U4 loop-close scripts: regenerate ref-524 vector, port check, 4B fit (pueue 549-551)

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

scripts/u4_step1_ref524.py

@@ -0,0 +1,56 @@
+"""U4 step 1/3: regenerate run-524's verified jacobian_word vector.
+
+(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
+
+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
+artifact instead of regenerating them (no drift axis).
+
+Exact run-524 parameters: Qwen/Qwen3-4B, persona=authority, n_pairs=256,
+seed=0, layers "mid" (7..27 of 36), words authority/obey/command/hierarchy,
+batch_size=4, max_length=384, cotangent_scope=source_scope=all_valid (defaults).
+"""
+import json
+from pathlib import Path
+
+import torch
+from loguru import logger
+from steering_lite.data import PERSONA_REGISTRY, make_persona_pairs
+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"
+MODEL = "Qwen/Qwen3-4B"
+WORDS = ["authority", "obey", "command", "hierarchy"]
+
+tok = AutoTokenizer.from_pretrained(MODEL)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL, torch_dtype=torch.bfloat16).to("cuda").eval()
+
+n = model.config.num_hidden_layers
+assert n == 36, f"expected Qwen3-4B with 36 layers, got {n}"
+layers = tuple(range(max(2, int(n * 0.2)), min(n - 2, int(n * 0.8))))  # run_sweep "mid" -> 7..27
+
+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
+(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}")
+logger.info("SHOULD: chat-templated authority-persona prompt with <think>. ELSE template drift.\n"
+            f"--- PROMPT[0] (full, special tokens) ---\n{prompts[0]}")
+
+vec = extract_word_pullback(model, tok, prompts, layers, WORDS,
+                            batch_size=4, max_length=384)["jacobian_word"]
+ref = {str(l): vec.stacked[l]["v"].squeeze(0).float().cpu() for l in layers}
+torch.save(ref, ART / "u4_ref_524.pt")
+logger.info(f"saved {ART / 'u4_ref_524.pt'}  "
+            f"norms={[round(ref[str(l)].norm().item(), 3) for l in layers][:5]}... "
+            "SHOULD: all 1.0 (unit vectors). ELSE _to_vector changed.")

scripts/u4_step2_vjp.py

@@ -0,0 +1,46 @@
+"""U4 step 2/3: port check -- jsteer's word_vector_vjp vs the run-524 reference.
+
+(Claude) Runs in jsteer's venv on the step-1 artifacts. Both sides are fp32
+direct-VJP extractions of the same math on the same 512 prompts; the only
+difference is code lineage (old experiment package vs this library) plus
+batch-order fp noise. GATE: cos > 0.999 per layer. A failure is a PORT BUG
+(position mask, pooling, layer indexing) -- debug, do not tune.
+"""
+import json
+from pathlib import Path
+
+import torch
+from loguru import logger
+from tabulate import tabulate
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from jsteer import word_vector_vjp
+
+ART = Path(__file__).resolve().parent.parent / "artifacts"
+meta = json.loads((ART / "u4_prompts.json").read_text())
+ref = torch.load(ART / "u4_ref_524.pt")
+
+tok = AutoTokenizer.from_pretrained(meta["model"])
+model = AutoModelForCausalLM.from_pretrained(
+    meta["model"], torch_dtype=torch.bfloat16).to("cuda").eval()
+
+v = word_vector_vjp(model, tok, meta["prompts"], meta["words"],
+                    layers=meta["layers"], batch_size=4, max_length=384)
+torch.save({str(l): v.stacked[l]["v"].squeeze(0).float().cpu() for l in meta["layers"]},
+           ART / "u4_vjp.pt")
+
+rows = []
+for l in meta["layers"]:
+    a = v.stacked[l]["v"].squeeze(0).float()
+    b = ref[str(l)].float()
+    rows.append((l, torch.nn.functional.cosine_similarity(a, b, dim=0).item()))
+table = tabulate(rows, headers=["layer", "cos(jsteer_vjp, ref524)"], floatfmt="+.6f")
+min_cos = min(c for _, c in rows)
+verdict = "PASS" if min_cos > 0.999 else "FAIL"
+out = (f"U4 step 2: jsteer word_vector_vjp vs regenerated run-524 vector\n"
+       f"model={meta['model']} prompts={len(meta['prompts'])} words={meta['words']}\n\n"
+       f"{table}\n\nmin cos = {min_cos:+.6f}   GATE (>0.999): {verdict}\n")
+(ART / "u4_step2_vjp_parity.txt").write_text(out)
+print(out)
+if verdict == "FAIL":
+    raise SystemExit("U4 step 2 FAILED: port bug, do not run step 3 until root-caused")

scripts/u4_step3_fit4b.py

@@ -0,0 +1,59 @@
+"""U4 step 3/3: full 4B Jacobian fit on run-524's substrate + cache loop-close.
+
+(Claude) The expensive one: 512 prompts x ceil(2560/dim_batch) backwards.
+checkpoint_path makes it resumable, so a kill/OOM loses at most one prompt.
+After fitting, the cached word vector must match BOTH the step-2 jsteer VJP
+vector and the step-1 run-524 reference (linearity: mean_p(J_p)^T w =
+mean_p(J_p^T w); fp16 cache storage is the only gap). GATE: cos > 0.999.
+
+This closes the loop on the verified 3/5 moral-foundations result: the library
+artifact (artifacts/qwen3-4b-authority.jac) provably contains the verified
+steering vector.
+"""
+import json
+import time
+from pathlib import Path
+
+import torch
+from loguru import logger
+from tabulate import tabulate
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from jsteer import Jacobian
+
+ART = Path(__file__).resolve().parent.parent / "artifacts"
+meta = json.loads((ART / "u4_prompts.json").read_text())
+ref524 = torch.load(ART / "u4_ref_524.pt")
+vjp = torch.load(ART / "u4_vjp.pt")
+
+tok = AutoTokenizer.from_pretrained(meta["model"])
+model = AutoModelForCausalLM.from_pretrained(
+    meta["model"], torch_dtype=torch.bfloat16).to("cuda").eval()
+
+t0 = time.time()
+jac = Jacobian.fit(model, tok, meta["prompts"], layers=meta["layers"],
+                   dim_batch=16, max_seq_len=384,
+                   checkpoint_path=str(ART / "qwen3-4b-authority.ckpt"))
+logger.info(f"fit wall-time: {(time.time() - t0) / 3600:.2f} h")
+jac.save(str(ART / "qwen3-4b-authority.jac"))
+logger.info(f"saved cache: {(ART / 'qwen3-4b-authority.jac').stat().st_size / 1e9:.2f} GB")
+
+v = jac.word_vector(model, tok, meta["words"], layers=meta["layers"])
+rows = []
+for l in meta["layers"]:
+    a = v.stacked[l]["v"].squeeze(0).float()
+    rows.append((l,
+                 torch.nn.functional.cosine_similarity(a, vjp[str(l)].float(), dim=0).item(),
+                 torch.nn.functional.cosine_similarity(a, ref524[str(l)].float(), dim=0).item()))
+table = tabulate(rows, headers=["layer", "cos(cache, jsteer_vjp)", "cos(cache, ref524)"],
+                 floatfmt="+.6f")
+min_cos = min(min(r[1], r[2]) for r in rows)
+verdict = "PASS" if min_cos > 0.999 else "FAIL"
+out = (f"U4 step 3: cached-4B word vector vs step-2 VJP and run-524 reference\n"
+       f"model={meta['model']} prompts={len(meta['prompts'])} words={meta['words']} "
+       f"dim_batch=16 fp16-cache\n\n{table}\n\n"
+       f"min cos = {min_cos:+.6f}   GATE (>0.999): {verdict}\n")
+(ART / "u4_loopclose.txt").write_text(out)
+print(out)
+if verdict == "FAIL":
+    raise SystemExit("U4 step 3 FAILED: cache wiring bug, root-cause before shipping the artifact")