For expression and solver data flow, see the architecture guide.

Status distinguishes modeling infrastructure from tested end-to-end solves. In-progress code remains on main; roadmap items have no delivery commitment.

Current support

Problem family	Status
LP	Supported and tested
Convex QP	Supported through available conic reformulations
Convex QCQP	Supported for available SOC/rotated-SOC formulations, including ellipsoidal and squared-inequality constraints
SOCP, SDP	Supported and tested
Exponential/power cones	Partial modeling and canonicalization; asymmetric-cone solver support in progress
Mixed-integer programming (MIP)	Planned; integer/binary modeling and solving are not implemented

QCQP means quadratically constrained quadratic programming. Support is limited to convex formulations expressible through the existing APIs, not arbitrary nonconvex quadratics.

Atom support boundary

Atoms	Required cone	Solve status
`Exp`, `Log`, `Softplus`, `LogSumExp`, `KLDivergence`	Exponential	In progress
`GeoMean`	Power	In progress

Evaluation/canonicalization tests do not establish solver support. Both asymmetric cones still have placeholder step-length routines and incomplete solver operations. GeoMean currently emits a power cone, not an alternative SOC reformulation.

Remaining work

QCQP: expand tests combining convex quadratic objectives with multiple quadratic constraints, checking canonicalization, feasibility, and optima.
Exponential/power cones: complete cone operations and a compatible asymmetric-cone algorithm; audit atom formulations and add end-to-end solves.
Mixed integer: define integer/binary domains, discrete search, bounds/gap reporting, and verification against known integer optima.
Capability diagnostics: reject unsupported cone/domain combinations early with actionable errors. Current asymmetric solves may fail in template instantiation.

Promote support only after modeling, canonicalization, solver behavior, and result validation are tested together. Keep API status notes linked to this page.