Quickstart

PMH is the matching principle from main.pdf: estimate $\hat{\Sigma}_{\text{task}}$, train with matched PMH on $h$, run Step 5 on deploy holdout. Pick the closest task in T1–T7 order.

pip install matching-pmh torch
pip install "matching-pmh[sklearn]"   # T1
pmh-train try --quick                 # train + ship verdict (see docs/START.md)
pmh-train doctor
pmh-train evaluate --demo
pmh-train route --list

---

T1 — frozen features (sklearn, paper block 1)

Task doc · Notebook

from pmh import load_g2_demo_arrays, evaluate_baseline_vs_pmh

xs, ys, xt, yt = load_g2_demo_arrays()
print(evaluate_baseline_vs_pmh(xs, ys, xt, yt, preset="t1_synthetic_sklearn").summary())

---

T4A — vision domain shift (PyTorch, block 4)

Task doc · Notebook

from pmh import check_applicability

print(check_applicability(stack="pytorch", has_target_domain=True).summary())

---

Step 5 + benchmark arms

from pmh import evaluate_robust_fit, format_five_step_guide
from pmh.benchmark.protocol import run_benchmark_protocol

print(format_five_step_guide("domain_shift"))
# report = run_benchmark_protocol(...)  # matched / wrong / isotropic + geometry

Ship only when matched beats wrong-direction and isotropic on deploy holdout (WHEN_PMH_HELPS).

---

Next

Goal	Doc
Deploy-change table (T1–T7)	README
All task pages	tasks/index.md
Short theory spine	PRINCIPLE.md
Theory + theorems	main.pdf
Honest expectations	Will PMH help?