Limitations and support¶
Unlisted is unsupported. The authoritative envelope is the repository's support matrix, generated from the registered envelope with the evidence for every row (a test, a manifest check or a document). This page states its substance in plain words.
Each code below links to its explanation in Messages, and each limitation row of the support matrix has its own section in Limitations.
Supported¶
- Property bases: the five of Property bases, each within its domain.
- Units: the models of the unit catalogue, each on the bases listed there.
- Topologies: acyclic flowsheets; recycles (single, nested, high-gain) by tearing with safeguarded Anderson acceleration and restart, or merged into one equation-oriented region; cross-unit specifications; exact zero flows. More than one loop or region per plan is not exercised by a registered case.
- Solve policies:
T06-revision-v2(the default for revision solves),T04-W12, and by explicit idT08-warm-v1,RRC-revision-v1(Revision continuation) and the experimentalT08-ptc-v1. - Interfaces: Python, CLI, HTTP and MCP, as in Interfaces.
- Platforms: Linux x86-64 and aarch64, Python 3.13, with the pinned dependency set.
Not supported, and how each is refused¶
| Request | What happens |
|---|---|
| A property method other than the selectable bases | not representable: a revision has no property-method member; another provider's components are components_unsupported |
| A model not in the registry, or on a basis it does not bind on | model_unsupported(<id>) at binding, with the bases it does support |
| A state outside a basis's domain | INITIALIZATION_FAILED, or invalid_trial_state inside a solve; never extrapolated |
Optimisation (validate with task: optimization, the optimize job) without the audited nlp environment, or off Linux x86-64 |
unsupported (nlp_solver_unavailable); the library returns an UNSUPPORTED report |
Optimisation of a coupled (torn) flowsheet or a legacy_eo document |
UNSUPPORTED (OPTIMIZATION_ROUTE_UNAVAILABLE) |
A change with task: optimization (commit_change, preview_change) |
task_unsupported(optimization) |
| Dynamics or time-varying specifications | not representable; a time member is a schema refusal |
| Gauge pressure, unregistered unit spellings | refused at validation |
| Second derivatives through the property callback | reported absent; never zeros |
| A specification pairing other than one freed column per promoted target | UNSUPPORTED_RANK_STRUCTURE at plan construction |
| Container or installer artifacts, container replay | not provided |
Known limitations worth knowing first¶
- A
VERIFIEDcertificate certifies residuals at registered tolerances; the solution-error bound is recorded, not promised. Energy checks share the property provider with the solver. - Bitwise reproduction is reported, never promised across platforms.
- Pseudo-transient continuation (
T08-ptc-v1) is experimental. - Pure-fluid stream regimes are declared; there is no regime discovery, and a closure in another
regime is refused
phase_regime_mismatch. A dormant single-phase thermal outlet certifiesUNVERIFIED(rank-deficient). - The column: see Distillation column.
- Cancellation is not cooperative inside the verifier.
- Optimisation holds phase and assessment regimes per start and verifies local stationarity only; its boundary, zero-flow and robustness limits are in Optimising a revision.
- Agent evidence: one pinned model and configuration, over MCP only, not out of sample.
Verification, validation, optimality¶
| Status | |
|---|---|
| Numerical verification | established against independent references (for example the SYN-001 oracle against 20-digit reference values) and by blind comparison with IDAES, DWSIM and BioSTEAM on shared cases — which share the equations, so this is verification |
| Empirical validation | the Wilson basis against measured vapour–liquid equilibrium (within its bands; methanol's liquid enthalpy and heat capacity not within band); IAPWS-95 and CoolProp carry their own published validation; the ammonia loop, its reactor and every SYN-001 result are not empirically validated |
| Optimality | not established |
| Human review | numerical and process-modelling review by a person is pending in every evidence manifest |
What each work package has and has not shown is recorded in its evidence manifest
(evidence/).