The Jaxstro research-software assessment scorecard is the living companion to this ranked investment assessment. Here, state of the art means unusually strong alignment among scientific scope, numerical contracts, JAX behavior, independent evidence, provenance, discoverability, and maintenance cost, not the largest feature list.
Delivered strengths¶
| Dimension | Maturity | Evidence-backed assessment |
|---|---|---|
| Scientific breadth | validated | Units, constants, coordinates, geometry, numerics, spatial operations, spectra, atmospheres, parameter bridges, and provenance cover recurring foundation needs. |
| Ownership discipline | ratified | Architecture and decisions keep generic mechanics here while physical interpretation and scientific acceptance remain downstream. |
| Numerical robustness | validated | Rootfinding, interpolation, quadrature, linear algebra, distributions, meshes, and special functions expose boundary and failure contracts in Validation. |
| Conditioning | validated | Denominator, slope, covariance, positive-definite, and removable-limit checks are explicit. |
| AD honesty | ratified | Smooth, zero, blocked, surrogate, validation-only, value-first, and certified implicit paths remain distinct. |
| JAX transform coverage | validated | Selected jit, vmap, lax.map, grad, JVP/VJP, scan, and PyTree behavior is tested with named exceptions. |
| Dimensional safety | implemented | jaxstro.units is canonical; jaxstro.quantity is implemented, and one- and multidimensional quadrature have an alpha opt-in adapter with heterogeneous coordinate axes without claiming ecosystem adoption. |
| API cohesion | validated | Grouped owner pages, export checks, typed results, and focused tests expose stable contracts. |
| Serialization | implemented | Quantity and provenance records have explicit round trips, but no public root-result replay serializer is validated. |
| Performance and compilation evidence | validated | Rootfinding and spectra use units-explicit evidence records. One-dimensional quadrature adds truth-gated matched-library labels and two-suite optimization acceptance. Phase B adds calibrated comparator labels and an immutable host baseline; its analytic dimension-16 memory trigger does not establish an observed bottleneck, and the measured memory campaign remains pending. |
| Evidence freshness | validated | Computational artifacts, scientific policy, and source-backed provenance remain distinct and freshness checked. |
| Provenance | validated | Runtime manifests and source-backed cards answer different questions while sharing deterministic identity and digest mechanics. |
| Research workflow coverage | implemented | Three executable investigations connect public APIs to contracts, indexed evidence, limitations, and warranted claims. |
| Limitation coverage | implemented | Workflow and contract records preserve explicit non-claims, but callable-level coverage remains uneven. |
| Accessibility | implemented | The docs gate checks alt text and rendered-route integrity; broader user evidence remains future work. |
| Discoverability | implemented | Semantic start, methods, representations, workflows, API, validation, and project routes answer distinct questions. |
| Downstream reuse | implemented | Known sibling use motivates shared primitives, but pinned adoption and compatibility evidence is not yet generated. |
High-confidence gaps¶
Evidence depth is uneven across modules.
Transform support and batching cost remain distributed across prose and tests.
Quantity adoption awaits downstream parity, serialization, performance, ergonomics, and migration-cost evidence.
Research workflow coverage is narrow relative to the public callable surface.
Downstream reuse lacks a generated symbol-to-project compatibility map.
Runtime priority¶
Future runtime ownership follows this exact order:
The active exception is to complete the approved jaxstro.quad program before this queue.
Phase A and the Phase B runtime families are implemented: finite
hyperrectangles, tensor and Genz-Malik integration, Smolyak sparse grids,
deterministic and randomized Sobol methods, first-order accepted-formula
replay, heterogeneous alpha quantity axes, and family-specific evidence.
The exhaustive release, strict rendered-site, clean-wheel, consolidated review,
and observed-memory gates remain open. Phase C geometries and downstream
adoption remain separate work.
jaxstro.mljaxstro.uncertaintyjaxstro.signalconsumer-driven ecosystem adapters
fields only after two consumers
Lineax owns general iterative linear solving, Optimistix owns general nonlinear systems, and Diffrax owns adaptive differential equations. Jaxstro owns its approved adaptive quadrature program. Quadax remains an independent comparison implementation for validation and matched benchmarks. Jaxstro may add a narrow consumer-driven adapter only when it adds a scientific contract or evidence layer absent from the delegated owner.
Now¶
1. Expand contract and evidence coverage¶
Impact. Researchers can discover supported transforms, limitations, and evidence per consequential public callable.
Evidence gate. Every claimed cell links to a test and limitation; unverified cells remain explicitly unclaimed.
2. Expand executable research workflows¶
Impact. More public APIs are exercised from question through computation, audit, and warranted claim.
Evidence gate. Coverage is generated from the schema-2 workflow registry and distinguishes contract links, indexed evidence, and limitation links.
3. Decide quantity adoption¶
Impact. Resolve dual-surface maintenance without destabilizing downstream science.
Evidence gate. Pinned downstream parity, serialization, performance, ergonomics, and migration-cost results precede the decision.
4. Generate downstream adoption evidence¶
Impact. Reuse and adapter claims become observable.
Evidence gate. Records derive from pinned imports, compatibility tests, and last-verified revisions.
Next¶
1. Complete the jaxstro.quad Phase B release gate¶
Impact. The implemented multidimensional families become release-ready without overstating memory, performance, geometry, quantity-adoption, or statistical guarantees.
Evidence gate. The exhaustive source/test/docs/clean-wheel gate, strict rendered-site audit, one consolidated checkpoint review, and observed process/device-memory campaign pass from clean committed source.
2. Build the minimal jaxstro.ml evidence-complete slice¶
Impact. Shared preprocessing, data plans, and fixed-step training become auditable without duplicating Equinox or Optax.
Evidence gate. Deterministic restart, leakage, mask, key, and manifest checks pass on a bounded reference workflow.
3. Add jaxstro.uncertainty¶
Impact. Scientific maps gain reusable propagation mechanics without claiming posterior inference.
Evidence gate. Analytic linear cases and documented nonlinear failure boundaries support each method.
4. Add jaxstro.signal¶
Impact. Sample-axis and spectrum conventions become explicit and reusable.
Evidence gate. Normalization, units, windows, phase, and delay identities are independently checked.
Later¶
1. Add consumer-driven ecosystem adapters¶
Impact. A real consumer can gain units, telemetry, provenance, or evidence without forking the delegated solver.
Evidence gate. The consumer documents the missing contract and verifies the adapter against Lineax, Optimistix, Quadax, or Diffrax behavior.
2. Revisit multidimensional fields¶
Impact. Shared topology and conservation primitives could serve multiple simulation domains.
Evidence gate. Two consumers first agree on topology, staggering, boundary, conservation, capacity, and JAX shape contracts.
Evidence required¶
Accept a future SOTA claim only when the relevant layer includes a contract, a typed boundary, supported transforms and cost semantics, independent validation, reproducible artifacts where metrics matter, explicit limitations, and downstream adoption evidence when reuse justifies ownership.