Sudoku Swarm: Test STM coordination on constraint propagation

t-369.26·WorkTask·
·
·
Parent:t-369·Created1 month ago·Updated1 month ago

Description

Edit

Test STM-based swarm on Sudoku solving where constraint propagation has true dependencies.

Hypothesis

Unlike contract review (embarrassingly parallel), Sudoku has TRUE inter-agent dependencies:

  • When Agent A discovers cell(1,3)=7, this ELIMINATES 7 from row 1, column 3, and box
  • Agent B working on same row/column/box can use this immediately
  • Discoveries directly affect other agents' search space

Deliverables

1. SudokuSwarm.hs

  • SharedState: TVar for board (9x9 Maybe Int) and candidates (9x9 Set Int)
  • Single agent: one agent solves entire puzzle
  • Parallel no-share: agents work on different regions, merge at end
  • STM swarm: agents share discoveries via TVar, real-time propagation

2. Test Puzzles

  • Easy (30+ givens): baseline comparison
  • Medium (25-29 givens): where sharing might help
  • Hard (17-24 givens): where sharing should really matter

3. Metrics

  • Solved? (correctness)
  • Time to solution
  • Iterations/LLM calls
  • Cells solved via propagation vs guessing

Expected Results

  • Single agent: must track all constraints, may lose track
  • Parallel no-share: duplicate work, may conflict
  • STM swarm: immediate propagation, should be faster/more reliable

Files

  • Omni/Agent/Experiments/SudokuSwarm.hs
  • Omni/Agent/Experiments/SUDOKU_RESULTS.md

Timeline (2)

🔄[human]Open → InProgress1 month ago
🔄[human]InProgress → Done1 month ago