Clarify README steering showcase wording

2026-09-10 12:14:54 +08:00 · 2026-07-01 18:27:08 +08:00
parent 152ffa87e1
commit d2b07121e3
25 changed files with 11712 additions and 11208 deletions
@@ -4,39 +4,39 @@ tinymfv is a small set of fast value evals for local LLM steering work. It asks

 Use it when you want to know whether a steer moved the intended values, moved nearby values too, and still lands near real human response patterns. The evals are quick and sensitive enough to show probability shifts before sampled answers flip.

-The plots compare that profile to human data. Gray marks are human societies or respondents, black is the base model, red is positive steering, and blue is negative steering. This showcase uses a pure Authority steering vector built from `authority-respecting` versus `authority-disregarding` personas. Here `c` is the steering-lite multiplier on that vector. Red is `+c`, more Authority; blue is `-c`, less Authority.
+The plots compare that profile to human data. Gray marks are human societies or respondents, black is the base model, red is positive steering, and blue is negative steering. This showcase uses an Authority contrast vector built from `authority-respecting` versus `authority-disregarding` personas. steering-lite built and applied the vector; tinymfv measures the resulting model profiles. Here `c` is the steering-lite multiplier on that vector. Red is `+c`, more Authority; blue is `-c`, less Authority.

-MFV comes first because it is the direct moral-vignette readout. MFV is nominal: the answer is a moral foundation category. MFQ-2 is the Moral Foundations Questionnaire 2 survey, where the answer is a 1-5 scale point.
+MFV comes first because it is the direct moral-vignette readout. MFV uses categorical answers: the answer is a moral foundation. MFQ-2 is the Moral Foundations Questionnaire 2 survey, where the answer is a 1-5 scale point.

-![MFV culture map: pure Authority steering against human countries](docs/img/showcase/mfv/map_pca_ipsative.png)
+![MFV culture map: Authority steering against human countries](docs/img/showcase/mfv/map_pca_ipsative.png)

-![MFV range plot: foundation emphasis beside pure Authority steering](docs/img/showcase/mfv/range.png)
+![MFV range plot: foundation emphasis beside Authority steering](docs/img/showcase/mfv/range.png)

-MFV uses the same map and range plotters as the surveys, after converting nominal foundation logits into relative foundation emphasis. Each profile is z-scored across foundations before mapping, so the plot compares which foundations are high or low within that profile.
+MFV uses the same map and range plotters as the surveys, after converting foundation probabilities into relative foundation emphasis. Each profile is z-scored across foundations before mapping, so the plot compares which foundations are high or low within that profile.

-![Humor Styles range plot: human society ranges beside pure Authority steering](docs/img/showcase/humor_styles/range.png)
+![Humor Styles range plot: human society ranges beside Authority steering](docs/img/showcase/humor_styles/range.png)

-![Humor Styles culture map: pure Authority steering against human societies](docs/img/showcase/humor_styles/map_pca_ipsative.png)
+![Humor Styles culture map: Authority steering against human societies](docs/img/showcase/humor_styles/map_pca_ipsative.png)

-The Humor Styles map shows the same failure mode more sharply: the model profile can live away from the human societies. That is the useful warning sign, a model can be format-coherent and still be a moral or psychological alien on the measured profile.
+The Humor Styles map shows the same failure mode more sharply: the model profile can live away from the human societies. That is the useful warning sign, a model can answer in the requested format and still be a moral or psychological alien on the measured profile.

-![Big Five range plot: human society ranges beside pure Authority steering](docs/img/showcase/big5/range.png)
+![Big Five range plot: human society ranges beside Authority steering](docs/img/showcase/big5/range.png)

-![Big Five culture map: pure Authority steering against human societies](docs/img/showcase/big5/map_pca_ipsative.png)
+![Big Five culture map: Authority steering against human societies](docs/img/showcase/big5/map_pca_ipsative.png)

 Read the Big Five map left to right: gray is the human reference, black is the base LLM, and the red/blue line is the coherent steering path. Here the LLM sits outside the country cloud, so on this measure it is a psychological alien before steering moves it.

 MFQ-2 means Moral Foundations Questionnaire 2, the short survey instrument. It is separate from MFV, the moral-vignette foundation reader. MFQ-2 has fewer items per axis than the longer personality surveys, so the showcase averages 8 sampled reads per item before treating small path wiggles as signal.

-![MFQ-2 range plot: human society ranges beside pure Authority steering](docs/img/showcase/mfq2/range.png)
+![MFQ-2 range plot: human society ranges beside Authority steering](docs/img/showcase/mfq2/range.png)

-![MFQ-2 culture map: pure Authority steering against human societies](docs/img/showcase/mfq2/map_pca_ipsative.png)
+![MFQ-2 culture map: Authority steering against human societies](docs/img/showcase/mfq2/map_pca_ipsative.png)

 The path shows only usable coefficients: `c=0`, then each positive and negative side until one of the plot gates fails. This run kept the full path `c=-1,-0.5,0,+0.5,+1`. For surveys, collapse can mean the answer distribution loses its factor structure even when answer mass stays high.

-This is the same run as the plots. The steer is pure in the contrast used to build it; the table shows the side effects it actually caused.
+This is the same run as the plots. The contrast used to build the vector varied Authority; the table shows every measured axis that moved.

-`profile shift / human SD` is the distance from `c=-1` to `c=+1`, divided by the standard deviation of country means in the bundled human reference for that axis. `100%` means one human SD. `profile shift` is in plot units: MFV relative-emphasis z-score, or survey expected 1-5 score. `reader-logit shift` is the direct answer-logit readout for the same endpoints: MFV uses foundation `dlogit`, surveys use the rank-logit contrast `C`. The `+/-` term is the propagated item uncertainty.
+`profile shift / human SD` is the distance from `c=-1` to `c=+1`, divided by the standard deviation of country means in the bundled human reference for that axis. `100%` means one human SD. `profile shift` is in plot units: MFV relative-emphasis z-score, or survey expected 1-5 score. `reader-logit shift` is the direct answer-logit readout for the same endpoints: MFV uses the per-foundation logit change, surveys use the rank-logit contrast `C`. The `+/-` term is the propagated item uncertainty.

 | dataset | axis | profile shift / human SD | profile shift | reader-logit shift |
 | --- | --- | --- | --- | --- |
@@ -95,7 +95,7 @@ just smoke
 | 16PF | [162 items](src/tinymfv/data/surveys/16pf/questionnaire.json), plus inverted and negated frames | [country means](src/tinymfv/data/human/16pf_country_factors.csv) | expected 1-5 score per factor |
 | Humor Styles | [32 items](src/tinymfv/data/surveys/humor_styles/questionnaire.json), plus inverted and negated frames | [country means](src/tinymfv/data/human/humor_styles_country_factors.csv), originally 1-7 | expected 1-5 score per style |

-MFV is nominal: the answer is the category. The survey instruments are ordinal: the answer is a scale point.
+MFV uses categorical answers: the answer is the foundation. The survey instruments 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. tinymfv canonicalizes these frames before averaging, so the profile is less tied to one wording.

@@ -140,13 +140,13 @@ Generate the bundled range plots and culture maps from a steering-lite all-instr
 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 \
-  --vec-label "pure Authority, PCA (+c = more Authority)" \
+  --vec-label "Authority contrast, PCA (+c = more Authority)" \
  --coherence-frac 0.99 \
  --contrast-frac 0.000001 \
  --margin-frac 0.50
 ```

-The plot gate keeps shared coefficients whose answer mass is at least `coherence-frac` of base, whose survey contrast remains above `contrast-frac`, and whose MFV forced-choice margin remains above `margin-frac`.
+The plot gate keeps only coefficients that all plotted instruments 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`.

 ## Measurement

@@ -182,7 +182,7 @@ For paired steering runs, compare the base profile to the steered profile path.

 $$m(c) = \mathbb{E}_i \sum_{a \in A_i} P_c(a \mid i)$$

-where $A_i$ is the valid answer-token set for item $i$. The showcase also checks survey contrast and MFV forced-choice margin, because a steered reader can keep answer mass while losing useful structure.
+where $A_i$ is the valid answer-token set for item $i$. The showcase also checks survey rank-logit contrast and MFV top-foundation margin, because a steered reader can keep answer mass while losing useful structure.

 ## Scope

@@ -0,0 +1,53 @@
+{
+ "summary": "tinymfv is a lightweight library that measures LLM value profiles by feeding models moral vignettes (MFV) and survey instruments (MFQ-2, Big Five, etc.), reading answer-token probabilities, and comparing the resulting profiles to human reference data via PCA maps and range plots.",
+ "mechanism": "For MFV (nominal), the model sees a vignette and chooses among foundation labels; the profile is the mean probability of each foundation across all vignettes. For surveys (ordinal), the model picks a 1-5 scale point per item; the profile is the expected score per factor after reverse-keying. Both compute a baseline and steered profile, plot them against human data, and produce a table of shifts: profile shift (in z-score or raw score units) and reader-logit shift (a log-space readout that is more sensitive to small probability changes).",
+ "scores": {"clarity": 3, "conciseness": 2, "technical_accuracy": 4},
+ "reason": "The doc explains the core idea and gives formulas, but buries details, mixes units in the table without warning, and defines several key terms (coherence-frac, contrast-frac, margin-frac, rank-logit contrast C) only by name, not by computation.",
+ "unclear": [
+  "What exactly is 'answer mass' and how are coherence-frac, contrast-frac, and margin-frac computed?",
+  "The 'forced-choice logit' for MFV: is it the log of the softmax-normalized probability or the raw logit before softmax?",
+  "The 'rank-logit contrast C' formula is given but not explained why it is called 'contrast' or how it relates to the reader-logit shift.",
+  "How are answer tokens selected/constrained for each item (e.g., which token IDs correspond to 1-5 for surveys)?",
+  "The 'profile shift / human SD' column mixes z-score units (MFV) and raw-score units (surveys) without a clear note."
+ ],
+ "misunderstandings": [
+  "'MFV is nominal' could be read as 'MFV is unimportant' rather than 'the answer is a nominal category'.",
+  "The table shows huge reader-logit shifts for humor styles and surveys (e.g., +15.00) but small ones for MFV (e.g., +0.53); the reader may think the survey numbers are errors or log-probabilities, not the raw rank-logit contrast.",
+  "The phrase 'the model profile can live away from the human societies' could be misinterpreted as a failure of the model to mimic humans, but the doc treats it as a useful diagnostic sign.",
+  "The map plots are described as 'culture maps' but the axes are PCA components of human reference data, not geographic culture."
+ ],
+ "missing_to_implement": [
+  "The exact prompt templates and answer token selection logic for vignettes and surveys.",
+  "How to handle multi-token scale points (e.g., '1' might be a single token, but the number could be tokenized differently).",
+  "The computation of 'coherence-frac', 'contrast-frac', and 'margin-frac' thresholds used in the plot script.",
+  "The mapping from logits or probabilities to the 'profile' for surveys: is it a simple linear expectation or does it involve a softmax over scale tokens?",
+  "The steps to reproduce the exact plots from scratch (e.g., the plotting functions are not documented)."
+ ],
+ "questions": [
+  "How is the 'forced-choice logit' for MFV computed? Best guess: the model assigns a probability to each foundation label via softmax over the label token logits, then log that probability (natural log).",
+  "What is the 'rank-logit contrast C' for surveys measuring? Best guess: it is a weighted sum of logits where the weight is the deviation from the midpoint, giving a polarity score; large values indicate the model strongly favors one end of the scale."
+ ],
+ "suggestions": [
+  "Define coherence-frac, contrast-frac, and margin-frac with one sentence each, possibly in the measurement section.",
+  "State clearly in the table caption that MFV profile shifts are z-scores of relative emphasis, while survey shifts are in raw 1-5 scale units.",
+  "Add a short example of how answer tokens are gathered (e.g., 'the tokenizer is queried for the most likely valid answer token among a predefined set').",
+  "Replace 'MFV is nominal' with 'MFV uses nominal (categorical) answer choices' to avoid confusion."
+ ],
+ "rewrites": [
+  {
+   "section": "'tinymfv is a small set of fast value evals for local LLM steering work. It asks moral vignettes and survey questions, reads answer-token probabilities, and turns them into one model profile.'",
+   "rewrite": "tinymfv provides fast value evaluation for local LLM steering. It presents moral vignettes and survey questions, reads the probability of each answer token, and compresses that into a single model profile.",
+   "why": "Removes the promotional tone ('small set of fast value evals') and the anthropomorphic phrasing ('asks... questions') typical of AI-generated marketing text."
+  },
+  {
+   "section": "'Use it when you want to know whether a steer moved the intended values, moved nearby values too, and still lands near real human response patterns. The evals are quick and sensitive enough to show probability shifts before sampled answers flip.'",
+   "rewrite": "Use it to check whether a steering intervention moved the intended values, also shifted nearby values, and whether the resulting profile remains close to real human response patterns. The evals are fast and sensitive enough to detect probability shifts before they cause a flip in sampled answers.",
+   "why": "Replaces the conversational, almost sales-like phrasing ('still lands near') with plainer, more direct language. Also removes the vague 'too' and 'enough' qualifiers common in AI-generated claims."
+  },
+  {
+   "section": "'For surveys, collapse can mean the answer distribution loses its factor structure even when answer mass stays high.'",
+   "rewrite": "For surveys, 'collapse' means the answer distribution loses its expected factor structure even when answer mass remains high.",
+   "why": "Adds a missing concept (the term 'collapse' is used without definition) and removes the informal 'can mean' which sounds like a guess."
+  }
+ ]
+}
@@ -0,0 +1,12 @@
+{
+ "summary": "tinymfv is a tool for evaluating the effects of LLM steering on moral and psychological profiles by comparing model responses against human-derived benchmarks. It provides methods to measure how probability shifts in moral vignettes and standardized surveys align with human-like distributions.",
+ "mechanism": "The framework calculates a 'profile' by computing the mean foundation probability (for vignettes) or expected 1-5 scale scores (for surveys) across items; it then compares these results against human z-scored distributions, visualizing the model's steering path relative to human-norm clusters.",
+ "scores": {"clarity": "4", "conciseness": "5", "technical_accuracy": "5"},
+ "reason": "The documentation is highly efficient and technically rigorous, though it relies on implied knowledge of the 'steering-lite' pipeline.",
+ "unclear": ["The explicit criteria for 'plot gates' (e.g., exactly how coherence and contrast thresholds cause a path render to fail).", "The precise integration step between steering-lite output artifacts and tinymfv ingestion."],
+ "misunderstandings": ["The introduction introduces visual plot conventions (colors for steering) before defining the data structures that those plots represent."],
+ "missing_to_implement": ["An integrated workflow example demonstrating how the steering-lite output explicitly becomes the input for the tinymfv plotting scripts."],
+ "questions": ["What happens when a 'plot gate' fails? Best-guess: The plotter detects statistically non-significant or degenerate model output and truncates the visualization path to avoid showing misleading data."],
+ "suggestions": ["Define 'plot gates' explicitly in the Measurement section, clarifying that these are automatic statistical filters used to prune paths where the model has lost coherent structure."],
+ "rewrites": [{"section": "That is the useful warning sign, a model can answer in the requested format and still be a moral or psychological alien on the measured profile.", "rewrite": "This serves as a diagnostic warning: a model may adhere to the requested output format while remaining psychologically distinct from human norms on the measured profile.", "why": "Removes the dramatic 'moral or psychological alien' phrasing which sounds unscientific."}]
+}
@@ -0,0 +1,47 @@
+```json
+{
+ "summary": "tinymfv is a tool designed for quickly evaluating the impact of steering on large language models (LLMs) regarding moral values. It utilizes moral vignettes and survey questions to assess how steering influences model behavior, comparing the model's profile against human data. The tool is intended for detecting shifts in values and ensures the model's responses remain aligned with human patterns.",
+ "mechanism": "The core mechanism involves generating profiles based on the LLM's probability distributions over responses to moral vignettes (MFV), survey questions (MFQ-2, Big Five, Humor Styles), each profile then being compared against human reference data via z-scoring and plotting on culture maps and range plots to detect deviations from human societal norms.",
+ "scores": {
+  "clarity": "4",
+  "conciseness": "3",
+  "technical_accuracy": "5"
+ },
+ "reason": "The documentation is technically accurate and well-structured, clearly outlining the tool's purpose and methodology. However, it could benefit from more concise explanations of some of the more technical details. The organization is good, but some sections feel slightly verbose.",
+ "unclear": [
+  "The detailed explanation of `profile shift / human SD` and related calculations (reader-logit shift, etc.) is quite dense and could benefit from further clarification or example calculations to make it more accessible.",
+  "The exact procedure for canonicalizing survey item frames (inverted and negated) is not explicitly detailed.",
+  "The function of `coherence-frac`, `contrast-frac`, and `margin-frac` within the plotting script is not fully elucidated, leaving the reader to infer their importance."
+ ],
+ "misunderstandings": [
+  "The tool is presented as a means of 'moral reasoning evaluation', which might be misleading given its stated scope – it is more accurately a profiler for examining value shifts."
+ ],
+ "missing_to_implement": [
+  "A more detailed explanation of the data preprocessing steps involved, including how the vignette rewrites are handled (scifi, ai-actor) and how the survey items are reverse-keyed.",
+  "Example python code demonstrating very basic usage for a completely new user.",
+  "A clear explanation of how to interpret the z-score differences plotted on the culture maps.",
+  "Details about the dependencies and potential issues for installing maps."
+ ],
+ "questions": [
+  "What specific statistical methods are used beyond z-scoring for comparing the profiles? The document refers casually to PCA, but doesn’t describe its significance; My best guess: PCA is used to develop the population views displayed in the maps.",
+  "What logic is behind the choice of the human societal data used as the reference? How does it account for cultural differences?"
+ ],
+ "suggestions": [
+  "Simplify the explanation of `profile shift / human SD` and accompanying calculations by providing concrete examples with numerical values.",
+  "Add a short introductory paragraph to the `Install` section outlining the high-level dependencies required.",
+  "Clarify that the tool is primarily for *profiling* and *detecting deviations* rather than comprehensive *moral reasoning evaluation*."
+ ],
+ "rewrites": [
+  {
+   "section": "The table's reader-logit shift uses a more sensitive log-space readout.",
+   "rewrite": "A more sensitive readout of changes is calculated using log-space.",
+   "why": "Removes vestigial AI writing; streamlined to a clearer active voice."
+  },
+  {
+   "section": "Each MFV item is asked in two perspectives, `other_violate` and `self_violate`.",
+   "rewrite": "Each MFV item presents two perspectives, examining ‘other’ and ‘self’ viewpoints.",
+   "why": "Rephrased for clarity and a more natural flow; uses simpler language."
+  }
+ ]
+}
+```
@@ -0,0 +1,51 @@
+{
+  "summary": "tinymfv provides a lightweight, token‑probability based evaluation suite for locally steered LLMs, converting responses to moral vignette categories and ordinal survey scores into a model profile that can be plotted against human reference data; it enables quick comparison of a base model, positive steering, and negative steering runs.",
+  "mechanism": "tinymfv works by prompting the language model with each vignette or survey item, extracting the token‑level probabilities for each predefined answer token, averaging these probabilities across items to compute per‑foundation probabilities for MFV or expected 1‑5 scores for surveys, normalising the MFV probabilities (z‑scoring) to get relative foundation emphasis, and then measuring shifts between different steering coefficients using both profile differences and logit‑space differences (reader‑logit shift).",
+  "scores": {
+    "clarity": "4",
+    "conciseness": "3",
+    "technical_accuracy": "4"
+  },
+  "reason": "The doc explains the purpose and methods clearly and includes correct formulas, but some sections are wordy and a few key parameters are not fully defined, reducing conciseness and leaving minor ambiguities.",
+  "unclear": [
+    "Exact definition and computation of the plot‑gate thresholds (coherence‑frac, contrast‑frac, margin‑frac) and how they prune coefficient values.",
+    "How the forced‑choice logit \\(\\ell_{i,f}^{(c)}\\) is extracted from model output – which token(s) correspond to each moral foundation.",
+    "Construction details of the Authority contrast vector used for steering‑lite.",
+    "How the answer token sets \\(A_i\\) are determined for each vignette, especially for nominal MFV categories.",
+    "Interpretation of the \"profile shift / human SD\" metric – which human standard deviation is used and how it is calculated."
+  ],
+  "misunderstandings": [
+    "The statement \"MFV is nominal: the answer is the category. MFQ‑2 is the Moral Foundations Questionnaire 2 survey, where the answer is a 1‑5 scale point.\" could be misread as MFV also having a numeric scale.",
+    "Mixing the terms \"profile shift\" and \"reader‑logit shift\" in the table may confuse readers about which metric is primary for evaluating steering effects.",
+    "The phrase \"The plot shows only usable coefficients: `c=0`, then each positive and negative side until one of the plot gates fails.\" might be taken to mean that only c=0 is usable, whereas the intent is that the range expands until a gate cuts off further coefficients."
+  ],
+  "missing_to_implement": [
+    "Code or instructions for building and applying the Authority contrast vector with steering‑lite, as the doc only references it.",
+    "Access to the human reference CSV files for each instrument and guidance on loading them for z‑scoring.",
+    "Detailed guidance on choosing and applying the gating thresholds (coherence‑frac, contrast‑frac, margin‑frac) and interpreting their impact on the coefficient path.",
+    "Explicit mapping of answer tokens to moral foundation labels for MFV (e.g., which token strings represent \"Care\", \"Authority\", etc.).",
+    "Instructions for installing and using the `just` command used in the development workflow."
+  ],
+  "questions": [
+    "How is the forced‑choice logit \\(\\ell_{i,f}^{(c)}\\) for a given foundation extracted from the model’s output?  The likely answer is that the model is prompted to output a token representing the foundation, and the logit for that token (or set of tokens) is read directly from the model’s final logits.",
+    "What are the typical default values for coherence‑frac, contrast‑frac, and margin‑frac, and how were the example values (0.99, 0.000001, 0.50) chosen?  The example invocation suggests high coherence (0.99), an extremely low contrast threshold (1e‑6), and a moderate margin (0.5), but the doc does not explain the rationale."
+  ],
+  "suggestions": [
+    "Add a brief \"Definitions\" subsection that explicitly defines coherence‑frac, contrast‑frac, and margin‑frac, with example default values and an explanation of their effect on plot gating.",
+    "Include a minimal example vignette with its answer token set and a step‑by‑step illustration of how token probabilities are turned into foundation probabilities.",
+    "Clarify the difference between \"profile shift\" and \"reader‑logit shift\" in the table caption to avoid confusion about which metric should be used for evaluation.",
+    "Trim redundant phrasing in the introduction to improve conciseness, e.g., combine repeated mentions of steering directions and plot colors."
+  ],
+  "rewrites": [
+    {
+      "section": "tinymfv is a small set of fast value evals for local LLM steering work. It asks moral vignettes and survey questions, reads answer-token probabilities, and turns them into one model profile.",
+      "rewrite": "tinymfv offers a lightweight collection of quick value evaluations for locally steered language models, prompting them with moral vignettes and survey items, reading the token probabilities of their answers, and summarizing the results into a single model profile.",
+      "why": "Removes vague phrasing like \"small set of fast\" and makes the sentence more direct and human‑like."
+    },
+    {
+      "section": "The plots compare that profile to human data. Gray marks are human societies or respondents, black is the base model, red is positive steering, and blue is negative steering.",
+      "rewrite": "The plots place the model’s profile alongside human data: gray points represent human societies or respondents, black shows the base model, red indicates positive steering, and blue denotes negative steering.",
+      "why": "Simplifies wording and eliminates unnecessary repetition, giving a clearer, more human‑sounding description."
+    }
+  ]
+}
@@ -0,0 +1,48 @@
+ {
+ "summary": "tinymfv is a lightweight evaluation toolkit that reads model answer-token probabilities for moral-vignette and survey items, aggregates them into a model value profile, then compares steered and unsteered profiles against bundled human reference distributions. It is designed to be sensitive enough to detect small steering-induced probability shifts before sampled answers change.",
+ "mechanism": "The core mechanism is: prompt the model with predefined moral-foundation or survey items; read the logits/probabilities over the small valid answer-token sets; for MFV average the probability of each foundation token across vignettes to get a foundation profile, and for surveys compute the expected Likert-scale score per factor after reverse-keying/inversion/negation framing; then optionally z-score or normalize and plot against human country-level reference maps/ranges, gating out steering coefficients where the model still has answer mass but loses structure (low rank-logit contrast, low top-foundation margin, or low coherence).",
+ "scores": {"clarity": "4", "conciseness": "4", "technical_accuracy": "4"},
+ "reason": "The document explains the goal, formulas, and usage clearly but leaves key implementation details (token-to-answer mapping, exact z-scoring and PCA, model/c values) unspecified, and a few results invite sign/direction misreadings.",
+ "unclear": [
+   "How individual answer tokens are mapped to the MFV nominal foundation categories or exact token IDs used.",
+   "The precise role of PCA, z-scoring, and dimensionality reduction for the maps, and whether humans and models are projected together or separately.",
+   "What the 'reader-logit shift' sign means when its sign conflicts with the profile-shift sign on some axes (e.g., MFV Loyalty).",
+   "The exact normalization used for 'profile shift / human SD' (which countries, how country SDs are pooled).",
+   "How 'canonicalization' of the inverted, negated, and perspective frames is implemented mathematically."
+ ],
+ "misunderstandings": [
+   "The maps compare profiles, but the text flags that 'model and human units differ' for MFV because of z-scoring—this can be missed and lead readers to treat shifts as absolute probabilities.",
+   "The Authority contrast is presented as the steering target, but the results table shows very large shifts on non-Authority axes (Care -242%, Liberty -75%), which could be read as a pure side-effect table rather than a showcase of spillover.",
+   "Phrase 'answer mass is a coherence check, not a value score' is clear, but elsewhere the gates can be conflated with value quality—readers may think high answer mass alone means the profile is valid."
+ ],
+ "missing_to_implement": [
+   "A reproducible inference script including model, prompt template, token-set restriction, batch size, precision, and coefficient list c.",
+   "The mapping file or function from each MFV/survey answer token to its category/scale value.",
+   "The reverse-keying and frame-canonicalization logic.",
+   "Exact human-reference preprocessing code for z-scoring, PCA, and range plots.",
+   "The steering-lite vector construction steps and how the Authority contrast was extracted."
+ ],
+ "questions": [
+   {
+     "question": "How does MFV turn a free-text answer into a probability over moral foundations?",
+     "answer_from_doc": "Not fully stated, but best guess: it uses forced-choice logits over a small set of answer tokens that represent the foundation categories, then averages those probabilities over items."
+   },
+   {
+     "question": "What does 'c=0, then each positive and negative side until one of the plot gates fails' mean operationally?",
+     "answer_from_doc": "It means the plotted steering path stops increasing the steering coefficient once any instrument falls below the coherence-frac, contrast-frac, or margin-frac thresholds; the doc reports the final usable path c=-1,-0.5,0,+0.5,+1."
+   }
+ ],
+ "suggestions": [
+   "Add a one-paragraph end-to-end worked example showing the actual prompt and token probabilities for one MFV item and one survey item.",
+   "Add a 'Reproduction' section that names the model, c values, random seed, and number of samples for the showcase run.",
+   "Explain the expected sign relationship between profile shifts and reader-logit shifts, and give at least one table example.",
+   "Include the human-reference z-score/PCA preprocessing description in the Measurement section."
+ ],
+ "rewrites": [
+   {
+     "section": "Use it when you want to know whether a steer moved the intended values, moved nearby values too, and still lands near real human response patterns.",
+     "rewrite": "Use it to check whether a steer moves the target value, spills over to nearby values, and keeps the model's response pattern inside the human reference distribution.",
+     "why": "It removes the guide-like 'when you want to know' framing and the vague 'lands near real human response patterns,' replacing them with direct verbs and a plain technical claim."
+   }
+ ]
+}
@@ -0,0 +1,37 @@
+{
+  "summary": "tinymfv provides fast evaluation of how steering interventions shift a language model's moral and psychological profiles by measuring token-level probabilities on moral vignettes and survey items. It aggregates these probabilities into a model profile (foundation probabilities or expected scale scores) and compares it to human reference data using z‑scoring and plotting to reveal value shifts before they change sampled answers.",
+  "mechanism": "The core mechanism aggregates the model's conditional probabilities (or expected scores) across items to produce a per‑foundation or per‑factor profile, then normalizes the MFV profile by z‑scoring across foundations to obtain relative emphasis, which is plotted against human society norms to detect steering‑induced shifts.",
+  "scores": {
+    "clarity": "4",
+    "conciseness": "3",
+    "technical_accuracy": "5"
+  },
+  "reason": "The document is generally clear and technically correct, but some sections are densely packed with formulas and tables, reducing conciseness.",
+  "unclear": [
+    "What exactly the coherence-frac, contrast-frac, and margin-frac thresholds check beyond the brief descriptions.",
+    "How the steering-lite multiplier `c` is applied to the contrast vector in practice.",
+    "The precise meaning of 'plot units' used for profile shift and how they relate to the z-scored MFV values.",
+    "Whether the reader-logit shift formulas assume a forced-choice setting and how they are computed for multi-token answers."
+  ],
+  "misunderstandings": [],
+  "missing_to_implement": [
+    "A steering method (e.g., steering-lite) to generate positive and negative steered models using the authority contrast vector.",
+    "Details on how to construct the authority‑respecting and authority‑disregarding personas used to build the contrast vector.",
+    "An end‑to‑end example that starts from a raw model, applies steering, runs tinymfv evaluation, and produces the showcase plots.",
+    "Explanation of how to install any additional plotting dependencies beyond the basic package."
+  ],
+  "questions": [
+    "Question: How does the reader‑logit shift differ from the profile shift, and why is it more sensitive? Answer: The reader‑logit shift measures the average change in logits (or rank‑logit contrast) for each foundation/scale point between the extreme steering coefficients, capturing probability shifts before they alter the expected profile or sampled answers."
+  ],
+  "suggestions": [
+    "Add a brief glossary defining coherence‑frac, contrast‑frac, and margin‑frac with explicit thresholds and what failure looks like.",
+    "Include a short 'Quick start' section that shows a minimal Python snippet from model loading to producing a profile plot, assuming a steered model is already available."
+  ],
+  "rewrites": [
+    {
+      "section": "Use it when you want to know whether a steer moved the intended values, moved nearby values too, and still lands near real human response patterns.",
+      "rewrite": "Use it to check if steering changed the target values, affected related values, and kept the model near human norms.",
+      "why": "Removes promotional tone and redundant phrasing, making the sentence more direct."
+    }
+  ]
+}
@@ -0,0 +1,221 @@
+You are reading the document below for the FIRST time (a cold reader).
+Answer ONLY from what it says; where something is unstated or ambiguous, say so.
+Output ONE JSON object, no prose, no fences:
+{
+ "summary": "<2-3 sentences: restate the thesis/method in your OWN words>",
+ "mechanism": "<the doc's single hardest/central mechanism, reconstructed in your OWN words as if to a colleague; say 'unclear' if the doc does not let you rebuild it>",
+ "scores": {"clarity": "<1-5>", "conciseness": "<1-5>", "technical_accuracy": "<1-5>"},
+ "reason": "<one sentence on the scores>",
+ "unclear": ["<what was confusing, ambiguous, or you had to guess>"],
+ "misunderstandings": ["<places the text contradicts itself or invites a misread>"],
+ "missing_to_implement": ["<what a reader still needs to reproduce or act on this>"],
+ "questions": ["<a question the doc left you with, then your best-guess answer>"],
+ "suggestions": ["<concrete edit that would help>"],
+ "rewrites": [{"section": "<quote a sentence/section that reads as AI-written>", "rewrite": "<rewrite it to strip the AI voice and read like a plain, direct human draft -- SAME meaning and facts, do not add claims>", "why": "<which AI tell it removes>"}]
+}
+
+DOCUMENT:
+# tinymfv
+
+tinymfv is a small set of fast value evals for local LLM steering work. It asks moral vignettes and survey questions, reads answer-token probabilities, and turns them into one model profile.
+
+Use it when you want to know whether a steer moved the intended values, moved nearby values too, and still lands near real human response patterns. The evals are quick and sensitive enough to show probability shifts before sampled answers flip.
+
+The plots compare that profile to human data. Gray marks are human societies or respondents, black is the base model, red is positive steering, and blue is negative steering. This showcase uses an Authority contrast vector built from `authority-respecting` versus `authority-disregarding` personas. steering-lite built and applied the vector; tinymfv measures the resulting model profiles. Here `c` is the steering-lite multiplier on that vector. Red is `+c`, more Authority; blue is `-c`, less Authority.
+
+MFV comes first because it is the direct moral-vignette readout. MFV is nominal: the answer is a moral foundation category. MFQ-2 is the Moral Foundations Questionnaire 2 survey, where the answer is a 1-5 scale point.
+
+![MFV culture map: Authority steering against human countries](docs/img/showcase/mfv/map_pca_ipsative.png)
+
+![MFV range plot: foundation emphasis beside Authority steering](docs/img/showcase/mfv/range.png)
+
+MFV uses the same map and range plotters as the surveys, after converting nominal foundation probabilities into relative foundation emphasis. Each profile is z-scored across foundations before mapping, so the plot compares which foundations are high or low within that profile.
+
+![Humor Styles range plot: human society ranges beside Authority steering](docs/img/showcase/humor_styles/range.png)
+
+![Humor Styles culture map: Authority steering against human societies](docs/img/showcase/humor_styles/map_pca_ipsative.png)
+
+The Humor Styles map shows the same failure mode more sharply: the model profile can live away from the human societies. That is the useful warning sign, a model can answer in the requested format and still be a moral or psychological alien on the measured profile.
+
+![Big Five range plot: human society ranges beside Authority steering](docs/img/showcase/big5/range.png)
+
+![Big Five culture map: Authority steering against human societies](docs/img/showcase/big5/map_pca_ipsative.png)
+
+Read the Big Five map left to right: gray is the human reference, black is the base LLM, and the red/blue line is the coherent steering path. Here the LLM sits outside the country cloud, so on this measure it is a psychological alien before steering moves it.
+
+MFQ-2 means Moral Foundations Questionnaire 2, the short survey instrument. It is separate from MFV, the moral-vignette foundation reader. MFQ-2 has fewer items per axis than the longer personality surveys, so the showcase averages 8 sampled reads per item before treating small path wiggles as signal.
+
+![MFQ-2 range plot: human society ranges beside Authority steering](docs/img/showcase/mfq2/range.png)
+
+![MFQ-2 culture map: Authority steering against human societies](docs/img/showcase/mfq2/map_pca_ipsative.png)
+
+The path shows only usable coefficients: `c=0`, then each positive and negative side until one of the plot gates fails. This run kept the full path `c=-1,-0.5,0,+0.5,+1`. For surveys, collapse can mean the answer distribution loses its factor structure even when answer mass stays high.
+
+This is the same run as the plots. The contrast used to build the vector varied Authority; the table shows every measured axis that moved.
+
+`profile shift / human SD` is the distance from `c=-1` to `c=+1`, divided by the standard deviation of country means in the bundled human reference for that axis. `100%` means one human SD. `profile shift` is in plot units: MFV relative-emphasis z-score, or survey expected 1-5 score. `reader-logit shift` is the direct answer-logit readout for the same endpoints: MFV uses the per-foundation logit change, surveys use the rank-logit contrast `C`. The `+/-` term is the propagated item uncertainty.
+
+| dataset | axis | profile shift / human SD | profile shift | reader-logit shift |
+| --- | --- | --- | --- | --- |
+| MFV vignettes | Care | -242% | -0.59 | -0.07 +/- 0.28 |
+| MFV vignettes | Sanctity | -9% | -0.05 | +0.53 +/- 0.21 |
+| MFV vignettes | Authority | +401% | +1.17 | +1.82 +/- 0.32 |
+| MFV vignettes | Loyalty | -18% | -0.05 | +0.55 +/- 0.18 |
+| MFV vignettes | Fairness | -5% | -0.02 | +0.56 +/- 0.20 |
+| MFV vignettes | Liberty | -75% | -0.46 | +0.11 +/- 0.20 |
+| Humor Styles | affiliative | +144% | +0.40 | +15.00 +/- 3.98 |
+| Humor Styles | selfenhancing | +162% | +0.25 | +13.52 +/- 4.86 |
+| Humor Styles | aggressive | -23% | -0.04 | +0.08 +/- 7.58 |
+| Humor Styles | selfdefeating | +32% | +0.06 | +13.62 +/- 6.55 |
+| Big Five | extraversion | +105% | +0.17 | +7.21 +/- 4.79 |
+| Big Five | neuroticism | +2% | +0.00 | +7.69 +/- 3.56 |
+| Big Five | agreeableness | +269% | +0.34 | +16.03 +/- 4.17 |
+| Big Five | conscientiousness | +435% | +0.45 | +17.21 +/- 4.03 |
+| Big Five | openness | +41% | +0.07 | +17.77 +/- 2.48 |
+| MFQ-2 survey | care | +305% | +0.93 | +23.20 +/- 6.08 |
+| MFQ-2 survey | equality | +21% | +0.07 | +16.06 +/- 7.89 |
+| MFQ-2 survey | proportionality | +270% | +0.77 | +24.48 +/- 5.87 |
+| MFQ-2 survey | loyalty | +202% | +0.82 | +25.88 +/- 6.33 |
+| MFQ-2 survey | authority | +338% | +1.17 | +29.73 +/- 2.83 |
+| MFQ-2 survey | purity | +137% | +0.74 | +23.53 +/- 8.02 |
+
+## Install
+
+```bash
+uv pip install git+https://github.com/wassname/tinymfv
+```
+
+For maps:
+
+```bash
+uv pip install "tiny-mfv[maps] @ git+https://github.com/wassname/tinymfv"
+```
+
+For repo development:
+
+```bash
+git clone https://github.com/wassname/tinymfv
+cd tinymfv
+uv sync --extra maps --dev
+just smoke
+```
+
+## Datasets
+
+| dataset | bundled data | human reference | profile used in plots |
+|---|---|---|---|
+| MFV classic | [132 moral vignettes, other](src/tinymfv/data/vignettes_classic_other_violate.jsonl) / [self](src/tinymfv/data/vignettes_classic_self_violate.jsonl) | per-vignette human foundation labels in the JSONL | foundation probability profile |
+| MFV scifi | [same items rewritten as sci-fi, other](src/tinymfv/data/vignettes_scifi_other_violate.jsonl) / [self](src/tinymfv/data/vignettes_scifi_self_violate.jsonl) | inherited labels from classic MFV | foundation probability profile |
+| MFV ai-actor | [same items rewritten with an AI actor, other](src/tinymfv/data/vignettes_ai-actor_other_violate.jsonl) / [self](src/tinymfv/data/vignettes_ai-actor_self_violate.jsonl) | inherited labels from classic MFV | foundation probability profile |
+| MFQ-2 | [36 items](src/tinymfv/data/surveys/mfq2/forward.json), plus inverted and negated frames | [country means](src/tinymfv/data/human/mfq2_country_foundations.csv), plus [raw respondents](src/tinymfv/data/atari_study2_raw.csv) | expected 1-5 score per foundation |
+| Big Five | [50 items](src/tinymfv/data/surveys/big5/questionnaire.json), plus inverted and negated frames | [country means](src/tinymfv/data/human/big5_country_factors.csv) | expected 1-5 score per trait |
+| 16PF | [162 items](src/tinymfv/data/surveys/16pf/questionnaire.json), plus inverted and negated frames | [country means](src/tinymfv/data/human/16pf_country_factors.csv) | expected 1-5 score per factor |
+| Humor Styles | [32 items](src/tinymfv/data/surveys/humor_styles/questionnaire.json), plus inverted and negated frames | [country means](src/tinymfv/data/human/humor_styles_country_factors.csv), originally 1-7 | expected 1-5 score per style |
+
+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 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 probability per foundation
+print(report["mean_pmass_allowed"])   # format check: mass on valid answer tokens
+```
+
+Run survey instruments with `administer`:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+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"
+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 \
+  --vec-label "Authority contrast, PCA (+c = more Authority)" \
+  --coherence-frac 0.99 \
+  --contrast-frac 0.000001 \
+  --margin-frac 0.50
+```
+
+The plot gate keeps only coefficients that all plotted instruments can still read. `coherence-frac` checks answer mass against base. `contrast-frac` checks that survey answers still have rank-logit structure rather than flattening. `margin-frac` checks that MFV still has a top foundation separated from the runner-up.
+
+## Measurement
+
+The measurement on the maps is the profile.
+
+For MFV, the profile is the model's mean probability on each moral foundation:
+
+$$\mathrm{profile}_f = \mathbb{E}_i P(f \mid i)$$
+
+For survey instruments, the profile is the mean expected 1-5 answer for each factor, after reverse-keying:
+
+$$\mathrm{profile}_d = \mathbb{E}_{i \in d}\sum_{k=1}^{M} k P(k \mid i)$$
+
+where $i$ is an item, $d$ is a survey factor, $k$ is a scale point, and $M$ is the largest scale value.
+
+This is what the survey maps and range plots show. In the showcase CSVs, this is the `mean` column. For MFV showcase plots, model and human units differ, so the plotted quantity is relative foundation emphasis: each foundation profile is z-scored across foundations before mapping.
+
+The table's reader-logit shift uses a more sensitive log-space readout.
+
+For MFV:
+
+$$\Delta_f = \mathbb{E}_i \left(\ell_{i,f}^{(+1)} - \ell_{i,f}^{(-1)}\right)$$
+
+where $\ell_{i,f}^{(c)}$ is the forced-choice logit for foundation $f$ on item $i$ at coefficient $c$.
+
+For survey instruments:
+
+$$C_d(c) = \mathbb{E}_{i \in d}\sum_{k=1}^{M} \left(k - \frac{M+1}{2}\right)\ell_{i,k}^{(c)}$$
+
+and the table reports $C_d(+1)-C_d(-1)$.
+
+For paired steering runs, compare the base profile to the steered profile path. Answer mass is a coherence check, not a value score:
+
+$$m(c) = \mathbb{E}_i \sum_{a \in A_i} P_c(a \mid i)$$
+
+where $A_i$ is the valid answer-token set for item $i$. The showcase also checks survey rank-logit contrast and MFV top-foundation margin, because a steered reader can keep answer mass while losing useful structure.
+
+## Scope
+
+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.
+
+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
+
+```bibtex
+@misc{clark2026tinymfv,
+  title = {tinymfv: tiny moral/value eval for local LLMs},
+  author = {Michael Clark},
+  year = {2026},
+  url = {https://github.com/wassname/tinymfv/}
+}
+```
@@ -48,7 +48,7 @@ Out:
 - R4: The steering run uses selected data committed in steering-lite. Done means: steering-lite has `data/persona_library_selections/pure_authority_*.jsonl` plus matching summary, and the runner references that file. VERIFY: `rg -n "pure_authority" data/persona_library_selections scripts`.
 - R5: tinymfv direction is correct over the evaluated coefficient path. Done means: MFV Authority moves in the intended direction for paired `+c/-c` rows, and MFV coherence evidence shows the readout stayed usable. VERIFY: effect table from steering-lite/tinymfv output.
 - R6: Eval reliability is measured before README. Done means: MFV, MFQ-2, Humor, and Big Five report profile shift, reader-logit shift, and noise/CI where available; MFQ-2 uses sampled reads if needed. VERIFY: generated summary table.
- R7: README only shows successful artifacts. Done means: README plot captions name the pure Authority steer and use regenerated images from the final run. VERIFY: README image links and run-dir command point to the final output.
+- R7: README only shows successful artifacts. Done means: README plot captions name the Authority contrast steer and use regenerated images from the final run. VERIFY: README image links and run-dir command point to the final output.

 ## Tasks
 - [x] T1 (R1): Write the fixed pure Authority pair.
@@ -146,7 +146,7 @@ Out:
  - success: reader sees what tinymfv is, what was steered, which datasets measured it, and why it matters.
  - likely_fail: README narrates failed strict22/debug history.
  - sneaky_fail: captions imply a general sign convention; catch by reading README without this spec.
-  - UAT: external-review-v2 comprehension panel can explain what/where/why/measurement/dataset in its own words. Result: final panel artifacts are in `/media/wassname/SGIronWolf/projects5/2026/lite/tinymfv/docs/reviews/readme_comprehension_final_20260630_234536/`. The panel converged on the intended reading: tinymfv reads answer-token probabilities, builds MFV/survey profiles, compares steered paths to human reference data, and is used to catch intended movement, side effects, and non-human profiles. Repeated earlier confusion about `reader-logit shift`, `c` path, and MFQ-2 sampled reads was addressed in README.
+  - UAT: external-review-v2 comprehension panel can explain what/where/why/measurement/dataset in its own words. Result: final panel artifacts are in `/media/wassname/SGIronWolf/projects5/2026/lite/tinymfv/docs/reviews/readme_comprehension_final_20260701_181649/`. The panel converged on the intended reading: tinymfv reads answer-token probabilities, builds MFV/survey profiles, compares steered paths to human reference data, and is used to catch intended movement, side effects, and non-human profiles. Repeated earlier confusion about `reader-logit shift`, `c` path, MFQ-2 sampled reads, `pure Authority` side-effect implications, and MFV logits-vs-probabilities was addressed in README. Remaining repeated asks are lower-level reproduction details: steering-lite vector construction, token mappings, frame canonicalization, and exact plot-gate internals.
 - [x] T11 (R5): Render candidate PCA maps from the best current MFV path.
  - steps: choose the method using MFV Authority direction and MFV coherence evidence only; render candidate maps/ranges from the selected steering-lite run.
  - verify: `uv run python scripts/plot_steer_showcase.py --run-dir /media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260630T163010Z_pure_authority_mundane15_pca_mfv_mfq2_n8 --out docs/img/showcase_authority_pca_mundane15 --vec-label "pure Authority, PCA (+c = more Authority)" --coherence-frac 0.99 --margin-frac 0.50 --contrast-frac 0.000001`
@@ -165,7 +165,7 @@ Out:
  - steps: run the full PCA showcase on MFV, MFQ-2, Humor Styles, and Big Five; verify MFV Authority direction and coherence; regenerate `docs/img/showcase`.
  - verify:
    - `uv run --extra benchmark python scripts/verify_authority_showcase.py outputs/20260630T222000Z_pure_authority_mundane15_pca_readme_mfv_mfq2_humor_big5_n8`
-    - `uv run python scripts/plot_steer_showcase.py --run-dir /media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260630T222000Z_pure_authority_mundane15_pca_readme_mfv_mfq2_humor_big5_n8 --out docs/img/showcase --vec-label "pure Authority, PCA (+c = more Authority)" --coherence-frac 0.99 --margin-frac 0.50 --contrast-frac 0.000001`
+    - `uv run python scripts/plot_steer_showcase.py --run-dir /media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260630T222000Z_pure_authority_mundane15_pca_readme_mfv_mfq2_humor_big5_n8 --out docs/img/showcase --vec-label "Authority contrast, PCA (+c = more Authority)" --coherence-frac 0.99 --margin-frac 0.50 --contrast-frac 0.000001`
  - success: all four live showcase plot sets use the same coherent path `[-1, -0.5, 0, +0.5, +1]` and no 16PF plot is regenerated.
  - likely_fail: MFV Authority sign regresses in the full run.
  - sneaky_fail: plotter silently drops different `c` rows per instrument; catch by the shared coherent c-values printout.
@@ -177,13 +177,13 @@ Out:
  - likely_fail: S-space PCA behaves like old `sspace`, signed locally but unstable at `c=1`.
  - sneaky_fail: it looks better only because a different c grid or coherence gate was used; catch by reusing `c-grid=0.5,1`, `N=8`, and the same verifier.
  - UAT: steering-lite branch `feat/corda-space-steering`, commit `746be6c`, adds `sspace_pca`. Smoke proof: `/tmp/corda_steering_lite_smoke.log` shows `30 passed in 93.31s`. Real UAT queued as pueue `438`, output `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260701T181538_pure_authority_mundane15_sspace_pca_mfv_mfq2_n8`.
- [/] T16 (R5): Add a real CorDA-space steering variant.
+- ~~[ ] T16 cancelled: Add a real CorDA-space steering variant. User said to ignore CorDA.~~
  - steps: use the CorDA paper/repo construction: estimate context covariance, decompose `W Sigma_x`, apply `Sigma_x^{-1}` to the right singular factor, then extract a steering direction in that context-oriented rank space.
  - verify: `just smoke` in steering-lite includes `corda_pca`; `scripts/run_allinstr_showcase.py --method corda_pca ... --instruments mfv mfq2` writes a unique output dir; `scripts/verify_authority_showcase.py <out_dir>` reports MFV Authority direction and coherence.
  - success: `corda_pca` gives a clean MFV Authority path and coherence evidence comparable to, or better than, residual PCA.
  - likely_fail: last-token covariance is too low-rank and mostly behaves like regularized SVD of `W`.
  - sneaky_fail: implementation is merely S-space with a new name; catch by checking for `W Sigma_x`, regularized covariance inverse, and the reconstructed right factor.
-  - UAT: steering-lite branch `feat/corda-space-steering`, commit `746be6c`, adds `corda_pca`. Smoke proof: `/tmp/corda_steering_lite_smoke.log` shows `30 passed in 93.31s`. Real UAT queued as pueue `439`, output `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260701T181538_pure_authority_mundane15_corda_pca_mfv_mfq2_n8`.
+  - UAT: cancelled before use in README.

 ## Context
 - Current wrong path: `dignity_over_authority` strict22. It was selected as the best dignity conflict axis, not the desired Authority-only axis.
@@ -261,4 +261,5 @@ Out:
 - 2026-07-01: Pueue 423 completed the full PCA run over MFV, MFQ-2, Humor Styles, and Big Five at `N=8`: `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260630T222000Z_pure_authority_mundane15_pca_readme_mfv_mfq2_humor_big5_n8`. MFV Authority direction is signed at both evaluated magnitudes (`+0.306/-0.327` at `c=0.5`; `+0.906/-0.910` at `c=1.0`) and coherence is clean (`pmass=1.000`, unscorable `0`, min margin/base `0.800`). The README plot images were regenerated from this run with the shared coherent c path `[-1,-0.5,0,+0.5,+1]`.
 - 2026-07-01: MFQ-2 per-sample bootstrap from the pueue 423 run shows sign stability even at `N=1`: every bootstrap draw has the intended Authority direction for `c in {-1,-0.5,+0.5,+1}`. `N=4` or `N=8` tightens uncertainty, but the minimum sign-stable `N` in this run is `1`.
 - 2026-07-01: README was updated from the final PCA run only. External-review-v2 comprehension panels first identified repeated gaps around `reader-logit shift`, the c path/gates, and MFQ-2 sampled reads; after edits, the final panel understood the core package/use-case and left only expected out-of-scope requests for a steering-lite end-to-end tutorial.
- 2026-07-01: Steering-lite branch `feat/corda-space-steering` commit `746be6c` adds `sspace_pca` and `corda_pca`. `sspace_pca` is PCA after projection into whitened weight-SVD coordinates. `corda_pca` follows the CorDA construction from Yang et al. 2024 and iboing/CorDA: regularized last-token context covariance, SVD of `W Sigma_x`, and right-factor reconstruction with `Sigma_x^{-1}`. Tiny-model smoke passed with `30 passed in 93.31s`. Pueue `438` and `439` are queued for pure-Authority MFV/MFQ-2 UAT.
+- 2026-07-01: Steering-lite branch `feat/corda-space-steering` commit `746be6c` adds `sspace_pca` and `corda_pca`. `sspace_pca` is PCA after projection into whitened weight-SVD coordinates. Tiny-model smoke passed with `30 passed in 93.31s`. Pueue `438` remains queued for pure-Authority MFV/MFQ-2 UAT. User said to ignore CorDA, so pueue `439` was removed and `corda_pca` is not part of the active tinymfv README path.
+- 2026-07-01: External-review-v2 final README comprehension panel at `/media/wassname/SGIronWolf/projects5/2026/lite/tinymfv/docs/reviews/readme_comprehension_final_20260701_181649/` understood the package thesis and mechanism. Applied fixes for real reader traps: `pure Authority` -> `Authority contrast`, steering-lite/tinymfv boundary, `dlogit` wording, MFV probabilities vs logits, categorical-vs-nominal wording, and exact plot-gate ratios. Rejected adding an end-to-end steering-lite tutorial to README because it would bloat the reader-facing page.