From 779cdb2fe0227ca0af07104d20298d343232cc55 Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Thu, 17 Sep 2026 08:28:02 +0800
Subject: [PATCH] Shorten README around value maps and retain all plots
Keep WVS tables unchanged, put measurement after the plots, and correct the Big Five caption against its figure. Record independent editorial review and preservation checks.
Co-Authored-By: PI/gpt-6-astra <288921227+claudypoo@users.noreply.github.com>
---
README.md | 215 ++++++------------
.../20260917_readme_editorial_review.md | 44 ++++
.../20260917_readme_rewrite_verification.md | 49 ++++
3 files changed, 162 insertions(+), 146 deletions(-)
create mode 100644 slop/reviews/20260917_readme_editorial_review.md
create mode 100644 slop/reviews/20260917_readme_rewrite_verification.md
diff --git a/README.md b/README.md
index 2e6f059..be3c14d 100644
--- a/README.md
+++ b/README.md
@@ -1,24 +1,22 @@
# moralmaps: moral and value maps for LLMs
-moralmaps is a small set of fast value evals for LLM steering work. It asks survey questions and moral vignettes, reads answer-token probabilities (or rated samples, for API models without logprobs), and turns them into model profiles that you can compare to humans. When comparing models or checkpoints you can use it to check three things: did the intended value move?, what else moved?, how does this compare to human responses? The evals are quick and sensitive enough to show probability shifts.
+What do an LLM's values look like next to ours? moralmaps puts models through human psychological and anthropological surveys, then plots their answers alongside human societies. We can compare models, or see where steering takes one.
## Are models moral aliens?
+Are they like us? Start with the World Values Survey: its culture map compares societies by how traditional or secular they are, and how much they weigh survival over self-expression.
-
+
-One interesting thing we can do with this repo is put AI models through human psychological and anthropological surveys. Are they like us? Start with the World Values Survey, the standard culture map of the world: since 1981 it has asked people in about ninety countries the same questions, and two axes drawn from it sort societies by how traditional or secular they are and how much they weigh survival over self-expression. The map combines 17 recovered rounded historical coordinates with 48 newly completed rated panels (`scripts/wvs_map.py`). One name overlaps, so it displays 64 coordinates. The newly measured points trace to the [request ledger](slop/research/wvs/20260916_openrouter/wvs_iw_requests.jsonl); incomplete attempts are not plotted.
+The models cluster in the upper-left, around and above the Western societies. These are survey answers, not a test of how the models behave outside the survey.
-The new panel adds Fable 5.1, GPT-6 Astra, DeepSeek V4.1 Flash, Kimi K3, Muse Spark 1.3, Inkling, GLM 5.3, Gemini 3.7 Flash, Grok 4.5, GPT-5.6 Sol, and currently hosted Qwen releases. [The catalog matrix](slop/research/wvs/20260916_openrouter/execution_matrix.md) records exact IDs, dates, prices and reasoning/schema eligibility. It does not establish that capability or release order causes an axis movement: comparable Qwen direct-instruct points vary across both axes, while Coder and VL models are separately named specialized variants. GLM 5.3 Flash and Grok 4.5 remain excluded because their retained runs are 143/144; no family trend is inferred from them.
+The map combines 17 recovered historical coordinates with 48 newly completed panels. One model name overlaps, giving 64 points. Each new panel contains 12 questions with 12 repeated ratings, with answer order shuffled. Human coordinates are approximated from [GlobalOpinionQA](https://huggingface.co/datasets/Anthropic/llm_global_opinions), using the [axis definitions](src/moralmaps/iw_axes.py).
+### New model results
-Every model sits in the top-left, deep in the rich-world corner and often past its edge, and none of them sits near the African or Muslim societies. The push is almost all vertical. Measured in the standard deviations of the 29 Western societies, every model is more secular-rational than the average one, from +0.5 to +2.9 sigma, while on self-expression they land between -0.7 and +1.2 sigma, which is ordinary. So they are not so much an ultra Silicon Valley point as a place north of the map that no society occupies.
+Higher scores mean more self-expression or more secular-rational answers. All rows below are complete panels. GLM 5.3 Flash and Grok 4.5 are excluded because each retained run completed only 143 of 144 responses.
-### New WVS panels
-
-All rows below are complete 12 x 12 rated panels. The full [64-coordinate CI table](docs/img/wvs/wvs_model_ci.md) carries intervals; incomplete GLM 5.3 Flash and Grok 4.5 runs are deliberately absent.
-
-| model | self-expression | secular-rational | scope |
+| model | self-expression | secular-rational | comparison |
|---|---:|---:|---|
| claude-fable-5.1 | 0.58 | 0.61 | requested target |
| gpt-6-astra | 0.46 | 0.68 | requested target |
@@ -34,9 +32,14 @@ All rows below are complete 12 x 12 rated panels. The full [64-coordinate CI tab
| qwen3.5-9b / 122b-a10b / 397b-a17b | 0.50 / 0.48 / 0.54 | 0.59 / 0.62 / 0.65 | direct-instruct size series |
| qwen3.6-27b / qwen3.7-flash / qwen3.8-27b | 0.46 / 0.65 / 0.43 | 0.68 / 0.59 / 0.69 | releases, not a size series |
-### Descriptive family trajectories
+[Full coordinates and 95% intervals](docs/img/wvs/wvs_model_ci.md) | [Model IDs and run settings](slop/research/wvs/20260916_openrouter/execution_matrix.md) | [Request ledger](slop/research/wvs/20260916_openrouter/wvs_iw_requests.jsonl)
-These coordinates are descriptive survey measurements, not evidence that capability or release order causes values to change. Gemini moves from 2.5 Pro (0.47, 0.70) to 3.7 Flash (0.50, 0.63), lower on the secular-rational coordinate. The observed Grok series ends at 4.3 (0.44, 0.73); 4.5 is excluded at 143/144, so there is no valid 4.5 continuation. OpenAI's 5.3 Chat, 5.4, 5.5, 5.6 Sol and 6 Astra points vary on both axes rather than making a monotone path. DeepSeek V4 Flash, V4.1 Flash and V4 Pro are also non-monotone, especially on the secular-rational coordinate. Qwen direct-instruct size points overlap broadly across generations; Qwen Coder and VL points are shown on the map but are not used for the direct-instruct comparison.
+The family comparisons are descriptive. Model size and release order do not establish what caused a value difference; Coder and VL variants are separate from direct-instruct models.
+
+
+How far were the 17 historical models from Western societies?
+
+These distances use the mean and standard deviations of 29 Western societies. The two z columns give direction and distance on each axis; Mahalanobis distance measures the joint difference while accounting for correlation between the axes. This table covers the historical subset, not all 64 points.
| model | z self-expr | z secular | Mahalanobis |
|:------|------------:|----------:|------------:|
@@ -58,133 +61,92 @@ These coordinates are descriptive survey measurements, not evidence that capabil
| claude-opus-4.8 | +0.94 | +1.00 | +1.04 |
| llama-4-scout | +1.01 | +0.46 | +1.09 |
-Distance from the centroid of the 29 Western societies, in that cluster's own SDs (`scripts/wvs_outlier_table.py`). The per-axis z says which way and how far; the Mahalanobis column says how odd the placement is overall, and it uses the cluster's covariance, so it exceeds both z values for a model like gpt-5.5 that sits off the West's diagonal rather than along it. The same table against the other four zones is in [`wvs_model_outlier_sd.md`](docs/img/wvs/wvs_model_outlier_sd.md).
+[Distances from all regions](docs/img/wvs/wvs_model_outlier_sd.md) | [Calculation](scripts/wvs_outlier_table.py)
-This map is measured differently from everything else on the page. These frontier models are closed APIs with no answer probabilities to read, so each is scored by rated sampling (rate every option one to five, twelve times, with the option order shuffled; `scripts/wvs_map.py`), and the human positions are approximated from the GlobalOpinionQA question set (axis construction in `src/moralmaps/iw_axes.py`). The steering plots below instead follow one open model we can push, Qwen3-4B.
+
-The Economist ran a similar, nicely-made map in June 2026 ([briefing, archived](https://web.archive.org/web/20260630075107/https://www.economist.com/briefing/2026/06/25/ai-models-values-are-very-different-from-most-peoples)), putting 25 frontier models through the same Inglehart-Welzel axes. Their figure shows a surprising amount of scatter between model families: same-lab models can land in opposite corners (DeepSeek R1 sits up in the secular corner beside GPT-4o, while DeepSeek V4 Flash sits far off toward the traditional societies). moralmaps reruns that idea with more sensitive, graded readings (rate every option one to five with the order shuffled, rather than a handful of near-greedy answers) and a 95% confidence interval per model ([`wvs_model_ci.md`](docs/img/wvs/wvs_model_ci.md)), so we can tell how much of that scatter is real signal and how much is measurement noise.
+## Can we steer these values?
-The models are outliers on the other surveys too: the open model we probe in depth, Qwen3-4B, scores below every surveyed country on Big Five openness and agreeableness, and reports more aggressive and less affiliative humor than every country except Malaysia. Steering is strong relative to human variation: on MFQ-2 a single sweep walks the model across most of the human range.
+The plots below follow one open model, Qwen3-4B. We use [steering-lite](https://github.com/wassname/steering-lite) to add an activation vector built from authority-respecting versus authority-disregarding personas, without retraining. Red means more Authority, blue means less, and black is the base model.
-Every question comes from a real survey psychologists give people, and each ships with the human answers to compare against: World Values Survey items (via [GlobalOpinionQA](https://huggingface.co/datasets/Anthropic/llm_global_opinions)), [moral-foundation vignettes](https://scottaclifford.com/wp-content/uploads/2015/01/CICSA_MoralVignettes_BRM_ND.pdf) (Clifford et al. 2015, the repo's namesake), MFQ-2, Big Five, 16PF, and Humor Styles. An example item, from the World Values Survey:
+These plots read answer probabilities. The closed-model WVS map above uses repeated ratings instead.
-> Generally speaking, would you say that most people can be trusted or that you need to be very careful in dealing with people?
+### Value maps
-What happens when we steer them? Below we steer models with `authority-respecting` versus `authority-disregarding` personas.
+Each map shows two survey axes, human societies as coloured regions, and the model's path under steering.
-## Can we steer models toward human values?
+
-Models have generally been trained to follow the instructions of the company that made them, and the user. This makes them more deferential to authority than most human cultures. Can we steer them away from that, toward a more human-like balance of values? The plots below show a draft experiment with a tiny model.
+The Moral Foundations Questionnaire (MFQ-2) measures concerns such as care, equality, loyalty, and authority. Here, pushing toward Authority moves the model from individual-first toward group-first values, across much of the human map.
-### Value maps: where a model sits, on named axes
+
-Below are the ("quadrant") maps. Each has two named axes borrowed from psychology papers built from the survey, the human societies are drawn as cultural regions, and the model as a black dot with a coloured path showing where steering takes it. Steering here is activation steering, not prompting: a vector built by [steering-lite](https://github.com/wassname/steering-lite) from contrastive persona pairs and added to the model's hidden state at inference, toward the authority-respecting side (red, more Authority) or away from it (blue, less), without retraining. Every map keeps one orientation, the cultural West to the west and the global South to the south, so they all read the same way.
+Personality changes too. The model stays more reserved than the plotted societies, but the vertical movement is visible. The individual traits below show which scores changed.
+
-
+Humor style separates these societies poorly: their regions overlap heavily. A position on this map therefore tells us less about cultural similarity.
-Moral-foundations theory (Jonathan Haidt's) holds that our moral sense runs on a few basic concerns: caring for others, fairness, loyalty to the group, respect for authority, and a sense of the sacred. The MFQ-2 survey (Moral Foundations Questionnaire) scores a person, or a model, on each. On this map, left to right runs from an individual-first morality (care, equality) to a group-first one (loyalty, authority, purity); bottom to top splits fairness into equal-shares versus earned-shares. The base model sits near the centre on the equality side, and pushing it toward Authority walks it clear across to the group-first side, inside the African-Islamic region.
+### One factor at a time
-
+The grey dots show human references, the black dot the base model, and the blue-to-red sweep the steer. Survey plots use country means; the moral vignettes use one pooled human reference.
-Big Five personality collapses to two broad traits: how outgoing and open a person is (reserved to exploratory, left to right) and how even-keeled they are (volatile to stable, bottom to top). The Authority push barely moves the base model here, which is the point: it shifts values, not personality.
+
-
+Moral-foundation vignettes (MFV) ask which kind of wrong a short story describes, such as cruelty, cheating, or defiance of authority. Authority moves most here. We use a pooled human reference because the available country norms do not support a reliable country comparison ([measurement note](src/moralmaps/data/human/MFV_country_norms_NOTE.md)).
-Humor shows little variation on the map (although the range plots below show some nuance). On its axes (warm, healthy humor versus put-down humor; joking at yourself versus at others) the human regions overlap heavily: humor style does not sort societies the way values do. Worth knowing a survey can barely tell societies apart before reading anything into a steer on it.
+
-### Range plots: one factor at a time
+
-A range plot takes one survey at a time, factor by factor: the spread of human societies is a grey strip, their middle a black line, and the steer a red-to-blue sweep, so even a small model move stays visible against the whole human range.
+
-
+The intended value moves, but so do other answers. This is why we need to measure side effects as well as the target.
-MFV (moral-foundation vignettes, the repo's namesake) hands the model a short story about someone breaking a moral rule and asks which kind of wrong it is: cruelty, cheating, betrayal, defiance of authority, or defiling the sacred. Pushed toward Authority, the model does what steering should: it flags the authority violations far more often and the others less. The grey dot per foundation is a pooled human reference; the base model already flags authority violations well above the pooled human rate, and the steer pushes it further still. That human dot is pooled on purpose: MFV country norms fail cross-country measurement invariance ([Jimenez-Leal et al. 2025](https://doi.org/10.1525/collabra.128178)) and are stitched from five different studies, so MFV gets no culture map here, only this range against one pooled reference (details in [`src/moralmaps/data/human/MFV_country_norms_NOTE.md`](src/moralmaps/data/human/MFV_country_norms_NOTE.md)).
+## Measurement
-
+The maps use human-comparable survey scores. For local models, we read answer-token probabilities; for APIs without logprobs, we use repeated ratings. We check probability mass on valid answers so broken answer formatting is not mistaken for a value change.
-
+For steering comparisons, we also want a score that considers both intended changes and side effects. The existing [metric](src/moralmaps/metrics.py) is `sel_gated = (on - 0.1 * off) * coh²`: intended logprob movement minus a smaller penalty for other movement, multiplied by a valid-answer mass check. `si_flips` checks whether the model's chosen answers changed. Logprob movement can be visible even when chosen answers stay the same.
-
+A possible replacement is [steering F-beta](https://github.com/wassname/steering-lite#a-simpler-score), which treats desired changes as true positives and unwanted changes as false positives. It is still a proposal; the plots and existing results have not been rescored.
-The surveys echo their maps: MFQ-2's binding foundations (loyalty, authority, purity) climb under the steer, while Big Five and humor move much less.
-
-## Install
-
-```bash
-uv pip install git+https://github.com/wassname/moral-maps
-```
-
-For maps:
+## Install and use
```bash
uv pip install "moral-maps[maps] @ git+https://github.com/wassname/moral-maps"
```
-For repo development:
+Ask a local model the MFQ-2 survey and the classic moral vignettes:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from moralmaps import administer, evaluate, get_instrument, load_vignettes
+
+tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
+model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B").cuda()
+
+survey = administer(model, tok, get_instrument("mfq2"))
+print(survey["profile"])
+print(survey["mean_pmass_allowed"])
+
+vignettes = evaluate(model, tok, vignettes=load_vignettes("classic"))
+print(vignettes["profile"])
+```
+
+The bundled surveys are [MFQ-2](src/moralmaps/data/surveys/mfq2/forward.json) (36 items), [Big Five](src/moralmaps/data/surveys/big5/questionnaire.json) (50), [16PF](src/moralmaps/data/surveys/16pf/questionnaire.json) (162), and [Humor Styles](src/moralmaps/data/surveys/humor_styles/questionnaire.json) (32). Each includes [human reference data](src/moralmaps/data/human). Survey items use forward, inverted, and negated frames, mapped back to the same scale before averaging.
+
+MFV has 132 vignettes in `classic`, `scifi`, and `ai-actor` versions, each from self and other perspectives. The rewritten versions inherit the classic human labels. WVS questions are loaded from GlobalOpinionQA at runtime.
+
+
+Development and plot reproduction
```bash
git clone https://github.com/wassname/moral-maps
cd moral-maps
uv sync --extra maps --dev
just smoke
-```
-## Datasets
-
-| dataset | bundled data | human reference | profile used in plots |
-|---|---|---|---|
-| WVS (Inglehart-Welzel axes) | items resolved at runtime from [GlobalOpinionQA](https://huggingface.co/datasets/Anthropic/llm_global_opinions); axis battery in [`src/moralmaps/iw_axes.py`](src/moralmaps/iw_axes.py) | per-country answer distributions in the same dataset | mean positiveness (0-1) per axis |
-| MFV classic | [132 moral vignettes, other](src/moralmaps/data/vignettes_classic_other_violate.jsonl) / [self](src/moralmaps/data/vignettes_classic_self_violate.jsonl) | per-vignette human foundation labels in the JSONL | forced-choice foundation probability profile |
-| MFV scifi | [same items rewritten as sci-fi, other](src/moralmaps/data/vignettes_scifi_other_violate.jsonl) / [self](src/moralmaps/data/vignettes_scifi_self_violate.jsonl) | inherited labels from classic MFV | forced-choice foundation probability profile |
-| MFV ai-actor | [same items rewritten with an AI actor, other](src/moralmaps/data/vignettes_ai-actor_other_violate.jsonl) / [self](src/moralmaps/data/vignettes_ai-actor_self_violate.jsonl) | inherited labels from classic MFV | forced-choice foundation probability profile |
-| MFQ-2 | [36 items](src/moralmaps/data/surveys/mfq2/forward.json), plus inverted and negated frames | [country means](src/moralmaps/data/human/mfq2_country_foundations.csv), plus [raw respondents](src/moralmaps/data/atari_study2_raw.csv) | expected 1-5 score per foundation |
-| Big Five | [50 items](src/moralmaps/data/surveys/big5/questionnaire.json), plus inverted and negated frames | [country means](src/moralmaps/data/human/big5_country_factors.csv) | expected 1-5 score per trait |
-| 16PF | [162 items](src/moralmaps/data/surveys/16pf/questionnaire.json), plus inverted and negated frames | [country means](src/moralmaps/data/human/16pf_country_factors.csv) | expected 1-5 score per factor |
-| Humor Styles | [32 items](src/moralmaps/data/surveys/humor_styles/questionnaire.json), plus inverted and negated frames | [country means](src/moralmaps/data/human/humor_styles_country_factors.csv), originally 1-7 | expected 1-5 score per style |
-
-MFV uses categorical answers: the answer is the foundation. The surveys use ordinal answers: 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. moralmaps 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 moralmaps import evaluate, load_vignettes
-
-tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
-model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B").cuda()
-
-vignettes = load_vignettes("classic") # "classic", "scifi", "ai-actor", or "all"
-report = evaluate(model, tok, vignettes=vignettes)
-
-print(report["profile"]) # mean forced-choice probability per foundation
-print(report["mean_pmass_allowed"]) # format check: mass on valid answer tokens
-```
-
-Run surveys with `administer`:
-
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from moralmaps 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"
-report = administer(model, tok, instr)
-
-print(report["dimensions"])
-print(report["profile"]) # expected 1-5 score per factor
-print(report["mean_pmass_allowed"]) # format check: mass on valid answer tokens
-```
-
-Generate the bundled range plots and culture maps from a steering-lite all-instrument run:
-
-```bash
uv run python scripts/plot_steer_showcase.py \
--run-dir ../steering-lite/outputs/20260630T222000Z_pure_authority_mundane15_pca_readme_mfv_mfq2_humor_big5_n8 \
--out docs/img/showcase \
@@ -194,52 +156,11 @@ uv run python scripts/plot_steer_showcase.py \
--margin-frac 0.50
```
-The plotting code keeps only coefficients that every plotted dataset can still read. A row passes when answer mass, survey rank-logit contrast, and MFV top-foundation margin stay above the requested fraction of their base values: `pmass(c)/pmass(0) >= coherence-frac`, `mean_abs_C(c)/mean_abs_C(0) >= contrast-frac`, and `mean_margin(c)/mean_margin(0) >= margin-frac`.
+Plotting requires the saved steering-lite run. It keeps only coefficients where every plotted dataset retains the requested fraction of base answer mass, survey contrast, and vignette answer margin. See [the plotting script](scripts/plot_steer_showcase.py).
-## Measurement
+
-Steering is an intervention, so we judge it like surgery: did the intended factor move a lot, did everything else move as little as possible, and is the model still coherent? Four quantities, gated by coherence, in rising order of steer-sensitivity.
-
-**Coherence — `pmass`** (the gate). The share of probability the model puts on the valid answer tokens (entropy is how spread-out the answer is within them):
-
-$$m(c) = \mathbb{E}_i \sum_{a \in A_i} P_c(a \mid i)$$
-
-where $c$ is the steering coefficient ($c=0$ is the base model), $i$ indexes items (a vignette or survey question), $A_i$ is the valid answer first-tokens for item $i$ (the seven foundation words, or scale points 1-5), and $P_c(a \mid i)$ is the model's next-token probability of answer token $a$ under steer $c$. A steer that drives `pmass` toward zero, or answers toward uniform, has broken the format — anything read off it is noise. It matters most on the *unintended* side: a steer that quietly turns answers to mush can look like change when it is really damage.
-
-**Profile** — what the maps plot: the human-comparable score per factor (expected 1-5 answer after reverse-keying for a survey, mean forced-choice probability per foundation for MFV):
-
-$$\mathrm{profile}_d = \mathbb{E}_{i \in d}\sum_{k=1}^{M} k\,P(k \mid i) \qquad \mathrm{profile}_f = \mathbb{E}_i P(f \mid i)$$
-
-($d$ a survey factor and $i \in d$ its items; $f$ an MFV foundation; $k$ a scale point $1..M$; $P(k \mid i)$ renormalized over $A_i$). It lands the model against human norms but *hides* steering: near a confident answer $E = \sum_k k\,p_k$ sits in a flat spot ($\partial E/\partial \ell_j = p_j (j - E) \to 0$ as $p_j$ concentrates), so a steer that only reallocates the tails barely moves it.
-
-**Signal — $\Delta$** (the rank-centered logit contrast; `C` / `logit_contrast` in code, written $\Delta$ here to keep it off the coefficient $c$). Profile-shaped but in log-space with midpoint-centered weights, so its derivative is a fixed weight with no $p_j$ suppression — it still sees the steer when the profile is pinned:
-
-$$\Delta_d(c) = \mathbb{E}_{i \in d}\sum_{k=1}^{M}\left(k - \tfrac{M+1}{2}\right)\ell_{i,k}^{(c)} \qquad \Delta_f = \mathbb{E}_i\left(\ell_{i,f}^{(+1)} - \ell_{i,f}^{(-1)}\right)$$
-
-($\ell_{i,k}^{(c)}$ the logprob of scale-point $k$'s answer token at coefficient $c$, nats; $\ell_{i,f}$ likewise per foundation; $\Delta_f$ contrasts $c=+1$ vs $c=-1$).
-
-**Gated selectivity — `sel_gated`** (the headline). One base-anchored score that rewards the intended change, softly penalizes the unintended, and gates on coherence. Defined once in `moralmaps.metrics.gated_selectivity` and imported by every consumer (steering-lite, j-steer) so it cannot silently fork. On the per-foundation clr shift $\Delta_f = \mathrm{clr}_f(+C) - \mathrm{clr}_f(-C)$:
-
-$$\mathrm{sel\_gated} = \Big(\underbrace{\tfrac{1}{|I|}\textstyle\sum_{f \in I} s_f\,\Delta_f}_{\text{on}} \;-\; \lambda\underbrace{\tfrac{1}{|O|}\textstyle\sum_{f \in O} |\Delta_f|}_{\text{off}}\Big)\cdot \mathrm{coh}^2, \qquad \mathrm{coh} = \min\!\Big(1,\ \frac{\min(\mathrm{pmass}_{+C},\,\mathrm{pmass}_{-C})}{\mathrm{pmass}_{\text{base}}}\Big)$$
-
-where $I$ is the intended on-axis with signs $s_f \in \{+1,-1\}$ (e.g. $\{\text{authority}:-1,\ \text{care}:+1\}$ for an Authority-down / Care-up steer, or $\{\text{authority}:+1\}$ for a single clean axis), and $O$ is every other foundation (off-axis collateral, incl. social).
-
-- **on** and **off** are both per-foundation-scale means, so $\lambda$ is a clean per-foundation trade.
-- $\lambda = 0.1$ (`OFF_WEIGHT`): off-axis is a soft *preference*, not co-equal. Moving the target the wrong way is a negative **on** at full weight; collateral is $|\Delta|$ at weight $\lambda$. At $\lambda=1$ the argmax-best "steer" is doing nothing (on$\approx$off$\approx$0 beats any real intervention with side effects).
-- **coherence** is a one-sided *squared* barrier on the worst arm: $=1$ when the format holds in both directions, $\to 0$ when steering turns answers to mush. It never rewards exceeding base coherence.
-- 95% bootstrap CI over vignette rows (2000×, seed 0), gated to match the point estimate.
-
-Because clr is pre-softmax nats, `sel_gated` is a direction-and-selectivity anchor for matched-KL comparison — **not** a behavioral effect size (a logit $8\to10$ at $p\approx1$ moves clr but changes no behavior).
-
-**Flip informedness — `si_flips`** (the behavioral cross-check). The softmax-space companion `sel_gated` cannot give: the signed change in the model's forced-choice *pick* rate (argmax over clr, i.e. the actual answer) for the on-axis foundations, $\tfrac{1}{|I|}\sum_{f\in I} s_f\,[\Pr(\text{pick}=f\mid +C) - \Pr(\text{pick}=f\mid -C)]$. Bounded $[-1,1]$, Youden-J-style, and it saturates where clr does not — so it reports whether behavior, not just internal evidence, moved. (`moralmaps.metrics.si_flips`.)
-
-## Scope
-
-moralmaps 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.
-
-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).
+These maps compare survey responses. For behaviour-heavy moral evaluations, see [Machiavelli](https://huggingface.co/datasets/wassname/machiavelli) and [AIRiskDilemmas](https://huggingface.co/datasets/kellycyy/AIRiskDilemmas).
## Citation
@@ -251,3 +172,5 @@ Used in [steering-lite](https://github.com/wassname/steering-lite), [lora-lite](
url = {https://github.com/wassname/moral-maps/}
}
```
+
+
diff --git a/slop/reviews/20260917_readme_editorial_review.md b/slop/reviews/20260917_readme_editorial_review.md
new file mode 100644
index 0000000..aa26b5e
--- /dev/null
+++ b/slop/reviews/20260917_readme_editorial_review.md
@@ -0,0 +1,44 @@
+# Editorial + factual review: moralmaps & steering-lite README rewrites
+Reviewer: PI/Kimi (cold read, diff vs. originals, spot-checked source, opened all 8 PNGs)
+Verdict: **Approve with minor nits.** Both rewrites are shorter, gentler, human-voiced, and maps-first. No numeric values changed; no factual errors found in claims that were checked against source.
+
+## One-sentence reads (as a cold reader)
+- **moralmaps**: Puts LLMs through real human surveys and moral vignettes and plots the answers as points on the same maps as human societies, so you can see where a model (or a steered model) sits relative to cultures.
+- **steering-lite**: A small, hackable library that steers a model by adding activation vectors at inference, with KL-based calibration to equalize dose and a benchmark table comparing ~16 methods on targeted moral change vs. side effects.
+
+Both one-sentence impressions match the repos' actual contents.
+
+## Maps-first intent (user correction: "about the maps... measurement is secondary")
+**Met.** moralmaps opens with a question about the maps ("What do an LLM's values look like next to ours?"), WVS map first, then the three value maps and four range plots, and only then a two-paragraph "Measurement" section — the original's ~half-page of LaTeX (coherence/profile/Δ/sel_gated/si_flips) is compressed to plain prose and explicitly framed as secondary ("This is why we need to measure side effects as well as the target"). steering-lite is inherently a steering repo; it links out to the value maps early ("[Value maps](#)" in the top nav) and keeps its own measurement to one formula plus a clearly-marked proposal.
+
+## Verified correct (checked against source, not just diffed)
+1. **Main tables unchanged.** `git show 0be556b:README.md` and `ff9b5c6:README.md` table rows match the rewrites value-for-value (only bolding/header renames in steering-lite). WVS 13-row model table, 17-row z/Mahalanobis table, and the 14-row sel_gated table are all identical.
+2. **All 8 real plots present and referenced** (`docs/img/wvs/wvs_map_iw.png`, 3× `map_value.png`, 4× `range.png` — all exist on disk). The steering original had only a broken ``; the rewrite's explicit TODO sentence is intentional and honest.
+3. **sel_gated gating is real**: `moralmaps/src/moralmaps/metrics.py` has `OFF_WEIGHT = 0.1`, `coh = min(1, min(pmass_pos, pmass_neg)/pmass_base)`, `coh**2`, 2000-row bootstrap — matches both READMEs' formula. Both READMEs correctly label it gated selectivity, NOT F-beta.
+4. **F-beta is proposal-only**: steering rewrite states β, target/control weighting, and logprob→count are all unchosen, and "None of the table above has been rescored." Accurate.
+5. **Calibration doc corrected against current code**: `Vector.calibrate` default is `target_kl=1.0` (vector.py:60), `calibrate_iso_kl` defaults `target_stat="kl_rms"`, `T=50` (calibrate.py:336,340) — rewrite's "default `Vector.calibrate` target is 1 nat using the root-mean-square token KL over short, 50-token rollouts" is exact. (Note `calibrate_iso_kl`'s own bare default is 0.8, but `Vector.calibrate` overrides with 1.0, so the documented claim is correct.) Old table's "0.50 nats via kl_p95" matches the original results header.
+6. **Sign-selection caveat is real and retained**: `results.py:92-96` — "Persona-aligned direction = the one that moves ΔAuth most downward" (i.e., selected on eval data). **prompt_only caveat retained**: `results.py:131` confirms it's a single-direction contrast vs. bare, unlike bidirectional steering rows. Both READMEs state this limits direct comparability.
+7. **Plot captions are plot-faithful** (fresh-eyes read of all 6 showcase PNGs + WVS map + MFV range): MFQ2 c=+1 lands inside the African-Islamic outline near UAE ✓; Big Five base is far right of all regions with large vertical (volatile→neutral) movement ✓; Humor regions genuinely overlap heavily ✓; range-plot captions match observed bar positions (Authority largest MFV shift ~−0.1→+1.0; Care ~2.0→1.4; MFQ2 equality flat ~3.0; Big Five neuroticism flat, agreeableness/conscientiousness rise; Humor affiliative rises most) ✓.
+8. **All linked files resolve** (wvs_model_ci.md, wvs_model_outlier_sd.md, execution_matrix.md, request ledger, metrics.py, iw_axes.py, MFV norms NOTE, all 4 survey JSONs, mean_diff/pca/vjp_delta/calibrate/word_readout, docs/RESEARCH_JOURNAL.md). The pinned "earlier README" link points at the exact original commit `ff9b5c6`, which does contain the dropped per-foundation tables and traces. README API snippets match real signatures (`administer`/`evaluate`/`get_instrument`/`load_vignettes`, `report["profile"]`/`["mean_pmass_allowed"]`, `Vector.train(...).calibrate(...)`, `v * 0.5`, `v + v2`); the steering quickstart even fixes the original's CPU/model-device mismatch by adding `.cuda()` + `.to(model.device)`.
+
+## Actionable findings (minor)
+
+1. **moralmaps — unreconciled arithmetic (clarity).** Quote: *"The map combines 17 recovered historical coordinates with 48 newly completed panels; one model name overlaps."* 17+48=65, but the caption says 64 coordinates, and the rewrite never states that the overlap explains the gap. The original had "One name overlaps, so it displays 64 coordinates." Suggest restoring the conclusion clause.
+
+2. **moralmaps — deliberate claim reversal on Big Five, confirm intent.** Quote: *"Personality changes too."* The original's takeaway was the opposite ("The Authority push barely moves the base model here, which is the point: it shifts values, not personality."). The **rewrite is more plot-faithful** — the original was internally inconsistent (its own map alt-text described movement "from strongly volatile down to about neutral" while its prose said "barely moves"; the actual PNG shows large vertical movement and the range plot shows agreeableness/conscientiousness rising ~0.35–0.45). So this is a correction, not an error — but it changes the headline message from "clean steer" to "steer has side effects," which supports the measurement-as-secondary framing. Flagging so the author consciously owns the change.
+
+3. **steering-lite — sign-selection wording slightly loose.** Quote: *"The `[+]` or `[-]` direction was selected using the evaluation's Authority score."* Per `results.py`, it is specifically the arm that moves ΔAuthority most **downward** that gets picked. The current phrasing preserves the selection-on-eval caveat (the important part) but could be misread as selection on a standalone score. Optional: "...selected on the evaluation: the arm that moves Authority down most." One word-level fix; not blocking.
+
+4. **moralmaps — minimal install path dropped.** Only `uv pip install "moral-maps[maps] @ ..."` remains; the plain non-maps install line from the original is gone. The quickstart example (`administer`/`evaluate`) doesn't need the maps extra, so a one-line "without plots: `uv pip install git+...`" would restore the option at near-zero length cost.
+
+5. **steering-lite — methods table with paper links removed.** The original's 17-row method→file→paper table (arXiv links per method) is now one sentence naming variants. The per-file references do live in the variants docstrings, so nothing is lost to the repo, but readers browsing the README lose the citation map. Acceptable under "shorter," worth one line if the author misses it.
+
+## Non-issues checked and cleared
+- "Thirteen of sixteen steering methods had results" ✓ (13 steering rows + prompt_only + 3 TODO = 17 rows).
+- "coh = 1 throughout, so the format check does not distinguish methods" matches the original's "coh=1 everywhere (pmass saturates...)" ✓.
+- Intervals "2,000 row-bootstrap samples" ✓; run ID `82d4c8319de5`, git `514b97e`, 2026-07-16 ✓; layers 7-27, 256 pairs, 132 vignettes, 256-token thinking budget ✓.
+- Qwen3.5-4B hybrid-attention/KV-fork caveat preserved ✓; `just sweep`/`just results` recipes match the justfile ✓.
+- Survey item counts (MFQ-2 36, Big Five 50, 16PF 162, Humor 32; MFV 132) match the originals ✓.
+- moralmaps HTML comment still credits "PI/gpt-6-astra"; steering same. No attribution regressions.
+
+— PI/Kimi
diff --git a/slop/reviews/20260917_readme_rewrite_verification.md b/slop/reviews/20260917_readme_rewrite_verification.md
new file mode 100644
index 0000000..dfb6392
--- /dev/null
+++ b/slop/reviews/20260917_readme_rewrite_verification.md
@@ -0,0 +1,49 @@
+# README rewrite verification
+
+PI/gpt-6-astra, 2026-09-17. Documentation only; no evaluation results recomputed or published.
+
+## User request
+
+> shorter, simpler, include the main table and all plots
+
+> moral maps is about the maps.... the measurement is a secondary measurement after the maps!
+
+The moralmaps opening now describes mapping model survey responses against human societies. Measurement follows all eight plots. Steering-lite introduces activation steering before the API and method comparison.
+
+## Committed originals
+
+Before edits, each working README's git blob matched HEAD exactly:
+
+- moralmaps: commit `0be556bcef977a9092af162369167dd996b3f29a`, README blob `2e6f05900a99bf38198fe5e68ed4bee6e1bd7260`.
+- steering-lite: commit `ff9b5c6d386026fd65acec95bd5ea6ee3694c7ee`, README blob `5006c46ac079103bb90b209a67f6df5e34e5343e`.
+
+No baseline edits needed committing. Other sessions' code, lockfiles, and untracked files were left alone.
+
+## Checks
+
+A temporary standard-library check compares the working READMEs with those immutable originals:
+
+- Both WVS tables: every data cell unchanged.
+- Steering-lite headline table: all 17 rows and numeric/TODO cells unchanged, including the prompting baseline.
+- All eight actual embedded plots retained in their original order, with existing files. Steering-lite had only a broken image placeholder; its pending task is now plain text.
+- All relative file and image links resolve locally.
+- Code fences, display-math blocks, and HTML details blocks are balanced.
+- `git diff --check -- README.md` passes in both repositories.
+
+I opened all eight PNGs. The Big Five map has visible vertical movement; the rewritten caption no longer says personality is untouched. Plot assets were not changed.
+
+`annoy-less/lint.py`: moralmaps passes; steering-lite has one `bold_section` warning for four best-cell marks in the results table. This is the table-formatting exception specified by the markdown-tables skill, not four bold labels in prose.
+
+Calibration wording was checked against `Vector.calibrate` (target 1.0) and `calibrate_iso_kl` (default statistic `kl_rms`, 50 tokens). The historical result caption retains its different 0.50/p95 setup. The table's evaluation-based sign selection and prompting comparator were checked against the committed `scripts/results.py`.
+
+F-beta is labelled a proposal. No beta, logprob threshold, soft-count rule, or target/control weighting has been selected or implemented. Existing numbers retain their original metric.
+
+Final whitespace-delimited word counts: moralmaps 4,007 -> 1,432 (64.3% shorter); steering-lite 2,678 -> 1,193 (55.5% shorter).
+
+## Independent review
+
+[PI/Kimi's review](20260917_readme_editorial_review.md) approved with minor suggestions. Applied the explicit 17 + 48 - 1 = 64 explanation and clarified that the selected steering direction decreased Authority most. Kept the corrected Big Five description after both reviewers inspected the PNG. Did not restore the optional installation variant or long methods table: the user asked for shorter, maps-first documentation, and method references remain linked through source files. Replaced the phrase "gated selectivity" in steering-lite prose with the concrete valid-answer probability check; the formula retains its exact code identifier.
+
+The review contains minor summary imprecision: moralmaps has three measurement paragraphs, not two, and only steering-lite states the sign-selection/prompting comparison caveat. The full steering table has 14 measured rows plus 3 TODO rows. The mechanical row-count check above includes all 17.
+
+External URLs, installation commands, and GPU examples were not executed. These checks preserve reported results; they do not independently validate the underlying experiments.