The limits of the language

A model can only say what the language has words for. This page says which words can be added, and which cannot. Read it before you ask for a new operator, block or keyword. For the rules a model itself has to obey, read the ten rules.

How a new construct enters

A request for something new is one of three kinds, and the kind decides what it costs to add.

• A macro is a template with arguments, written in the file under macros:, that is substituted into an expression before anything reads it. Adding one costs nothing: it uses only operators that exist, so no engine has to change. Most requests turn out to be a macro. See macros.
• A primitive is an operator built into the language: sum, sum_back, at, shift, and the where comparisons. A file cannot add one. Adding one here is the expensive kind: every engine that builds models has to implement it, and the typesetter has to print it in LaTeX, Typst and Markdown.
• A formulation is a block that expands into ordinary variables and constraints before the model is built. piecewise: is the only one. It costs as much as a primitive to build, but composes as freely as a macro, because the rest of the model only sees the variables and constraints it emitted.

A request that is none of the three is refused, and the table of refusals records it with what to write instead.

What a new primitive has to satisfy

A macro must be able to call it. Everything a modeller might pass in goes in the value of a keyword argument, such as over=snapshot, never in the key. A macro can write over=d and let the caller supply d. It could not do that if the dimension were the keyword itself.

An operator may read the whole table. It pays one full pass over the data. sum(p, over=g) reads one row per generator. shift(p, along=t, offset=1) reads one row, the one before it. x * y * a reads the rows of a that pair an x with a y. Each reads a bounded number of rows per output row, so an engine builds the model one chunk of rows at a time.

One kind of operator reads every row to produce one row. The engine then reads the whole table before it builds any chunk, and the chunks stop being independent. That is the price, and a request for such an operator names it. Reading only the coordinate labels costs nothing. "The last snapshot" looks at the list of snapshots, not at the data.

An operator that calls itself is refused. Nothing bounds how far it expands, so no number of passes over the data is enough.

The operator	Allowed?
filters rows on a column they already carry	yes
joins each row against a parameter or a relation	yes
reads a fixed number of neighbouring rows	yes
reads only the coordinate labels	yes
reads every row	yes, at one full pass before any chunk builds
calls itself	no, and the message names what to write instead

Degree is not a third test. p * q at one coordinate is a join of a table with itself, so the objective and the constraints take it. Two things limit the quadratic case:

• Where it stands. More solvers and file formats take a quadratic objective than a quadratic constraint. Which ones is the separate question below.
• A product of two sums. sum(x, over=i) * sum(y, over=j) multiplies every term of the first sum by every term of the second, and the file does not say how many terms either sum has. It is refused. x[i] * y[j] * a[i, j] is allowed, because the table a says which pairs exist.

A new primitive is finished when to_program lowers it, the typesetter prints it in all three formats, and an engine's build of a model that uses it matches the same model written out by hand.

Three kinds of refusal

The language refuses it because…	Examples	Can it change?
one solver cannot take it	indicator constraints (#220); a quadratic constraint. sos: was in this group, and entered: a solver with sets takes it as one, and a solver without gets binaries	yes, solver by solver
the file would stop being the artifact	arbitrary Python, whose content no loader can check and no typesetter can print	no
this project puts the work elsewhere	data preparation such as resampling; helpers for one domain; Python that decides which declarations exist	it could; this project does not want it to

Three things never appear inside one model: an if, a loop, and a set of declarations that depends on the data. dims: [snapshot] does not know how many snapshots there are, and it does not need to. A dimension computed before the model loads is fine: a cycle basis for Kirchhoff's voltage law is a graph algorithm run in data preparation, and its result arrives as a parameter. What no model can hold is work that needs the solver's answer before it can write the next row, such as cuts added during a solve. A tool can still loop over models: a rolling horizon, Benders decomposition and successive substitution each build a model, solve it, and build the next (Track 2).

Solver capability

The test above asks whether an engine can build the operator. Whether a given solver then accepts the result is a second question, and the language does not answer it. If it did, one solver's limits would be written into the language, and every other solver would inherit them.

Two examples show why the questions stay apart:

• HiGHS has no special-ordered sets. Gurobi does. An engine handing a model to HiGHS rewrites each set as binaries and big-M rows; one handing it to Gurobi passes the set through.
• A quadratic constraint is accepted by some solvers only when it is convex, and convexity depends on the numbers, which the file does not have.

So sos: entered the language on the first question alone: a set only names columns a variable has already declared. Each engine then decides how to hand it to its solver. A set rewritten as binaries returns no duals, where the native set does, so the engine reports which it did, and you choose the solver (#925, #928).

A set is a declaration under sos:, not an expression:, because no expression can write it: x_i * x_j == 0 for every pair is degree 2, and "at most two of these are non-zero" is not arithmetic at all.

What counts as data preparation

The table below refuses data preparation as a language feature. But from inside a model, a column you computed in pandas and a column the language could have derived look the same: a parameter arrives, and a constraint reads it. One sentence tells them apart:

Data preparation computes what the model cannot know. The language derives what it can from data the model already has.

A cycle basis is the first kind. It needs the network's topology, which only the data has, so cycle_incidence arrives as a parameter. A minimum up time is the second kind. min_up_time is a column the model already binds, and the window "the last min_up_time hours" follows from it, so sum_back(window=min_up_time) reads the width off the column and you ship no window mask (#849).

Calliope made this trade-off the other way. A Calliope component carries several where-guarded versions of one equation, so a cyclic and a non-cyclic store are one block. Here they are two constraints, and which applies is a column in the data (#711). What this buys is that the list of declarations is fixed before any data is read, which is what lets an engine build the model one chunk of rows at a time.

Deliberate non-primitives

What has been asked for and refused, with the reason and what to write instead. That another tool has a feature is not by itself a reason to add it.

Request	Why refused	Instead
Resampling, clustering, file IO, unit conversion	not math	do it in data preparation, and pass a parameter
Unit checking at load	a unit: MW on a parameter is a claim that nothing checks against the column. A clean pass would only mean that no two annotated operands disagreed. It would also need a grammar of units that the language then has to maintain (#125)	convert to one unit system in data preparation, and name it in the description:
Array operations such as merge and reindex	there is no end to them	data preparation
Helpers for one domain, such as reduce_carrier_dim	writes one field's vocabulary into the language	a component library of macros over the operators that exist
A vocabulary for tracked metrics: impacts:, effects:, a costs axis	a named expression already does this	an impact dimension and one named expression. Cap it with a constraint, whose dual is the shadow price; weight it in the objective; read it back after the solve (#124)
** with a variable in the base or the exponent	the exponent would decide the degree, and to_spec reads no data. p ** n is linear at n = 1, quadratic at n = 2, and refused at n = 3	x * x for a square. ** over parameters and numbers is allowed (#1175)
Normalisation, x / sum(x)	dividing by a variable is not a polynomial, and no solver takes it	write the ratio as a constraint, or fix the denominator
An if, a loop, or declarations that depend on the data	to_spec could no longer read the file without the data	where: masks and dims: dimensions. A tool may loop over models
A Python API for building models	the model is the file you review and diff	YAML, or a dict with the same keys (below)

Composition (component libraries)

A component library is a set of templates, such as a boiler, a battery and a line, that agree on how ports and flows are named. You merge the templates you need into one file, wire the components together with a connectivity table in the data, and close the system with one sum(by=) balance.

The topology is data. Adding a second battery is a row in a table, not a second block of YAML, so the file grows with the number of component types and not with the number of components.

Merging happens before to_spec. Every function here takes a dict as well as a path, so a model assembled in Python is checked exactly as a file is, and Spec.to_yaml() writes the file a reviewer reads. A dict may hold only what a file may hold. There is no Python API that builds models any other way.

Two requests were closed against this design. A built-in merge (#30) and namespaces so that two templates can each declare a p (#29) are both things a library does before it hands over a dict. Arithmetic in bounds:, which signed and bidirectional flows need, is still open as #31. A component whose number of ports is only known at run time belongs in the library, which emits more rows or more templates, and never one block of YAML per component.