Skip to content

compounding_engine.md

Latest commit

 

History

History
83 lines (67 loc) · 2.68 KB

compounding_engine.md

File metadata and controls

83 lines (67 loc) · 2.68 KB
Raw

Multi-Episode Compounding Engine

scripts/run_mathgraph_compounding_engine.py is the repo-native ETP compounding runner. It exercises the loop:

episode -> constructors -> finite checking -> residuals -> obstruction atlas
-> repair constructors -> Lawbook update -> next episode

The runner is intentionally verifier-boundary strict. PQ-IR rows, route policies, residual obstructions, repair families, and Lawbook reuse signals are advisory scheduling objects. They cannot prove claims and cannot promote truth.

FALSE recovery is counted only when a finite magma satisfies the source equation globally and violates the target equation at a concrete witness assignment. Finite-search failure is recorded as residual evidence only. TRUE requires a proof-producing route; this runner emits no TRUE terminal forms.

Tiny Demo

python scripts/run_mathgraph_compounding_engine.py \
  --out-dir /tmp/mathgraph_compounding_demo \
  --episodes 2 \
  --tiny-demo

The tiny demo uses a built-in equation set and identity TRUE controls, so it is a wiring and boundary check rather than a real ETP performance claim.

Full ETP / SAIR Run

python scripts/run_mathgraph_compounding_engine.py \
  --equations /content/equations.txt \
  --matrix /content/etp_matrix_full_best_bool.npy \
  --out-dir /content/drive/MyDrive/MathGraph_Compounding_Run \
  --episodes 4 \
  --train-false 18000 \
  --eval-false 14000 \
  --eval-true 9000 \
  --route-train-false 18000 \
  --route-eval-false 10000 \
  --max-n 5 \
  --repair-steps 30 \
  --seed 20260524

The script refuses real mode when the equation or matrix files cannot be loaded.

Outputs

The output directory contains:

  • lawbook.sqlite
  • compounding_summary.json
  • compounding_report.md
  • gate_results.csv
  • cross_episode_policy_summary.csv
  • cross_episode_policy_eval.csv
  • cross_episode_obstruction_summary.csv
  • constructor_bank_manifest.csv
  • repair_gain_curve.csv
  • repair_selected_constructors.csv
  • quotient_repair_family_lawbook.csv
  • obstruction_atlas.csv
  • residual_queue.csv
  • true_proof_template_summary.csv

Metrics

  • yield_rate: fraction of held-out FALSE pairs recovered by finite countermodel checking.
  • true_contamination_count: number of TRUE controls incorrectly hit by a FALSE route. This should be zero.
  • obstruction_entropy: residual basin dispersion.
  • lawbook_reuse_rate: advisory constructor-family reuse from prior episodes.
  • episode_lift_vs_baseline: improvement over the generic route.

Compounding is evidence, not proof, when later episodes improve yield, residuals, obstruction entropy, Lawbook reuse, or named obstruction coverage while preserving zero TRUE contamination.