Skip to content

Task 007: Homework for Meeting #10

Priority: HIGH Status: DONE Due: Monday 2026-03-23 (Meeting #10) Source: Meeting #9 homework assignment

Assignment

From Meeting #9:

  1. Get agents to improve sparse parity using DMC (Data Movement Complexity) as the energy proxy
  2. Iterate on prompts and meta-approaches to go from "metric spec + problem spec" to experiments quickly

Checklist

DMC Baseline Sweep

  • Run DMC measurement for GF(2) Gaussian Elimination
  • Run DMC measurement for KM Influence Estimation
  • Run DMC measurement for SGD (standard config: n=20, k=3, hidden=200, lr=0.1)
  • Run DMC measurement for SMT Backtracking
  • Run DMC measurement for Fourier/Walsh-Hadamard
  • Compile comparison table: method / accuracy / DMC / ARD / wall time

DMC Optimization

  • Pick the most promising method for DMC optimization
  • Run at least one experiment targeting DMC reduction
  • Record results in findings format (hypothesis, method, result, key number)
  • Update DISCOVERIES.md if findings answer an open question

Presentation Prep

  • Prepare results summary (table + key insight)
  • Note any differences between DMC ranking and ARD ranking
  • Record video or prepare demo if presenting async

Context

DMC is already tracked in MemTracker alongside ARD. Formula: DMC = sum(sqrt(stack_distance)) for all float accesses. Current baseline from CLAUDE.md: ARD 4,104 / DMC 300,298.

Key question: does optimizing DMC lead to different algorithmic choices than optimizing ARD? If the rankings change, that's a finding.

Files

  • src/sparse_parity/tracker.py -- MemTracker with DMC
  • src/sparse_parity/cache_tracker.py -- CacheTracker with LRU simulation
  • src/sparse_parity/harness.py -- Locked evaluation harness (DO NOT MODIFY)