Verify accounting identities of a solved scenario — verify

Recomputes core model identities (commodity balance and its sign per commodity limtype, the source/sink total layers, cost decomposition, and the discounted objective) from the solution tables and the scenario's own maps, and reports any violations. A valid optimal solution from any backend must pass every check; a failure indicates a solver, template, writer, or reader defect – or a corrupted solution.

Usage

verify_solution(
  object,
  checks = "default",
  tol_rel = 1e-06,
  tol_abs = 1e-06,
  verbose = FALSE
)

Arguments

object: a solved scenario (must carry modOut variables).
checks: character. Individual check names, or the sentinels "default" (the cheap set, run everywhere) and "all". See verify_checks() for the vocabulary, each check's group, and which data path it derives its expectation from. Defaults to "default".
tol_rel, tol_abs: numeric tolerances; a row violates its identity when |lhs - rhs| > tol_abs + tol_rel * max(|lhs|, |rhs|).
verbose: logical, report each check and its elapsed time as it runs. The opt-in checks join the whole solution table on a large model, where silence for minutes is not useful.

Value

an object of class solution_verification: a list with ok (logical), scenario (name), checks – per check a list with status ("ok", "violated", "skipped"), n (rows checked), violations (data.table of offending rows with lhs, rhs, diff), stats, reason when skipped, and report for report-only checks – and divergence.

print() takes detail = c("full", "issues"); "issues" shows only violated, skipped and reporting checks. The full set requested is always computed either way – an issues-only compute mode could hide a check that silently did not run.

divergence is a data.frame, one row per check, worst first: status, n_compared, n_violated, max_abs, median_abs, mean_abs and max_rel. It summarises how far each identity came from closing even when the check passed, so tolerances can be set from the noise the backends actually produce rather than guessed. max_abs is what a tol_abs must clear; median_abs beside it separates one bad cell from a uniformly loose check. Aggregate identities (objective, cost) sum over the whole model, so their absolute divergence grows with model size while max_rel stays flat – which is why the relative term carries the tolerance on large models.

Examples

if (FALSE) { # \dontrun{
scen <- solve_model(mod, solver = solver_options$glpk)
vs <- verify_solution(scen)
vs$ok
print(vs)
} # }