17 requirements · evidence first

Product roadmap

A requirement-by-requirement view of what exists, what is only partially qualified, and what still needs engineering.

Current boundary (October 4, 2026): Axiom exposes only paper-derived basic and multilevel matching through one Durable/Service path. Packed is graph storage, not a third matcher. Hosted Linux resource qualification currently fails: Multilevel runs out of memory during startup; Basic completes one million updates at ~9,054/s but fails during disk-pressure recovery replay. Durable external IDs are supported for the fixed vertex universe; dynamic vertex lifecycle is not implemented. These are not passing gates. The 10k/s target remains unproven. Hardware power-loss is deferred; billion-vertex support is unproven.

“Partial” means only the evidence stated exists; it does not imply the full requirement is met. Historical native-engine results do not qualify Basic or Multilevel.

#StatusRequirementEvidence and remaining work
1PartialPaper-engine performance hardeningRecursive deepcopy is removed; recent work indexes I3 crossings, H-tilde sources, and Extend projections. Remaining graph-sized snapshots, hierarchy work, and aggregate memory budgets still need migration and qualification.
2In qualificationDurable basic and multilevel modesDurable and Service select one of the two persisted paper modes. The former native matching engine is removed; exact recovery, failure-path and per-mode performance qualification remain.
3PartialRepeatability, skew and adversarial qualificationEarlier power-law and 60-second rates are historical results from the removed native matcher and do not qualify Basic or Multilevel. Repeatable paper-mode uniform/skewed, adversarial, hierarchy-boundary and recovery matrices remain.
4PartialService and deployment qualificationPaper-backed local recovery, bounded-resource tests and operational guidance exist. Deployment supervision, broader skew/burst and explicit quota/alert policy remain. Hardware power-loss testing is deferred.
5PartialDurable external vertex IDsTyped string, UUID, and signed-64-bit integer mappings persist in SQLite and are never reused. Installed-package, corruption, sustained lookup, and lifecycle qualification remain.
6Not doneDynamic vertex lifecycleThe vertex universe is fixed at construction; durable add/remove semantics are absent.
7Implemented; qualification activeAtomic batch updatesDurable Service has bounded atomic submit_batch with one SQLite commit boundary, exact retry and rollback tests. Installed-release and workload qualification remain.
8Implemented, boundedAtomic multi-query read snapshotsService.read_snapshot returns immutable partner and edge-membership answers under one committed version with a 4096-query cap, stale-version rejection and serialized owner execution. This is not long-lived MVCC or historical topology retention.
9Not doneFirst-class bipartite APINo typed partitions or same-side edge rejection contract exists.
10Not doneEvent subscriptionsVersioned change delivery, bounded subscriber queues and backpressure semantics are absent.
11PartialObservabilityService admission/group counters and benchmark histograms exist. Stable metrics/exporter, storage/recovery health and alert contracts remain.
12PartialDeterministic replay and historical debuggingDurable/Service history exports bounded retained operations with versions and chained digests; recovery replays and verifies state. Public import/replay verification and historical graph queries remain.
13Not doneMatch explanationsThere is no version-linked explanation of partner changes or algorithm repair evidence.
14PartialCross-platform packagingCI/release matrices define 20 CPython wheels across five platform targets; hosted evidence is pending. Windows verifies native/paper APIs only because POSIX Durable/Service are unsupported there. Windows ARM64 and musllinux are excluded.
15Not doneNetwork transportService is local and threaded; authenticated HTTP/gRPC transport and end-to-end network semantics are absent.
16Future scopeWeighted and constrained matchingWeighted, maximum-cardinality and capacity-constrained matching remain separate future engine contracts.
17PartialMachine-enforced release gatesResearch, quality and packaging gates exist. Paper-mode durable qualification and complete mode-specific release gates do not.

Implementation-level decisions and benchmark artifacts are maintained alongside this roadmap in the repository. A claim is complete only when its stated gate passes on the relevant engine and packaging path.

NextRead research status