docs/img/README.md is a captioned gallery: each figure states what the dots are, the
named axes and their source paper, the readout (rated sampling N=12 for the frontier
panel, logprobs N=8 for the steering showcase), and for the steer the base model
(Qwen3-4B), the Authority PCA vector, and +c/-c directions. Regenerate all maps with the
edge-inset + roomier margins.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The four pole signposts (Adaptive / Self-directed ...) draw on the crosshair edges after
label placement, so a zone name near the frame collided with them (African-Islamic vs
Adaptive on the humor map). Feed each pole in as a short keep-out band of obstacle points.
Regenerate committed WVS + showcase maps.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Region labels maximised clearance, which fled to the FARTHEST open space -- the polygon
name ended up floating far from its hull. Instead take the nearest perimeter ring that
has a clear slot (emptiest spot on that ring), so the name hugs its own hull edge. Marker
spacing beyond the glyph is now anisotropic (spacing_x=0.5, spacing_y=0.3): text stacks
tighter vertically, so pull labels in more up-down than left-right. Marker clearance pad
is unscaled, so nothing lands on its own star. Re-render WVS + showcase.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The per-instrument value maps (big5/humor/mfq2) now route labels through
labelplace.allocate_labels: zone names in clear air (no white boxes), country
labels adjacent, steer path unchanged. Ipsative/range figures re-emitted from the
same run for a consistent set.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Fresh-eyes QA caught the "East Asia" zone label hugging its (central, crowded) red hull
edge. Region labels only searched 0-0.9 text-heights off the perimeter; widen to 3.0 so
the clearance search can escape a surrounded hull into open space.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The 17-model IW map now renders through labelplace.allocate_labels; save the artifact
into the repo (png for viewing, svg for hand-editing labels, CI table as companion).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
plot_value_map now takes a steer= dict and draws the steer as a connected
black->red(+c)/blue(-c) path (matching the ipsative map's trajectory) instead
of three disconnected model dots. draw_zone_hulls returns its label anchors so
textalloc routes country/model labels around zone names. Regenerated the
showcase (mfq2/big5/humor value maps are new; mfv ipsative picks up Netherlands).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Russian validation of the Knutson/Kruepke Realistic Moral Vignettes
(norm violation / social affect / intention), sent by the authors.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Per feedback the individual-respondent contours filled the frame (within >>
between variance). Now each IW zone is a ~1.6-sigma covariance ellipse over its
member country-mean dots, with an eigenvalue floor so 1-2 country zones get a
visible blob instead of a dot/line (fixes big5 SG/PK orphans). PCA now fits on the
country means M so the axes are between-country and zones separate. mfq2/big5/mfv
read cleanly; humor still overlaps (real negative result: humor country profiles
don't cluster the IW way).
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Probe + artifact: the WVS subset of Anthropic/llm_global_opinions is 353 questions
over 90 countries (212 questions with >=40 countries), matching tinymfv's MC +
human-anchor shape and dense enough for an Economist-scale map. Documents the
selections parse recipe and the open axis-definition fork (literal IW 10-question
factor model vs shared-question ipsative PCA) before model-run compute.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
read_api.read_items_sampled samples N chat completions at temperature and uses the
empirical answer frequency as the per-item categorical p, emitting the same row
shape the logprob reader does -- so E/profile/entropy flow through the identical
per_item_categorical + reducers and a frontier model without logprobs drops onto
the same map. pmass_allowed becomes the parse rate (sampling coherence gate); C/LO
are omitted by design (log of a frequency has -inf zeros). This is the Economist's
'average of ten responses' method.
UAT (docs/reviews/p3_api_sampling_uat.md): E_mc == E_logprob to <=0.01 (unbiased),
llama-3.1-8b sampled E lands on the same [1,5] scale, parse-rate gate flags
off-format draws.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
respondent_profiles now returns countries alongside profiles; plot_ipsative_pca
gains respondent_zones -> a p90 Gaussian ellipse per IW zone of the projected
Atari respondent cloud (edge-only, no scipy). mfq2 uses these real-respondent
ellipses; the other instruments keep country-mean hulls. Journal notes the
finding: individual profiles overlap across cultures (within >> between variance),
so only the country-mean hull reproduces the Economist's clean zone blobs.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
