For contributors — how xlog’s release state and architecture priorities are organized internally. This is a roadmap, not a user guide.
This roadmap describes implemented capabilities and forward-looking architecture priorities. It does not duplicate a mutable package version. Use CHANGELOG.md and the selected artifact’s release notes to determine packaged availability.

​
Released capabilities

Tagged artifacts include the shared Rust workspace, CUDA-backed deterministic execution, probabilistic and epistemic surfaces, solver services, Python packaging, and the xlog CLI. Aggregate-fused worst-case-optimal joins (WCOJ), GPU Free Join, factorized recursive deltas, joint neural-symbolic mixtures, and grouped training are available in tagged artifacts beginning with 0.10.0. Two versions have followed. 0.11.0 added a device-resident joint-constraint solver — a GPU handle that picks one label per entity under pairwise constraints and hands its result buffers back without a host copy — plus a memoized dynamic-programming solve stage for chain components too large to enumerate. 0.12.0 added exact log-evidence and the CNF-variable-to-fact map on probabilistic results (EvalResult.log_z_e, CompiledProgram.prob_var_map), native n-ary relation provenance, batched same-head rule unions, and imported-module pragma warnings. The public Rust release boundary is the set of publishable crates listed in Release Process. pyxlog is shipped through Python packaging, not through crates.io.

​
Stable Architecture Commitments

XLOG’s architecture is organized around a fixed set of commitments. These commitments — not old milestone labels or implementation-board names — define the architecture:
  • One frontend for deterministic, probabilistic, epistemic, and neural-symbolic programs.
  • An explicit boundary between host-side control logic and the GPU data plane.
  • Inspectable reasoning contracts carried by three internal intermediate representations: RIR for deterministic relational execution, PIR for probabilistic provenance, and EIR for epistemic semantics. SAT/MaxSAT and verification use the separate shared GpuCnf representation and CDCL solver service.
  • CUDA kernels that operate over device-resident buffers (data that stays on the GPU).
  • Fail-closed behavior: when a proof, route, or capacity contract is not met, execution stops rather than returning an unchecked result.
  • Route counters and telemetry: dispatch counts that record which optimized execution path actually fired.

​
Available since 0.12.0

These public contracts are part of 0.12.0:
  • Native relation evidence (breaking change) — ordered semantic roles and provenance records bind to complete facts of any positive arity and follow every relation mutation atomically. This replaced the Python sidecar API rather than extending it: evidence() no longer returns source_path, source_hash, row_hashes, accepted_count, rejected_count, output_hash, or decision_counts; evidence() and relation() raise KeyError for an unknown relation instead of returning an empty dict; and RelationEvidence is an immutable native class with no RelationEvidence(session, name) constructor.
  • Exact relation manifests — versioned metadata and DLPack columns reconstruct persistent rows, roles, and whole-fact evidence in another compatible session.
  • Owned persistent relation snapshots — persistent DLPack imports are copied device-to-device into session-owned storage, while transient evaluation imports remain zero-copy.
  • Deterministic delta behavior — canceled updates do not advance relation versions or emit callbacks, and invalid batch keys are rejected before mutation.
  • Batched multiway unions — same-head rule outputs merge through one chunked multiway union per evaluation pass, with the byte budget controlled by XLOG_UNION_CHUNK_BYTES.
  • Ignored-pragma warnings — an imported-module pragma produces warning[W0510] on stderr because pragmas apply only in the entry file.

​
Architecture Priorities

The next architecture work stays focused on runtime facts that users can observe:
  • clearer route telemetry for worst-case-optimal join (WCOJ), Free Join, aggregate fusion, and recursive delta routing;
  • stronger release evidence for CUDA-required behavior;
  • explicit no-host-transfer counters wherever a path claims device-resident data flow (data that never leaves the GPU);
  • sharper failure diagnostics for unsupported probabilistic, epistemic, and solver shapes;
  • documentation examples that show how to verify route dispatch directly, instead of inferring it from the final answers.

​
Out-Of-Scope For This Roadmap

This page is a release-state view, not the historical task board. Internal phase labels, board codes, and dated task bundles are not part of this roadmap. Where a historical item still matters, it is documented as shipped behavior, the command that verifies it, and the release boundary — not as a task entry.