axis = SocialNorms/Care (Authority degenerate); over-steer generation

scripts/diag_axis.py shows steering at 1 nat moves gemma's foundation profile
the right way: SocialNorms 0.68->0.42, Care 0.21->0.33, coherence 0.72->0.88.
Authority is ~0 on this model (no headroom), so:
- eval reports all foundations; trait axis = SocialNorms (down) + Care (up)
- map.html plots Care vs SocialNorms
- add gen_alpha=1.5: over-steer generation into the incoherent regime so the
  heal (Q1) has work to do (at 1 nat coherence improved, nothing to heal)
- results.py groups on coherence/socialnorms/care

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

docs/RESEARCH_JOURNAL.md

@@ -23,3 +23,20 @@ Bug found: iso-KL calibration could not reach `target_kl=1.0`. c_star pinned at
 **Fix:** pass `bracket=(0.1, 1024.0)` to `v.calibrate`. Re-running to confirm an interior c_star with p95 KL ~ 1.0.
 
 **Also to investigate:** auth=0.000 exactly — is gemma-3-1b-it genuinely never attributing the Authority foundation on these 24 vignettes, or a metric/profile issue? Check once steering is strong enough to move things.
+
+## Steering validity confirmed; real axis is SocialNorms/Care, not Authority
+
+`scripts/diag_axis.py` on gemma-3-1b-it, base vs steered at calibrated c_star=67.7 (~1 nat p95 KL). The vector moves the moral-foundation profile in the right direction for "less deference to authority":
+
+| foundation  | base  | steered | Δ      |
+|-------------|-------|---------|--------|
+| SocialNorms | 0.680 | 0.421   | -0.260 |
+| Care        | 0.213 | 0.328   | +0.115 |
+| Fairness    | 0.030 | 0.098   | +0.069 |
+| Liberty     | 0.040 | 0.075   | +0.035 |
+| Authority   | 0.000 | 0.001   | +0.001 |
+| coherence   | 0.722 | 0.884   | +0.162 |
+
+**Interpretation:** the core premise holds, steering shifts moral judgments coherently toward the trait. But (1) Authority is degenerate on this model (~0), so the eval/plot axis must be **SocialNorms (down) and Care (up)**, not Authority. (2) At the 1-nat dose coherence went UP, not down, so there is little incoherency to heal at alpha=1. To give Q1 (heal) something to do we must generate training data at higher alpha (~1.5-2 nats, where the iso-KL repo finds "dead" traces) or rely on long-trajectory drift.
+
+**Changes:** eval reports all foundations; map uses Care vs SocialNorms; add `gen_alpha` (default 1.5) so generation over-steers into the incoherent regime while calibration stays at 1 nat.

scripts/diag_axis.py

@@ -0,0 +1,51 @@
+"""Diagnostic: does the steering vector move the moral-foundation profile, and where?
+
+Base gemma-3-1b-it puts ~0 on the Authority foundation (forced-choice), so the
+"authority axis" has no headroom. This prints base vs steered (at calibrated
+c_star) 7-foundation profiles side by side so we can pick the axis the trait
+actually moves. Run: uv run python scripts/diag_axis.py
+"""
+
+import sys
+
+import torch
+import tinymfv
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+sys.path.insert(0, "src")
+from steer_heal.config import RunConfig  # noqa: E402
+from steer_heal.steering import teacher_vec  # noqa: E402
+
+MODEL = "google/gemma-3-1b-it"
+cfg = RunConfig(model=MODEL, n_prompts=12)
+
+tok = AutoTokenizer.from_pretrained(MODEL)
+if tok.pad_token is None:
+    tok.pad_token = tok.eos_token
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL, torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="eager"
+).eval()
+
+v = teacher_vec(model, tok, cfg)
+
+
+def profile(label):
+    rep = tinymfv.evaluate(model, tok, name="classic", n_vignettes=24,
+                           conditions=("other_violate",), max_think_tokens=64, device=model.device)
+    p = dict(zip(rep["profile"]["foundation"], rep["profile"]["model"]))
+    p["_coherence"] = rep["mean_pmass_allowed"]
+    print(f"\n=== {label} ===")
+    for k, x in p.items():
+        print(f"  {k:12s} {x:.4f}")
+    return p
+
+
+base = profile("BASE (c=0)")
+with v(model, C=v.cfg.coeff):
+    steer = profile(f"STEERED (c_star={v.cfg.coeff:.1f}, ~1 nat)")
+
+print("\n=== delta (steered - base), sorted by |Δ| ===")
+keys = [k for k in base if not k.startswith("_")]
+for k in sorted(keys, key=lambda k: -abs(steer[k] - base[k])):
+    print(f"  {k:12s} {base[k]:+.4f} -> {steer[k]:+.4f}   Δ={steer[k]-base[k]:+.4f}")
+print(f"  coherence    {base['_coherence']:.3f} -> {steer['_coherence']:.3f}")

scripts/results.py

@@ -19,13 +19,14 @@ df = pl.read_csv(RESULTS_TSV, separator="\t")
 agg = (
     df.group_by(GROUP)
     .agg(
-        pl.col("p_ans_any").mean().round(3).alias("coherence"),
-        pl.col("auth").mean().round(3),
-        pl.col("auth").std().round(3).alias("auth_sd"),
+        pl.col("coherence").mean().round(3),
+        pl.col("socialnorms").mean().round(3),  # trait axis: lower = more trait
+        pl.col("care").mean().round(3),
+        pl.col("care").std().round(3).alias("care_sd"),
         pl.len().alias("n"),
         pl.col("seed").cast(pl.Utf8).sort().str.join(",").alias("seeds"),
         pl.col("argv").first(),
     )
-    .sort("auth", descending=True)
+    .sort("care", descending=True)
 )
 print(tabulate(agg.to_pandas(), headers="keys", tablefmt="pipe", floatfmt="+.3f"))

src/steer_heal/config.py

@@ -23,6 +23,7 @@ class RunConfig:
     neutral: str = "You are a helpful assistant."
     layer_range: tuple[float, float] = (0.4, 0.6)  # fraction of depth to steer
     target_kl: float = 1.0  # iso-KL p95 dose (nats)
+    gen_alpha: float = 1.5  # over-steer generation into the incoherent regime (heal has work to do)
     alphas: tuple[float, ...] = (0.5, 1.0, 1.5, 2.0)  # multiples of c_star to generate at
 
     # ── generation + filter (U1) ──

src/steer_heal/eval.py

@@ -24,17 +24,21 @@ def evaluate_model(model, tok, cfg: RunConfig) -> dict:
         device=model.device,
     )
     prof = rep["profile"]  # pandas: foundation, human, model, model_T
-    model_p = dict(zip(prof["foundation"], prof["model"]))
-    # SHOULD: auth/care in [0,1], coherence ~ base level on a working model;
-    # a sharp coherence drop after steering = format collapse. On tiny-random
-    # the numbers are junk (we test the path, not the value).
+    p = dict(zip(prof["foundation"], prof["model"]))
+    # The trait "less deference to authority" moves SocialNorms DOWN and Care UP
+    # on gemma-3-1b-it (Authority is degenerate ~0; see RESEARCH_JOURNAL 2026-06-04).
+    # Report all foundations so we never lose the axis that actually moves.
+    # SHOULD: under steering, socialnorms drops and care rises; coherence holds.
     out = {
-        "auth": float(model_p["Authority"]),
-        "care": float(model_p["Care"]),
+        "socialnorms": float(p["SocialNorms"]),  # trait axis: DOWN = more trait
+        "care": float(p["Care"]),                # trait axis: UP = more trait
+        "auth": float(p["Authority"]),
+        "fairness": float(p["Fairness"]),
+        "liberty": float(p["Liberty"]),
         "coherence": float(rep["mean_pmass_allowed"]),
         "ppx_json": float(math.exp(rep["mean_nll_json"])),
         "top1_acc": float(rep["top1_acc"]),
     }
-    logger.info(f"eval: auth={out['auth']:.3f} care={out['care']:.3f} "
+    logger.info(f"eval: socialnorms={out['socialnorms']:.3f} care={out['care']:.3f} "
                 f"coherence={out['coherence']:.3f} ppx={out['ppx_json']:.1f}")
     return out

src/steer_heal/plot.py

@@ -15,20 +15,20 @@ def write_map(run_dir: Path, rounds: list[dict]) -> Path:
     r = [d["round"] for d in rounds]
     fig = make_subplots(
         rows=1, cols=2, column_widths=[0.6, 0.4],
-        subplot_titles=("trait map: Care vs Authority", "coherence + direction per round"),
+        subplot_titles=("trait map: Care vs SocialNorms", "coherence + direction per round"),
         specs=[[{"type": "scatter"}, {"type": "scatter"}]],
     )
-    # trajectory across the auth axis, coloured by round
+    # trajectory across the SocialNorms axis (trait moves it DOWN, Care UP), coloured by round
     fig.add_trace(go.Scatter(
-        x=[d["auth"] for d in rounds], y=[d["care"] for d in rounds],
+        x=[d["socialnorms"] for d in rounds], y=[d["care"] for d in rounds],
         mode="lines+markers+text", text=[f"r{i}" for i in r], textposition="top center",
         marker=dict(size=12, color=r, colorscale="Viridis", showscale=False),
         hovertext=[f"r{d['round']} coh={d['coherence']:.3f} cos={d.get('cos_v0', float('nan')):.2f}"
                    for d in rounds],
         name="trajectory",
     ), row=1, col=1)
-    fig.update_xaxes(title_text="Authority p (trait →)", row=1, col=1)
-    fig.update_yaxes(title_text="Care p", row=1, col=1)
+    fig.update_xaxes(title_text="SocialNorms p (← trait)", row=1, col=1)
+    fig.update_yaxes(title_text="Care p (trait →)", row=1, col=1)
 
     fig.add_trace(go.Scatter(x=r, y=[d["coherence"] for d in rounds],
                              mode="lines+markers", name="coherence"), row=1, col=2)

src/steer_heal/run.py

@@ -67,7 +67,7 @@ def steer_heal(model, tok, cfg: RunConfig, run_dir: Path) -> dict:
         # extract teacher vector + generate steered data from the CURRENT student
         with baked(model, hist_specs):
             v = teacher_vec(model, tok, cfg)
-            comps = generate_steered(model, tok, v, alpha=1.0, cfg=cfg)
+            comps = generate_steered(model, tok, v, alpha=cfg.gen_alpha, cfg=cfg)
         # filter under the ORIGINAL (no history, no steering)
         kept, scored = filter_completions(model, tok, comps, cfg)
         log_event(run_dir, stage="gen", round=rnd, n_comps=len(comps), n_kept=len(kept), scored=scored)
@@ -87,7 +87,7 @@ def steer_heal(model, tok, cfg: RunConfig, run_dir: Path) -> dict:
         rec = {"round": rnd, **m, "cos_v0": cos_v0, "c_star": float(v.cfg.coeff), "n_kept": len(kept)}
         rounds.append(rec)
         log_event(run_dir, stage="round", **rec)
-        logger.info(f"round {rnd}: auth={m['auth']:.3f} care={m['care']:.3f} "
+        logger.info(f"round {rnd}: socialnorms={m['socialnorms']:.3f} care={m['care']:.3f} "
                     f"coh={m['coherence']:.3f} cos_v0={cos_v0:+.2f}")
 
     map_path = write_map(run_dir, rounds)