy rises as confidence is earned, per person and per category. Adoption is a controlled ramp with an evidence trail behind every step. The ramp has three stages, and the rules never change between them.