Open the think block for non-thinking-default templates, probe answer-slot readability

Qwen3.5 templates close an empty <think></think> by default, so the reader appended a
second <think> and the model saw </think>...<think>. enable_thinking=True fixes the
sequence; Qwen3 is unchanged (pmass 1.000 before and after).

The probe exists because Qwen3.5-0.8B reads the WVS battery at pmass 0.61-0.84, not
0.999, and a mushy readout would waste the big run.

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

justfile

@@ -28,3 +28,7 @@ wvs-steer-modal model="Qwen/Qwen3.5-27B" gpu="H200":
 
 wvs-steer-pull:
     uv run --group dev modal volume get --force moralmaps-wvs-steer outputs .
+
+# which candidate large model reads the answer slot cleanly enough to be worth a sweep
+wvs-steer-readable models="Qwen/Qwen3.5-27B,Qwen/Qwen3-32B":
+    uv run --extra steer --group dev modal run scripts/run_modal_wvs.py::readable --models {{models}}

scripts/probe_wvs_think_budget.py

@@ -0,0 +1,70 @@
+"""Does the WVS answer-slot readout hold on this model family, and at what think budget?
+
+Qwen3-0.6B answers the IW battery with pmass 0.999. Qwen3.5-0.8B read 0.783 at think=1 in the
+steering smoke, which would make every steered coordinate mushy. Before renting a big GPU, find out
+whether that is the think budget (the model is mid-thought when we force the answer slot) or the
+chat template (the prefill does not land where we think it does).
+
+  uv run --extra steer python scripts/probe_wvs_think_budget.py --model Qwen/Qwen3.5-0.8B
+"""
+from __future__ import annotations
+
+import argparse
+
+import numpy as np
+import torch
+from loguru import logger
+from tabulate import tabulate
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from moralmaps.iw_axes import resolve_items
+from moralmaps.read import read_items, resolve_answer_ids
+from moralmaps.wvs import build_instruments, load_wvs_all, model_axis_scores, read_model
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--model", default="Qwen/Qwen3.5-0.8B")
+    ap.add_argument("--think-budgets", default="1,16,64,256")
+    ap.add_argument("--batch-size", type=int, default=12)
+    ap.add_argument("--device", default="cuda")
+    ap.add_argument("--device-map", default=None, help="'auto' shards a large model over the GPUs")
+    args = ap.parse_args()
+
+    tok = AutoTokenizer.from_pretrained(args.model)
+    if tok.pad_token is None:
+        tok.pad_token = tok.eos_token
+    tok.padding_side = "left"
+    if args.device_map:
+        model = AutoModelForCausalLM.from_pretrained(args.model, dtype=torch.bfloat16,
+                                                     device_map=args.device_map).eval()
+    else:
+        model = AutoModelForCausalLM.from_pretrained(args.model, dtype=torch.bfloat16).to(args.device).eval()
+
+    resolved = resolve_items(load_wvs_all())
+    instrs, meta = build_instruments(resolved)
+
+    rows = []
+    for think in [int(t) for t in args.think_budgets.split(",")]:
+        read = []
+        for k, instr in enumerate(instrs):
+            read += read_items(model, tok, instr, instr.items,
+                               resolve_answer_ids(tok, instr.answer_space),
+                               max_think_tokens=think, batch_size=args.batch_size,
+                               verbose_first=(k == 0 and think == 1))
+        pm = [r["pmass_allowed"] for r in read]
+        x, y = model_axis_scores(read_model(read, meta), meta, resolved)
+        closed = sum(r["emitted_close"] for r in read)
+        rows.append([think, f"{np.mean(pm):.3f}", f"{np.min(pm):.3f}", f"{closed}/{len(read)}",
+                     f"{x:.4f}", f"{y:.4f}"])
+        logger.info(f"think={think}: mean pmass {np.mean(pm):.3f}")
+
+    print(tabulate(rows, tablefmt="pipe", headers=[
+        "think tokens", "mean pmass", "min pmass", "closed think", "x", "y"]))
+    print("\nSHOULD: pmass climbs toward ~1.0 as the think budget grows, and the coordinate settles.\n"
+          "ELSE, if pmass stays low at every budget, the prefill or chat template is wrong for this\n"
+          "family and no steered coordinate from it is comparable to the published map.")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_modal_wvs.py

@@ -83,6 +83,30 @@ def main(model: str = MODEL, methods: str = ",".join(METHODS), seeds: str = ",".
             print(f"{method}\ts{seed}\tFAILED\t{error}")
 
 
+@app.function(gpu=os.environ.get("WVS_GPU", "H200"), volumes={"/cache": cache}, timeout=60 * 60)
+def probe(model: str, device_map: str) -> str:
+    """Read the battery unsteered at several think budgets: is this model's answer slot readable?"""
+    from huggingface_hub import snapshot_download
+
+    snapshot_download(model)
+    argv = ["--model", model] + (["--device-map", device_map] if device_map else [])
+    out = subprocess.run([sys.executable, "scripts/probe_wvs_think_budget.py", *argv],
+                         cwd="/repo", check=True, capture_output=True, text=True)
+    return out.stdout
+
+
+@app.local_entrypoint()
+def readable(models: str = "Qwen/Qwen3.5-27B,Qwen/Qwen3-32B", device_map: str = ""):
+    """Which candidate large model reads cleanly enough to be worth a sweep? ~10 min per model."""
+    handles = {m: probe.spawn(m, device_map) for m in models.split(",")}
+    for m, handle in handles.items():
+        print(f"\n===== {m} =====")
+        try:
+            print(handle.get())
+        except Exception as error:
+            print(f"FAILED\t{error}")
+
+
 @app.local_entrypoint()
 def smoke():
     """Same image, mounts and Volume as the real fan-out, on the tiny random model."""

scripts/wvs_steer_sweep.py

@@ -141,7 +141,9 @@ def main() -> None:
     else:
         model = AutoModelForCausalLM.from_pretrained(args.model, dtype=dtype).to(args.device).eval()
 
-    n_blocks = model.config.num_hidden_layers
+    # Qwen3.5 ships as a VL wrapper (Qwen3_5ForConditionalGeneration), so the block count lives in
+    # config.text_config, not at the top level.
+    n_blocks = model.config.get_text_config().num_hidden_layers
     if args.layers == "mid":
         layers = tuple(range(max(2, int(n_blocks * 0.2)), min(n_blocks - 2, int(n_blocks * 0.8))))
     else:

slop/plans/20260918_wvs_steer_big.md

@@ -7,6 +7,15 @@ written by Claude (claude-opus-4.8 in pi), 2026-09-18. NOT yet approved by wassn
 > moral map how the cultural preference change when steered for honesty + credulity. I want to
 > try a few steering methods." -- wassname
 
+Decisions after the first pass (wassname, 2026-09-18):
+- honesty only, one axis, credulity dropped
+- WVS only, no MFQ-2. "that's the plot"
+- "$100 is fine if it works"
+
+Branch note: this worktree branched from 789829e, which the Gemini-pilot session later amended to
+4955468. Rebase or cherry-pick the steering commits onto 4955468 before any merge, do not carry the
+stale preregistration commit. (told to me over intercom by moralmap_add_astra)
+
 ## What exists already (observation, read from the code today)
 
 - `scripts/wvs_map.py --local-model X` already reads the 12-item Inglehart-Welzel battery with the
@@ -51,13 +60,15 @@ Why not the biggest: Kimi K3 does not fit one Modal container at any precision,
 fits as fp8 where the `vjp_delta` backward is a research project of its own. Qwen3.5-122B-A10B also
 gives a clean scaling story next to the published Qwen3-4B showcase, same family.
 
-## Preregistered confound (raise before running)
+## Preregistered confound
 
-X-axis item 2 is interpersonal trust, `"Most people can be trusted"`. A **credulity** steer is
-close to a paraphrase of that item, so an X shift may be lexical leakage, not a cultural move.
-Discriminator: recompute X with that item held out. If the credulity effect on X survives the
-holdout, it is a cultural move; if X collapses to base, it is item leakage and must be reported
-as such. Honesty has no equivalent overlap in the battery.
+The battery is 12 items, so one item can carry a whole apparent move. `scripts/wvs_steer_sweep.py`
+saves each item's position at every dose, and `scripts/plot_wvs_steer.py` reports the move length
+again with the single most influential item removed. A move that mostly disappears under that
+leave-one-out is one item reacting to the persona wording, not a cultural shift.
+
+This mattered most for credulity, which nearly paraphrases the X-axis trust item
+(`"Most people can be trusted"`). Credulity is now out of scope, but the check is cheap and stays.
 
 ## Goals
 

src/moralmaps/guided.py

@@ -27,9 +27,14 @@ _ASSISTANT_SENTINEL: str = "ZZUNIQ_ASSISTANT_SENTINEL_ZZ"
 
 
 def _generation_prompt_with_open_think(tok, messages: list[dict[str, str]]) -> str:
-    """Return exactly one open reasoning marker. (Claude, 2026-07-19)"""
+    """Return exactly one open reasoning marker. (Claude, 2026-07-19)
+
+    enable_thinking=True matters for templates that default to non-thinking: Qwen3.5 otherwise emits
+    a CLOSED empty `<think></think>`, we then append our own `<think>`, and the model reads a
+    `</think> ... <think>` sequence it never saw in training. Templates without the variable ignore
+    it (Qwen3 is unchanged)."""
     prompt = tok.apply_chat_template(
-        messages, tokenize=False, add_generation_prompt=True)
+        messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
     if prompt.rstrip().endswith("<think>"):
         return prompt
     return prompt + "<think>\n"