Know what is stable before you depend on it
Review product version, capability status, open release gates, dependencies, evidence sources, known limits, and the exact boundary between local checks and external promotion evidence.
Capability and Release Center
This center is projected from the checked-in version manifest and capability stability registry. It does not infer a release from a passing local build.
Promotion reports are review inputs only. They do not mutate the manifest, approve a release, create missing G3/G11 evidence, or promote experimental work directly to stable.
- Language version
- 0.2.0
- Capabilities
- 58
- Stable
- 2
- Preview
- 35
- Experimental
- 21
- Open gates
- 135
Product preview readiness
Requirements passedThe overall preview label is valid only while all four product-spine requirements remain satisfied.
- Stable parser and statevector core3 checked-in evidence references
- Usable workspace product spine4 checked-in evidence references
- Version-aligned tooling product spine3 checked-in evidence references
- Transparent stable, preview, and experimental separation3 checked-in evidence references
Release history
Manifest history and the latest per-capability review date (2026-09-24) are shown separately from CI or operational promotion evidence.
v0.2.02026-09-26PreviewPreviewCoordinated 0.2.0 preview for the expanded language, numerical library, training workflows and toolchain. New sessions use the default language profile; saved feature choices and supported 0.1.0 artifacts remain readable. Capability statuses and artifact format versions are unchanged.
v0.1.02026-07-15ExperimentalPreviewPromoted the overall N/M product label to preview after the stable parser/statevector core, usable workspace spine, version-aligned tooling, and gate-transparent capability matrix were verified; experimental N/M 0.2 capabilities remain explicit opt-ins.
Strict G3/G11 evidence evaluator
The bounded evidence reader and recomputation tools are downloaded only when you choose to inspect an exact candidate.
Capability contracts and promotion gates
Expand any capability to inspect its contract version, owner, surfaces, dependencies, limitations, G1-G12 gate states, and evidence references.
Showing 58 of 58 capabilities.
Exact-candidate G3 evidence
The automated matrix stays open until 5 frozen suites produce all 12 Node shards, 4 Python shards, plus 1 production Chromium shard in one workflow run and attempt for one repository and commit. It accounts for 53 capabilities without converting a local run into remote evidence.
Separate hardware/provider evidence remains explicit: hardwareExecution.
Trust boundary: The JSON is not a cryptographic signature and does not replace retained GitHub run artifacts or real hardware/provider operations evidence.
The machine-readable promotion evaluator applies preview-to-stable-only, checks dependency closure, and only proposes candidates. It never edits the manifest or approves a release.
Reviewed production G11 evidence
The compact operational contract binds 6 reviewed production suites to one repository and candidate SHA, covering 7 capabilities without retaining raw events in the readiness report.
Capabilities without sufficient shared operational coverage remain on explicit evidence tracks: hardwareExecution.
Trust boundary: The JSON is not a signature, anonymous event counts do not prove distinct users, it does not replace each retained privacy-reviewed source export, and it cannot cover hardware without its separate provider evidence track.
qaoaRuntime9/10 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- qml
Why this is not more stable
- Contract 0.1 is a local MaxCut simulator bounded to five nodes, ten ordered edges, three layers, and 100 GD/SPSA steps; it does not establish large-graph optimality or hardware performance.
- The local maximum-workload performance ratchet is complete and configured on Windows/Linux Node 20/22; remote conformance evidence for the current candidate remains open.
Promotion gates
Dependencies
trainBlock
Evidence
12 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15
qngOptimizer9/10 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- qml
Why this is not more stable
- Contract 0.1 uses a bounded forward-state fidelity diagonal approximation; full, block-diagonal, noisy, sampled, and hardware-derived quantum geometric tensors are outside this preview.
- QNG is limited to eight parameters and 100 local steps. The local maximum-load ratchet is complete; the remote Windows/Linux Node 20/22 candidate matrix must pass before stable promotion.
Promotion gates
Dependencies
trainBlock
Evidence
12 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15
trainingStudio9/10 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- qml
Why this is not more stable
- Training remains local, bounded to 100 steps and eight parameters, and has no checkpoint recovery or early stopping.
- The local maximum-profile performance ratchet is complete and configured on Windows/Linux Node 20/22; remote conformance evidence for the current candidate remains open.
Promotion gates
Dependencies
trainBlock, qmlLosses
Evidence
13 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15
trainabilityDiagnostics9/10 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- qml
Why this is not more stable
- Gradient-variance thresholds are heuristic and must not be interpreted as proof of trainability or quantum advantage.
- The local maximum-profile performance ratchet is complete and configured on Windows/Linux Node 20/22; remote conformance evidence for the current candidate remains open.
Promotion gates
Dependencies
trainBlock
Evidence
12 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15
modelPackages10/11 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- qml
Why this is not more stable
- FNV-1a integrity detects accidental/tampered content but is not an author signature; signed release artifacts remain outside contract 0.1.
- Contract 0.1 accepts the current language major/minor and preserved 0.1.0 model artifacts, the statevector engine, fourteen qubits, and the explicit noise registry; new engines require a new runtime compatibility contract.
- The Windows/Linux Node 20/22 conformance and maximum-carrier workflow is configured, but remote matrix evidence remains open.
Promotion gates
Dependencies
trainBlock, learningPackages
Evidence
15 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15
qmlCopilot9/11 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- qml
Why this is not more stable
- Contract 0.1 is limited to five deterministic dataset intents, bounded local templates, and finite smoke-training evidence; it does not claim model accuracy, convergence, quantum advantage, or arbitrary prompt support.
- The 20 EN + 20 TR golden set, maximum-boundary local ratchet, QML-mode production browser flow, and mode-specific strict operations evaluator are complete. Current-candidate remote platform parity and a reviewed real 7-8 day window with 20 live-provider attempts, at least 90% externally verified outcomes, bounded p95 latency and daily cost, plus correlated feedback remain open.
Promotion gates
Dependencies
verifiedCopilot, trainBlock
Evidence
20 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15
pennylaneExport7/8 stability gates passedPreview
- Contract
- 0.2.0
- Owner
- tooling
Why this is not more stable
- The 0.2 executable support contract is pinned to PennyLane 0.45.1, Python 3.11-3.14, the analytic default.qubit device, generated Bell probabilities, bounded Hamiltonian training, and frozen embedded-dataset supervised training; plugins, hardware devices, sampled/noisy execution, external datasets, and other PennyLane versions are not certified.
- The local Python 3.11-3.14 downstream matrix is green. The current candidate still requires all four remote Python jobs before cross-platform parity is evidence-backed.
- The local G6 ratchet fills the 200,000-byte public export carrier with the frozen five-qubit, eight-parameter QML limits and 8,100 deterministic gates; it measures N/M-to-Python generation, not Python environment installation or device execution.
Promotion gates
Dependencies
parser
Evidence
9 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-26
mcpServer9/10 stability gates passedPreview
- Contract
- 0.2.0
- Owner
- tooling
Why this is not more stable
- Contract 0.2 is limited to newline-delimited stdio and seven local tools, including the read-only shared capability catalog; prompts, resources, sampling, remote transports, arbitrary filesystem/network access, and third-party client certification are excluded.
- The local Core/standalone parity and maximum-carrier ratchet are complete and configured on Windows/Linux Node 20/22, but remote matrix evidence for the current candidate remains open.
Promotion gates
Dependencies
parser, cli
Evidence
11 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-26
sourceDebugger10/12 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- tooling
Why this is not more stable
- Contract 0.1 records a bounded local timeline and the CLI is non-interactive; durable timeline persistence and remote debugging remain out of scope.
- The strict privacy-minimized G11 evaluator is complete, but exact-candidate remote matrix evidence and a reviewed real 14-day window with 100 sessions plus 100 navigation operations remain open.
Promotion gates
Dependencies
statevectorRuntime, stabilizerRuntime
Evidence
19 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15
experimentRecords10/11 stability gates passedPreview
- Contract
- 0.1.0
- Owner
- qml
Why this is not more stable
- Source is omitted by default but hashes remain pseudonymous identifiers; records are local artifacts, not anonymous telemetry.
- Comparison identities intentionally exclude runtime metadata in contract 0.1; callers comparing cross-runtime results must inspect runtime fields, which remain covered by whole-record integrity.
- The Windows/Linux Node 20/22 conformance and maximum-carrier workflow is configured, but remote matrix evidence remains open.
Promotion gates
Dependencies
trainBlock
Evidence
14 source references. Open the capability contract for the complete list.
Last reviewed: 2026-07-15