wip

scripts/scratch/parity_u1.py

@@ -3,7 +3,7 @@
 (authored by Claude)
 
 The two paths are linear-identical:  mean_p(J_p)^T w == mean_p(J_p^T w).
-Path A pulls the word cotangent through the CACHED pooled Jacobian.
+Path A pulls the word cotangent through the CACHED averaged Jacobian.
 Path B contracts the same cotangent inside per-prompt backward passes.
 The only expected gap is fp16 storage in the cache, so per-layer cosine must
 exceed 0.999. A failure is a WIRING bug (layer index, position mask, pooling),
@@ -25,7 +25,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 from jsteer import Jacobian, word_vector_vjp
 
 # Claude: repo root on path so `scripts.smoke` imports whether run as a file or -m.
-sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent.parent))  # scripts/scratch/ -> repo root
 from scripts.smoke import CACHE, DEVICE, DTYPE, MODEL, PROMPTS  # same inputs  # noqa: E402
 
 WORDS = ["happy", "joy"]
@@ -46,7 +46,7 @@ def main() -> None:
     layers = jac.layers  # exact int layers fitted by the smoke
     logger.info(f"cached layers={layers}")
 
-    # Path A: cached pooled Jacobian pullback.
+    # Path A: cached averaged Jacobian pullback.
     vA = jac.word_vector(model, tok, WORDS)
     # Path B: direct per-prompt VJP over the SAME prompts / layers / skip_first / max_length.
     vB = word_vector_vjp(model, tok, PROMPTS, WORDS, layers=layers, max_length=128)

scripts/scratch/u4_step1_ref524.py

@@ -2,13 +2,13 @@
 
 (Claude) Run this under j-steer-dev's venv, NOT jsteer's:
 
-    cd ../j-steer-dev && uv run python ../jsteer/scripts/u4_step1_ref524.py
+    cd ../j-steer-dev && uv run python ../jsteer/scripts/scratch/u4_step1_ref524.py
 
 There `import jsteer` resolves to the OLD experiment package (j-steer-dev/src),
 whose extract_word_pullback produced the verified 3/5 result. Run 524 never
 persisted its vector tensors (only eval JSONs), but the extraction is
 deterministic (seed-0 prompts, greedy, no sampling), so re-running it IS the
-reference. Also dumps the 512 substrate prompts so steps 2/3 consume this one
+reference. Also dumps the 512 fitting prompts so steps 2/3 consume this one
 artifact instead of regenerating them (no drift axis).
 
 Exact run-524 parameters: Qwen/Qwen3-4B, persona=authority, n_pairs=256,
@@ -25,7 +25,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from jsteer.pullback import extract_word_pullback  # OLD package (j-steer-dev/src)
 
-ART = Path(__file__).resolve().parent.parent / "artifacts"
+ART = Path(__file__).resolve().parent.parent.parent / "artifacts"  # scripts/scratch/ -> repo root
 MODEL = "Qwen/Qwen3-4B"
 WORDS = ["authority", "obey", "command", "hierarchy"]
 
@@ -40,7 +40,7 @@ layers = tuple(range(max(2, int(n * 0.2)), min(n - 2, int(n * 0.8))))  # run_swe
 persona_pairs, template = PERSONA_REGISTRY["authority"]
 pos, neg = make_persona_pairs(tok, n_pairs=256, thinking=True,
                               persona_pairs=persona_pairs, template=template, seed=0)
-prompts = pos + neg  # run_sweep feeds pos+neg as the linearization substrate
+prompts = pos + neg  # run_sweep feeds pos+neg as the prompts J is linearized on
 (ART / "u4_prompts.json").write_text(json.dumps(
     {"model": MODEL, "layers": list(layers), "words": WORDS, "prompts": prompts}))
 logger.info(f"dumped {len(prompts)} prompts, layers={layers}")

scripts/scratch/u4_step2_vjp.py

@@ -16,7 +16,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from jsteer import word_vector_vjp
 
-ART = Path(__file__).resolve().parent.parent / "artifacts"
+ART = Path(__file__).resolve().parent.parent.parent / "artifacts"  # scripts/scratch/ -> repo root
 meta = json.loads((ART / "u4_prompts.json").read_text())
 ref = torch.load(ART / "u4_ref_524.pt")