How monce-optimization is metered.
| Budget | EC2 cost (approx) | Typical instance size | Use case |
|---|---|---|---|
| 1s | ~$0.001 | <500 vars | Small puzzles, demos |
| 5s | ~$0.005 | 500–5,000 vars | Standard Sudoku, small graphs |
| 10s | ~$0.01 | 5k–50k vars | Medium scheduling, coloring |
| 30s | ~$0.03 | 50k–200k vars | Large industrial instances |
| 60s | ~$0.06 | 200k+ vars | Maximum speculative compute |
The natural metering unit. Clause count after reduction × seconds of compute. This captures both problem complexity (more clauses = harder) and compute investment.
token_cost = uniform_clauses * budget_seconds
# Sudoku: 8,300 * 5 = 41,500 tokens
# PHP(12,11): 2,541 * 10 = 25,410 tokens
# Job-shop (100 jobs): ~500,000 * 30 = 15,000,000 tokens| Instance | Without reduction | With reduction | Savings |
|---|---|---|---|
| PHP(12,11) | ~32s budget needed | ~1s budget needed | 97% |
| Rand(250, 4.26) | ~85ms | ~33ms | 61% |
| Sudoku (hard) | 5s budget | 2s budget | 60% |
The uniform reduction doesn't just make solving faster — it makes it cheaper. Lower budgets needed means fewer EC2 seconds billed.
npdollars is speculative: you pay for time, not guaranteed success. The budget is a ceiling on effort. The uniform reduction improves the odds of success within a given budget by presenting the solver with a more tractable topology.
Think of it as: reduction = free preprocessing that converts expensive speculative compute into cheaper deterministic progress.