Tiny-instance validation

P5.6 checks correctness by exhaustive enumeration on a tiny formulation. It is a proof for that finite instance only; it is not an optimality claim for a realistic corridor.

The committed formulation is notes/reference/p5-formulation.v1.json and the committed optimum is notes/reference/p5-tiny-optimum.v1.json. The tiny instance has 6 variables and 324 assignments.

Enumerate and compare

enumerate_assignments returns every complete variable-to-option assignment in stable order. exhaustive_optimum evaluates all assignments twice:

  1. one exact evaluation per assignment;

  2. one exact batch evaluation.

TinyOptimum.batch_agreement is true only when every pair matches. The best feasible scalar is kept; ties are counted and resolved by the lexicographically smallest assignment. If no feasible assignment exists, the optimum fields are None.

from brushline.optimization.validation import (
    exhaustive_optimum,
    validate_tiny_instance,
)

optimum = exhaustive_optimum(evaluator, weights)
assert optimum.batch_agreement
report = validate_tiny_instance(evaluator, weights=weights, search=run)
assert report.replay_ok

Known optimum

The committed tiny optimum is balanced-scalarized:

  • 324 assignments;

  • 72 feasible assignments;

  • unique best assignment;

  • best assignment {0: 1, 1: 1, 2: 0, 3: 0, 4: 0, 5: 0};

  • best vector: present cost 100, near-zero service risk, service attainment 1, uncertainty 0.3253.

The best plan mows the shrub segment in both periods. A no-action plan misses the hard service height after growth.

Search evidence

validate_tiny_instance compares an optional SearchRun with the exhaustive optimum:

  • gap is search scalar minus optimum scalar for a feasible search result;

  • gap is None for an infeasible search or an instance with no feasible assignment;

  • reaches_optimum is true only at a zero gap;

  • replay_ok re-evaluates the search assignment and compares feasibility and scalar value;

  • expected_ok compares a committed optimum’s assignment, objective, and assignment count.

A positive gap is reported, not hidden.

Reports

TinyOptimum uses brushline.tiny-optimum.v1. TinyValidationReport uses brushline.tiny-validation.v1 and embeds the optional search run. Both have JSON writers and loaders.

Limits

  • 324 assignments is a unit test, not a production search.

  • The exhaustive result depends on the declared weights and contexts.

  • Batch agreement checks exact evaluator plumbing, not model correctness.

  • No MIP/CP bound is claimed; the finite enumeration is the bound for this instance.