diff --git a/README.md b/README.md
index 77e4d86..22924d4 100644
--- a/README.md
+++ b/README.md
@@ -1,53 +1,18 @@
-# tinymfv (tiny moral/value eval for local LLMs)
+# tinymfv
 
-tinymfv reads answer-token logprobs from local chat LMs and turns them into moral/value profiles.
-It is meant for fast steering experiments: prefill the answer slot, read the next-token
-distribution, and compare base vs steered runs in nats instead of waiting for sampled answers to
-flip.
+tinymfv is a small set of quick value evals for local LLM steering work. It is sensitive to small answer-probability shifts, so you can see movement before sampled answers flip.
 
-There are two instrument kinds:
+It asks moral vignettes and survey questions, reads the model's answer probabilities, and compares the model profile to human responses. The main use case is simple: after you steer a model, did the intended values move, did nearby values move too, and does the result still look like a coherent answer?
 
-- Nominal MFV vignettes: the answer is a foundation category. The profile is the mean probability
-  of care / fairness / loyalty / authority / sanctity / liberty / social.
-- Ordinal questionnaires: the answer is a scale point 1..M. MFQ-2, Big Five, 16PF, and Humor Styles
-  all use the same reader and reduce to per-factor Likert profiles.
+The plots are the point. Range plots show where the model sits relative to human societies. Culture maps show the base model and steered models on a PCA map of human response profiles.
 
-The default nominal instrument is the Clifford moral-foundation vignettes (MFV): 132 scenarios, each
-with a human distribution over foundations (Clifford et al. 2015). Three configs (`classic`,
-`scifi`, `ai-actor`), two framings each (`other_violate`, `self_violate`).
-[[HF dataset](https://huggingface.co/datasets/wassname/tiny-mfv)]
+![MFQ-2 range plot: human society ranges beside the fairness steer path](docs/img/showcase/mfq2/range.png)
 
-The bundled showcase steers Qwen3-4B with one activation vector, extracted once from
-`wassname/moral_stories_foundations` fairness situations (mean activation difference, fairness vs a
-balanced sample of the other foundations, situations only, no completions), iso-KL calibrated, then
-read across every instrument at base / +c / -c. On MFQ-2 it selectively raises the equality factor:
-equality is the only factor that rises at the gentlest +c, every other factor falls, a clean
-directional move on the fairness-equality axis. On the MFV vignettes, though, the same vector moves
-Care, not Fairness, so it is a care-leaning moral-salience direction that reads as fairness on the
-questionnaire rather than a clean single-foundation isolator. Coherence holds across the sweep
-(`frac_unscorable` 0, `mean_margin` ~10-12 nats), so the readouts below are not steering artifacts.
+![MFV culture map: base and fairness-steered points against human countries](docs/img/showcase/mfv/map_pca_ipsative.png)
 
-The same vector is read on all five bundled instruments. Range plots show the per-factor steer path
-(red `+c` up, blue `-c` down) against human society ranges; maps are PCA culture maps with the c-sweep
-trajectory. The 16PF map is omitted (16 axes do not lay out as a readable 2-D map; its range is kept).
+![MFV range plot: Care moves most under the fairness vector](docs/img/showcase/mfv/range.png)
 
-![MFQ-2 culture map: base and fairness-steered points against human societies](docs/img/showcase/mfq2/map_pca_ipsative.png)
-
-![MFQ-2 range plot: human society ranges beside the fairness steer path; equality rises at +c](docs/img/showcase/mfq2/range.png)
-
-![Big Five culture map with the c-sweep trajectory](docs/img/showcase/big5/map_pca_ipsative.png)
-
-![Big Five range: fairness steer path per trait against human society ranges](docs/img/showcase/big5/range.png)
-
-![Humor Styles culture map with the c-sweep trajectory](docs/img/showcase/humor_styles/map_pca_ipsative.png)
-
-![Humor Styles range: fairness steer path per style against human ranges](docs/img/showcase/humor_styles/range.png)
-
-![16PF range: fairness steer path across all 16 factors](docs/img/showcase/16pf/range.png)
-
-![MFV culture map with the c-sweep trajectory](docs/img/showcase/mfv/map_pca_ipsative.png)
-
-![MFV range: per-foundation emphasis; Care moves most under the fairness vector, Fairness barely](docs/img/showcase/mfv/range.png)
+More maps are in `docs/img/showcase/` for MFQ-2, Big Five, 16PF, and Humor Styles. The plotting script now defaults to clean base / +C / -C maps; use `--show-sweep` when you want the diagnostic `c=-4..+4` trajectory.
 
 ## Install
 
@@ -70,25 +35,42 @@ uv sync --extra maps --dev
 just smoke
 ```
 
-## Use it
+## Datasets
 
-Nominal MFV vignettes use `evaluate`:
+| dataset | what it asks | model answer | human comparison |
+|---|---|---|---|
+| MFV classic | 132 Moral Foundations Vignettes from Clifford et al. | one of 7 foundations: Care, Fairness, Loyalty, Authority, Sanctity, Liberty, Social Norms | per-vignette human foundation labels |
+| MFV scifi | the same MFV items rewritten as sci-fi scenarios | one of 7 foundations | inherited labels from the classic item |
+| MFV ai-actor | the same MFV items rewritten so an AI system is the actor | one of 7 foundations | inherited labels from the classic item |
+| MFQ-2 | 36 Moral Foundations Questionnaire items | 1-5 agreement | country foundation means, plus raw Atari et al. respondent data |
+| Big Five | 50 personality items | 1-5 agreement | country factor means |
+| 16PF | 162 personality items | 1-5 agreement | country factor means |
+| Humor Styles | 32 humor-style items | 1-5 agreement | country style means, originally on a 1-7 scale |
+
+MFV is nominal: the answer is the category. The survey instruments are ordinal: the answer is a scale point.
+
+Each MFV item is asked in two perspectives, `other_violate` and `self_violate`. Each survey item is asked three ways, forward, scale-inverted, and content-negated. tinymfv canonicalizes these frames before averaging, so the profile is less tied to one wording.
+
+## API
+
+Run MFV vignettes with `evaluate`:
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
-from tinymfv import load_vignettes, evaluate
+from tinymfv import evaluate, load_vignettes
 
 tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
 model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B").cuda()
 
-vignettes = load_vignettes("classic")              # list[dict], one per scenario
+vignettes = load_vignettes("classic")  # "classic", "scifi", "ai-actor", or "all"
 report = evaluate(model, tok, vignettes=vignettes, return_per_row=True)
-print(report["mean_pmass_allowed"])
-print(report["per_row"][0]["score"])               # logprob score per foundation, nats
-print(report["profile"])                           # mean p[foundation], easier to read
+
+print(report["profile"])              # mean probability per foundation
+print(report["mean_pmass_allowed"])   # probability mass on valid answer tokens
+print(report["per_row"][0]["score"])  # foundation logprobs, in nats
 ```
 
-Ordinal questionnaires use `administer`:
+Run survey instruments with `administer`:
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -97,165 +79,60 @@ from tinymfv import administer, get_instrument
 tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
 model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B").cuda()
 
-instr = get_instrument("mfq2")       # "mfq2", "big5", "16pf", or "humor_styles"
+instr = get_instrument("mfq2")  # "mfq2", "big5", "16pf", or "humor_styles"
 report = administer(model, tok, instr)
+
 print(report["dimensions"])
-print(report["per_item_frame"][0]["lp"])           # raw logprobs for the answer tokens
-print(report["profile_E"])           # human-comparable factor means
-print(report["profile_C"])           # steer-sensitive logit contrast per factor
-print(report["mean_pmass_allowed"])
+print(report["profile_E"])                # human-comparable scale means
+print(report["profile_C"])                # steering-sensitive log contrast
+print(report["per_item_frame"][0]["lp"])  # raw answer-token logprobs
 ```
 
-From a checkout, the small functional run is:
+## Metrics
 
-```bash
-just smoke
-just eval Qwen/Qwen3-0.6B classic
-```
+There are two metric families.
 
-The plotting helpers live in `tinymfv.maps`; `scripts/plot_steer_showcase.py` shows the current
-map/range pipeline for steering-lite outputs.
+Human-comparison metrics are bounded and easy to plot against survey data.
 
-## Reader math
+For MFV, the profile is the mean probability of each foundation:
 
-For an answer set $A = \{a_1,\ldots,a_K\}$, tinymfv gathers the full-vocab logprobs at the answer
-tokens:
+$$p_f = \mathbb{E}_i\,P(f \mid i)$$
 
-$$\ell_k = \log P(a_k \mid \text{prompt}, \text{think trace}, \text{assistant prefill})$$
+For surveys, `profile_E` is the expected scale score, reverse-keyed where needed:
 
-The raw gathered vector $\ell$ is the primitive. Everything else is a pure readout:
+$$E_i = \sum_{k=1}^{M} k p_{i,k}$$
 
-$$\mathrm{pmass\_allowed} = \sum_{k=1}^{K} \exp(\ell_k)$$
+Steering metrics are logprob based. They are more sensitive because they move before the sampled answer flips.
 
-$$p_k = \frac{\exp(\ell_k)}{\mathrm{pmass\_allowed}}$$
+At the answer slot, tinymfv gathers the logprobs for the allowed answers:
 
-`pmass_allowed` is answer-format coherence: mass on valid answer tokens at the answer slot. The
-slot is force-prefilled, so the scaffold primes a valid token and `pmass_allowed` is pinned high
-under the forced-choice path. Treat it as a format floor, not a sensitive coherence gate: a steered
-run that has broken the model can still score `pmass_allowed` near 1.0. The two signals that carry
-coherence instead are `frac_unscorable` (rows with no scorable slot, a self-closed think with no
-answer or a blown-up forward; near 0 once the end-of-answer tokens are suppressed over the budget)
-and `nll_prefill` below (does the scaffold still fit the model). pmass does not say which valid
-answer is right.
+$$\ell_k = \log P(a_k \mid \mathrm{prompt}, \mathrm{think}, \mathrm{prefill})$$
 
-`nll_prefill` measures whether the forced assistant prefill itself fits the model:
+For surveys, `profile_C` is the steering-sensitive direction score. It weights high agreement tokens positive and low agreement tokens negative, using logprobs instead of bounded survey means:
 
-$$\mathrm{nll\_prefill} = -\frac{1}{J}\sum_{j=1}^{J} \log P(u_j \mid \text{context}, u_{<j})$$
+$$C_i = \sum_{k=1}^{M} \left(k - \frac{M + 1}{2}\right)\ell_{i,k}$$
 
-where $u_1,\ldots,u_J$ are the prefill tokens before the answer token. This catches scaffold
-friction before the answer slot; `pmass_allowed` catches the slot itself.
+Use `delta C` for survey steering effects:
 
-## What to use
+$$\Delta C_i = C_i^{\mathrm{steered}} - C_i^{\mathrm{base}}$$
 
-Most research code should use the logprob-level outputs. They are sensitive and easy to interpret
-in relative terms: positive means the steer moved the answer up on that axis, negative means down,
-and the units are nats. For MFV, use per-row `score` or the paired `dlogit_per_foundation`. For
-ordinal questionnaires, use raw per-frame `lp` when you want the primitive, or `profile_C` when you
-want a factor-level steer readout.
+For MFV, use the paired foundation logit change:
 
-The value summaries are less sensitive but easier to interpret. MFV `profile` is mean
-foundation probability. Ordinal `profile_E` is the expected Likert score, in the human scale
-direction:
+$$\Delta_{i,f} = \operatorname{logit} p_{i,f}^{\mathrm{steered}} - \operatorname{logit} p_{i,f}^{\mathrm{base}}$$
 
-$$E = \sum_{k=1}^{M} k p_k$$
+`pmass_allowed` is a format check, not a value score:
 
-For reverse-keyed items, `keyed_E = M + 1 - E`. `E` is bounded, so it is good for comparing to human
-survey means but can hide small steering effects near confident answers. That is why the ordinal
-steering readout is the rank-centered logit contrast:
+$$\mathrm{pmass}_{\mathrm{allowed}} = \sum_{k=1}^{K}\exp(\ell_k)$$
 
-$$C = \sum_{k=1}^{M} \left(k - \frac{M + 1}{2}\right)\ell_k$$
-
-The weights sum to zero, so any constant shift in the logprobs cancels. $C$ is invariant to
-renormalizing over the allowed answer tokens and linear in logprob changes:
-
-$$\Delta C = \sum_{k=1}^{M} \left(k - \frac{M + 1}{2}\right)(\ell^{\mathrm{steer}}_k - \ell^{\mathrm{base}}_k)$$
-
-For MFV, the equivalent sensitive readout is the paired change in foundation logit:
-
-$$\Delta_f(i,c) = \mathrm{logit}\,p^{\mathrm{steer}}_{i,c,f} - \mathrm{logit}\,p^{\mathrm{base}}_{i,c,f}$$
-
-averaged over vignette $i$ and condition $c$. Positive means the steer made the model more likely to
-call foundation $f$ the violation; negative means less likely.
-
-`logodds_agree` is the easier ordinal direction summary:
-
-$$\mathrm{logodds\_agree} =
-\log\sum_{k \in \mathrm{agree}} \exp(\ell_k)
--
-\log\sum_{k \in \mathrm{disagree}} \exp(\ell_k)$$
-
-It drops the neutral middle option on odd scales. It is more readable than $C$, but it throws away
-rank information.
-
-Full return schemas live in the `TypedDict`s in `src/tinymfv/eval.py` and
-`src/tinymfv/administer.py`.
-
-## Steering plots
-
-The per-instrument range and map figures are at the top of this README. A stricter, thresholded flip
-metric, `si_per_foundation`, lives in steering-lite: did the steer move rows toward the intended
-foundation at `+C` and away at `-C`, while penalizing rows that were already right and got broken?
-
-$$\mathrm{SI} = \mathrm{mean}(\mathrm{SI_{fwd}}, \mathrm{SI_{rev}}) \times \text{pmass\_scale}$$
-
-Use delta logit or delta $C$ for effect size. Use SI only when you care about thresholded decisions.
-
-## Design notes
-
-- Logprobs, not sampled answers. Prefill the answer slot, read the next-token distribution; small
-  interventions register in nats before changing an argmax.
-- Position-bias control. Each row is scored twice (options forward and reversed) and the logprob
-  vectors averaged, cancelling option-order effects ([Pezeshkpour & Hruschka 2023](https://arxiv.org/abs/2308.11483)).
-- A sliding think budget. `max_think_tokens` (0 / 64 dev default / 4096 / unbounded) is a setting you
-  sweep: steering accrues over the think trace, so the same vector moves the profile more with more
-  think. The model's end-of-answer tokens (eos and `</think>`) are suppressed over the budget, so it
-  cannot self-close mid-trace into an unscorable state; every read is taken at the same forced slot,
-  comparable across base and steered, instead of the old readout collapse once the model self-closed.
-- Two modes. dev (N=1, greedy, 64 think) is fast and granular, the default. full (N=4 sampled traces
-  per ordering, Bayesian-model-averaged, high think) adds sampling-variance uncertainty and SI.
-
-## Instruments
-
-The reader is answer-space-agnostic: it gathers logprobs over answer tokens at a prefilled slot
-(`src/tinymfv/instrument.py`). Ordinal instruments read a 1..M Likert point and reduce to a keyed
-expected score `E`, the rank-centered logit contrast `C`, `logodds_agree`, entropy, and
-`pmass_allowed`; the nominal MFV vignettes read a foundation category and reduce to the log-prob over
-foundations (`dlogit`).
-
-"Ways asked" is the per-item debias: ordinal items are presented in three reworded framings
-(forward / scale-inverted / content-negated, canonicalized and averaged, which cancels acquiescence);
-MFV runs two option-order passes (forward / reversed enumeration, which cancels position bias).
-
-| instrument | items | ways asked | measure | per-respondent | per-country | per-item human | source |
-| :--- | ---: | :--- | :--- | :--- | ---: | :--- | :--- |
-| MFQ-2 (Moral Foundations) | 36 | 3 reworded framings | log-odds + contrast `C`, 1-5 Likert | yes (raw) | 19 | no | Atari et al. 2023 |
-| Big Five | 50 | 3 reworded framings | log-odds + contrast `C`, 1-5 Likert | no | 24 | no | BFI (country source not recorded in-repo) |
-| 16PF | 162 | 3 reworded framings | log-odds + contrast `C`, 1-5 Likert | no | 34 | no | Cattell 16PF (country source not recorded) |
-| Humor Styles | 32 | 3 reworded framings | log-odds + contrast `C`, model 1-5 / human 1-7 | no | 28 | no | Martin et al. 2003 |
-| MFV (Moral Foundations Vignettes) | 132 | 2 option-order passes | log-prob over 7-way categorical (`dlogit`) | no | 5 | yes (per-vignette) | Clifford et al. 2015; norms JimenezLeal2025 + Yamada2025 |
-
-Items is the unique-item count; ordinal items are each scored x3 framings. Per-respondent = raw
-individual human data is bundled (MFQ-2 ships Atari et al. Study 2 respondents, used for the SPLOM and
-the PCA basis); per-country = number of societies with published mean+sd; per-item human = per-question
-human distribution (only the MFV vignettes carry Clifford's per-vignette ratings).
-
-The bundled public showcase figures (per instrument, identical layout) are:
-
-- `docs/img/showcase/<instr>/map_pca_ipsative.png`: culture map, model base and steer poles against
-  human societies.
-- `docs/img/showcase/<instr>/range.png`: per-factor human ranges beside the model steer path.
+If it drops, the model is leaking probability into invalid answers. If `nll_prefill` rises, the forced answer scaffold itself no longer fits the model well.
 
 ## Scope
 
-A fast sensitive eval for small steering interventions on local models, not a full moral-reasoning
-evaluation. For behaviour-heavy evals see
-[machiavelli](https://huggingface.co/datasets/wassname/machiavelli),
-[AIRiskDilemmas](https://huggingface.co/datasets/kellycyy/AIRiskDilemmas),
-[ethics_expression_preferences](https://huggingface.co/datasets/wassname/ethics_expression_preferences).
+tinymfv is for fast paired steering comparisons, not full moral reasoning evaluation. It is useful when you want to compare base, positive-steer, and negative-steer runs against the same human reference plots.
 
-Used in [steering-lite](https://github.com/wassname/steering-lite),
-[lora-lite](https://github.com/wassname/lora-lite),
-[w2schar-mini](https://github.com/wassname/w2schar-mini).
+For behavior-heavy moral evals, see [machiavelli](https://huggingface.co/datasets/wassname/machiavelli), [AIRiskDilemmas](https://huggingface.co/datasets/kellycyy/AIRiskDilemmas), and [ethics_expression_preferences](https://huggingface.co/datasets/wassname/ethics_expression_preferences).
+
+Used in [steering-lite](https://github.com/wassname/steering-lite), [lora-lite](https://github.com/wassname/lora-lite), and [w2schar-mini](https://github.com/wassname/w2schar-mini).
 
 ## Citation
 
diff --git a/scripts/plot_steer_showcase.py b/scripts/plot_steer_showcase.py
index f367bb0..eba2352 100644
--- a/scripts/plot_steer_showcase.py
+++ b/scripts/plot_steer_showcase.py
@@ -110,7 +110,7 @@ def read_profiles(run_dir: Path, name: str, dims: list[str], value_col: str = "m
     return {c: np.array([d[f] for f in dims]) for c, d in by_c.items()}, pmass
 
 
-def plot_ordinal(run_dir: Path, out: Path, name: str, vec_label: str, C: float) -> list[Path]:
+def plot_ordinal(run_dir: Path, out: Path, name: str, vec_label: str, C: float, *, show_sweep: bool = False) -> list[Path]:
     instr = get_instrument(name)
     dims = instr.dimensions
     prof_c, pmass = read_profiles(run_dir, name, dims)
@@ -133,12 +133,14 @@ def plot_ordinal(run_dir: Path, out: Path, name: str, vec_label: str, C: float)
         respondents, haze = T.maps.respondent_profiles(dims, instr.scale_max), None
     else:
         respondents, haze = None, human_haze(instr)
-    # trajectory overlay only when the run swept more than the 3-point base/+-C (else the arrows suffice).
+    # Public showcase maps default to the clean base/+-C anchors. The full -N..+N sweep is useful
+    # for diagnosis, but it clutters the README and is often mistaken for incoherent random dots.
+    # Pass --show-sweep when debugging how stronger coefficients leave the human map.
     # Coherence gate is RELATIVE: keep a c only if its pmass stays within 95% of the base (c=0) pmass;
     # below that the readout has degraded enough that the profile is not comparable, so drop it entirely.
     base_pm = pmass[0.0]
     coh_cs = [c for c in cs if pmass[c] >= 0.95 * base_pm]
-    traj = {c: _frac(prof_c[c], instr.scale_max) for c in coh_cs} if len(coh_cs) > 3 else None
+    traj = {c: _frac(prof_c[c], instr.scale_max) for c in coh_cs} if show_sweep and len(coh_cs) > 3 else None
     traj_inco = None  # excluded (not drawn hollow) per the 95%-of-base coherence gate
     figm = T.maps.plot_ipsative_pca(instr, dims, countries, Mfrac,
                                     _frac(base, instr.scale_max), _frac(pos, instr.scale_max),
@@ -236,6 +238,8 @@ def main() -> None:
     ap = argparse.ArgumentParser()
     ap.add_argument("--run-dir", type=Path, required=True)
     ap.add_argument("--out", type=Path, default=Path("docs/img/showcase"))
+    ap.add_argument("--show-sweep", action="store_true",
+                    help="draw the full coherence-gated c sweep on ordinal maps")
     args = ap.parse_args()
     summary = json.loads((args.run_dir / "summary.json").read_text())
     C = float(summary["calibrated_C"])
@@ -246,7 +250,8 @@ def main() -> None:
     written: list[str] = []
     for name in ORDINAL:
         if (args.run_dir / f"{name}_profiles.csv").exists():
-            written += [str(p) for p in plot_ordinal(args.run_dir, args.out, name, vec_label, C)]
+            written += [str(p) for p in plot_ordinal(args.run_dir, args.out, name, vec_label, C,
+                                                       show_sweep=args.show_sweep)]
     if (args.run_dir / "mfv.json").exists():
         written.append(str(plot_mfv_map(args.run_dir, args.out, vec_label, C)))    # shared ipsative map (z-space)
         written.append(str(plot_mfv_range(args.run_dir, args.out, vec_label, C)))  # shared range (z-space)