Independent Research August 2026

Testing a solver prototype before further development

Eric Luttrell’s independent research used generated configuration-and-allocation problems to evaluate a solver prototype. The work explored how to check proposed solutions and decide whether the approach warranted further development.

The question

Could a prototype combine specialized solvers, check their proposed solutions against the original problem, and perform well enough to justify further development?

Separate solution validity from practical performance, and compare all three methods against the same original problem constraints.

See the result
Benchmark results
MethodProblems solved
Combined prototype28 / 80
MiniZinc / CP-SAT80 / 80
Z3 baseline3 / 80
Same 80 problems; three measured repetitions per method. Solved means at least two valid, checked solutions within the time limit.

Method

Define the test. Preserve the evidence.

The research fixed the test protocol before measurement, compared three methods on the same configuration-and-allocation problems, and replayed accepted solutions through a shared source checker. A chained execution log preserved the individual results.

What the work produced

A completed benchmark, reproducible evidence records, a verification report, and a documented decision about the prototype’s viability.

Tools used Python · CaDiCaL · HiGHS · Z3 · MiniZinc / CP-SAT

Result

Evidence for the next decision.

The combined prototype solved 28 of 80 benchmark problems; the MiniZinc/CP-SAT baseline solved all 80. The completed campaign recorded no false accepts under its source checks, but the prototype missed its viability threshold. The result supported narrowing the research claim or redesigning the prototype before further development.

Decision: Narrow the research claim or redesign the prototype.

Evaluation

How the result was checked.

The completed campaign contains 735 execution records: 15 warmups and 720 measurements. Each of the three methods ran three measured repetitions on each of 80 problems. A problem counts as solved when at least two repetitions return a valid, checked solution within the 60-second limit.

Review the evidence checks
  • The completed August 16, 2026 run supplied the benchmark results; an earlier interrupted run was excluded.
  • A September 5 evidence review verified all 735 record hashes and all 80 input identities. It replayed all 343 accepted records through the existing source checker and reproduced the reported solved counts.
  • This review checked stored research evidence. It did not repeat the solver campaign or provide an independently implemented checker.

Scope

Where these findings apply.

  • These findings apply to the tested configuration-and-allocation problem family and the frozen implementation.
  • They do not establish production readiness, general solver superiority, or performance on customer systems.
  • This is independent research using generated test problems, not customer past performance. No false accepts in these checks is a bounded test result, not a guarantee for arbitrary inputs.

Need evidence for your next technical decision?

Start with the question, the system, and what you need to learn.

Discuss your question