What’s upcoming · no promised dates

Alpha state

Current Axiom alpha status, the work blocking qualification, and the engineering phases ahead.

Alpha, not production-ready. The latest hosted Linux resource gate fails both paper modes. Multilevel runs out of memory during million-vertex startup under a 512 MiB address-space cap. Basic completes one million updates and a backup, but runs out of memory during disk-pressure recovery. The 10k durable-updates/s target and billion-vertex capacity are not established.

What the team is working toward

First make the current two-mode product bounded and recoverable at its declared scale. Then broaden adversarial qualification and complete the durable user-facing contracts. No calendar dates are attached: a phase advances only when its evidence passes.

PHASE 01–04

Make the engines fit and recover

Reduce Basic/Multilevel memory peaks, eliminate avoidable graph-sized work, and pass constrained startup, update, backup, pressure, and replay tests before claiming scale.

Active · blocking
PHASE 05–07

Strengthen the durable user contract

Qualify external-ID recovery and lifecycle rules; stress atomic batches and version-coherent query snapshots under load, failure, and bounded-resource conditions.

Partially implemented
PHASE 08–12

Build application-facing capabilities

Add bipartite constraints, bounded versioned events, stable observability, deterministic replay/history, and factual matching explanations. Introduce transport only after local semantics are qualified.

Upcoming · API design first
PHASE 13

Gate releases on evidence

Verify reproducible packages and independent research, durability, performance, and deployment gates for each matcher mode. Keep weighted/maximum/capacity matching as separate future engines.

Release gate · open

Qualification before expansion

  • Pass the 512 MiB / 192 MiB hosted resource envelope for both Basic and Multilevel, including recovery—not only update throughput.
  • Repeat uniform, skewed, burst, overload, rollback, and restart workloads; retain latency, memory, storage, and deterministic-state evidence.
  • Publish each mode’s actual limits. Do not transfer removed native-engine benchmarks or guarantees to the paper modes.

Hardware power-loss qualification is deferred for this version. Maximum-cardinality, weighted, preference, and capacity matching are separate future-engine questions, not promises in this alpha.

Follow the detailed 13-phase engineering plan, the product roadmap, and the current qualification evidence.

NextCurrent evidence