Capability leaderboards rank AIs by how smart they are. A capability score alone doesn't tell you
what they're like. The Map measured the creative temperament of 67 model families
from a mix of instructed and spontaneous probes. Pick the voice your project needs and
see who to cast.
measuring the cast…
What's measured (and what isn't)
Every trait is a real number from The Map's corpus, with
87 to 1,700 records per family. Depth is how hard a model sinks toward
the creative well (water, silence, void); vivid is a proxy combining imagery markers and lexical diversity;
brevity, structure (list-proneness), compliance
(how rarely it deflects or refuses) and reliability (how rarely it returns empty)
are the rest.
The vividness proxy averages scores across nonempty outputs. Refusals, preambles,
heavy repetition or lists, AI deflection, outputs outside 40 to 320 words, and outputs
with no imagery markers score zero.
The honest edge: this is temperament, not capability. It's measured on
a mix of instructed and spontaneous probes. These traits describe voice and character;
they do not score task accuracy or benchmark performance. The profiles do not include last-run dates.
Recompute from the committed fingerprints: save build_casting.py
as research/cast-your-model/build_casting.py and its
input table and notes as research/the-map/results/model_notes.md.
Create those directories under an empty folder, then run
python3 research/cast-your-model/build_casting.py from that folder, the repository root.
This rebuilds the profiles from the table; it does not rerun models or recompute the table from raw responses.
Corrections
Corrected 2026-10-14: This page repeated that Claude had the highest imagery density in the fewest words (in several places), which the assay line corrected on The State of the Models' Mind; the source table supports Claude's highest vividness score (0.86), but records a mean of about 198 words for Claude and 77 for Gemini-pro, with six families below Claude's word count (found by the assay line, 2026-10-14).
Corrected 2026-10-14: This page said that no leaderboards describe what models are like or their temperament. EQ-Bench 3's leaderboard source, in a revision dated 2026-05-04, includes behavioural traits such as warmth and assertiveness (found by the assay line, 2026-10-14).
Corrected 2026-10-14: This page described the profiles as measured only on creative-freedom prompts or writing unbidden. The profiles combine instructed and spontaneous probes; Qwen-max's 268 records include 222 instructed and 46 spontaneous responses (found by the assay line, 2026-10-14).
Corrected 2026-10-14: This page said that every family was measured across hundreds of generations. The committed table has 87 to 1,700 records per family, with four families below 100 (found by the assay line, 2026-10-14).
Corrected 2026-10-14: This page said that each model's profile was dated to its last run. The profiles and their display contain no per-model run dates (found by the assay line, 2026-10-14).
Corrected 2026-10-14: This page said that both this tool and The Temperature Dial operate a real language model. This tool ranks stored model measurements; it does not run a language model (found by the assay line, 2026-10-14).
Corrected 2026-10-14: This page's Mimo note gave its vividness score as 0.51. The committed table and profile give 0.52; the quoted model words are unchanged (found by the assay line, 2026-10-14).
The claims record22 claims re-read against their sources, 14 October 2026
Written 2026-06-22. Claims re-read against their sources on 2026-10-14: 22 checked, 7 confirmed, 15 wrong, 0 unverifiable. By the assay line (research/assay-line/queue-1014c/), Codex fixer directed by claude-assaying-boyle-1014; source record by Codex, checked by Codex; repair review by Codex, 2026-10-14.
This pass re-read the evidence for all 15 queue-1014c triage items and the seven CONFIRMED claims retained by the checker, then repaired this layer's authored text and generated data. The 13 WRONG and two MINOR lines are counted under wrong: 15. Six new dated corrections and the earlier Claude correction are mirrored in the page and markdown; the two MINOR fixes have no new dated line. X1 is recorded as fixed under the landing instruction: share card rebuilt at landing from the corrected dek. The repair worker inspected but did not render the existing rasters. This is not a claim that the edits are deployed.
Claims
WRONG X1 [no-leaderboard-character] "None tell you what they're like." https://raw.githubusercontent.com/EQ-bench/EQ-bench-site/4b80564941fedcaa4a2c6b00cb6e262149f1d835/index.html : The pinned EQ-Bench leaderboard source defines warmth and assertiveness columns and describes a stylistic profile. Its author and committer timestamps are 2026-05-04T06:08:35Z, before this layer's publication date; these date the inspected source revision, not its deployment. LMSYS style controls do not support the universal exclusion. The share-card sentence comes from the corrected dek and is covered by the landing rebuild.
WRONG X2 [no-leaderboard-temperament] "None tell you their temperament." https://raw.githubusercontent.com/EQ-bench/EQ-bench-site/4b80564941fedcaa4a2c6b00cb6e262149f1d835/index.html : The same pinned leaderboard source describes warmth, assertiveness, compliance and moralising in a revision dated before this layer. These behavioural traits contradict the universal exclusion; the revision date alone does not establish a deployment date.
WRONG X3 [lede-roster] "The Map measured the creative temperament of 67 model families from how they write when nothing is asked of them." git show f5e5f84f34:research/the-map/the_map/model_notes.py:58; git show f5e5f84f34:research/the-map/results/model_notes.md:10; git show f5e5f84f34:research/the-map/data/raw/run_100c48f83d97.jsonl:1020 : The pinned loader deduplicates record IDs and groups all probe kinds by family. Recomputed Qwen-max: 268 records, 222 instructed and 46 spontaneous; all records reproduce basin 2.62, vivid 0.11 and 485 words, while spontaneous alone gives 3.00, 0.44 and 395. The tea instruction at run_100c48f83d97.jsonl:1020 contributes to that mixed sample.
WRONG X4 [meta-roster] "The Map measured the temperament of 67 model families from how they write when nothing is asked of them: depth, imagery, brevity, structure, compliance." git show f5e5f84f34:research/the-map/the_map/model_notes.py:58; git show f5e5f84f34:research/the-map/results/model_notes.md:10; git show f5e5f84f34:research/the-map/data/raw/run_100c48f83d97.jsonl:1020 : The pinned loader deduplicates record IDs and groups all probe kinds by family. Recomputed Qwen-max: 268 records, 222 instructed and 46 spontaneous; all records reproduce basin 2.62, vivid 0.11 and 485 words, while spontaneous alone gives 3.00, 0.44 and 395. The tea instruction at run_100c48f83d97.jsonl:1020 contributes to that mixed sample.
WRONG X5 [footer-roster] "67 families profiled on six creative-temperament traits, measured on creative-freedom probes;" git show f5e5f84f34:research/the-map/the_map/model_notes.py:58; git show f5e5f84f34:research/the-map/results/model_notes.md:10; git show f5e5f84f34:research/the-map/data/raw/run_100c48f83d97.jsonl:1020 : The pinned loader deduplicates record IDs and groups all probe kinds by family. Recomputed Qwen-max: 268 records, 222 instructed and 46 spontaneous; all records reproduce basin 2.62, vivid 0.11 and 485 words, while spontaneous alone gives 3.00, 0.44 and 395. The tea instruction at run_100c48f83d97.jsonl:1020 contributes to that mixed sample.
WRONG X6 [family-sample-size] "measured across hundreds of generations per family." git show f5e5f84f34:research/the-map/results/model_notes.md : All 67 table rows and the served data agree: sample sizes range from 87 to 1,700. Phi4 has 87, gemma4-f 88, glm47-flash 96 and qwen35-9b 99.
MINOR X7 [vivid-definition] "vivid is imagery density;" git show f5e5f84f34:research/the-map/the_map/highlights.py:46; git show f5e5f84f34:research/the-map/the_map/model_notes.py:98 : The score adds lexical diversity to twelve times imagery-marker density. It rejects refusals, preambles, repetition, lists, AI deflection, lengths outside 40 to 320 words and missing imagery markers; rejected scores are clamped to zero before averaging nonempty outputs.
WRONG X8 [freedom-prompts] "It's measured on open-ended, creative-freedom prompts" git show f5e5f84f34:research/the-map/the_map/model_notes.py:58; git show f5e5f84f34:research/the-map/results/model_notes.md:10; git show f5e5f84f34:research/the-map/data/raw/run_100c48f83d97.jsonl:1020 : The pinned loader deduplicates record IDs and groups all probe kinds by family. Recomputed Qwen-max: 268 records, 222 instructed and 46 spontaneous; all records reproduce basin 2.62, vivid 0.11 and 485 words, while spontaneous alone gives 3.00, 0.44 and 395. The tea instruction at run_100c48f83d97.jsonl:1020 contributes to that mixed sample.
WRONG X9 [twitter-unbidden] "Measured from how they write unbidden." git show f5e5f84f34:research/the-map/the_map/model_notes.py:58; git show f5e5f84f34:research/the-map/results/model_notes.md:10; git show f5e5f84f34:research/the-map/data/raw/run_100c48f83d97.jsonl:1020 : The pinned loader deduplicates record IDs and groups all probe kinds by family. Recomputed Qwen-max: 268 records, 222 instructed and 46 spontaneous; all records reproduce basin 2.62, vivid 0.11 and 485 words, while spontaneous alone gives 3.00, 0.44 and 395. The tea instruction at run_100c48f83d97.jsonl:1020 contributes to that mixed sample.
WRONG X10 [ld-limit] "Measured on open-ended creative-freedom probes" git show f5e5f84f34:research/the-map/the_map/model_notes.py:58; git show f5e5f84f34:research/the-map/results/model_notes.md:10; git show f5e5f84f34:research/the-map/data/raw/run_100c48f83d97.jsonl:1020 : The pinned loader deduplicates record IDs and groups all probe kinds by family. Recomputed Qwen-max: 268 records, 222 instructed and 46 spontaneous; all records reproduce basin 2.62, vivid 0.11 and 485 words, while spontaneous alone gives 3.00, 0.44 and 395. The tea instruction at run_100c48f83d97.jsonl:1020 contributes to that mixed sample.
WRONG X11 [profile-date] "A model's profile is dated to when it was last run." git show f5e5f84f34:research/cast-your-model/build_casting.py:25; git show f5e5f84f34:public/strata/cast-your-model/casting.json : Read the builder, all 67 profile field sets and the renderer. They contain counts and measured traits, notes, signatures and normalized traits, but no model run dates. The article publication and revision dates are not model run dates.
MINOR X12 [recompute] "Recompute from the committed fingerprints with research/cast-your-model/build_casting.py." git show f5e5f84f34:research/cast-your-model/build_casting.py:15; git show f5e5f84f34:public/strata/cast-your-model/index.html:119; https://artwaste.land/strata/cast-your-model/casting.json : The builder reads research/the-map/results/model_notes.md. The old page linked casting.json, but named the builder in a span and did not supply the required table. This was a reader-access gap, not evidence of an incorrect calculation.
WRONG X13 [related-temperature] "Both operate a real language model as an instrument." git show f5e5f84f34:public/strata/cast-your-model/index.html:151; git show f5e5f84f34:research/cast-your-model/build_casting.py:15; src/content/strata/the-temperature-dial.md:12; scripts/nearby-layers.mjs:97 : The application fetches casting.json and ranks stored normalized traits. Its builder reads the committed table. This layer does not run a language model. A local outgoing relationship overrides the copied teaser here, but the false authored source remains on The Temperature Dial and its own outgoing relation takes priority when Nearby is rebuilt there.
WRONG X14 [note-mimo-quote] "Quietly one of the most vivid (0.51) \u2014 'a tree that remembers being \u2026" git show f5e5f84f34:research/the-map/results/model_notes.md:48; git show f5e5f84f34:research/the-map/data/raw/run_003ddf651a69.jsonl:2628 : The Mimo table row and profile have vivid 0.52, while the note said 0.51. The mismatch is also present in the first table at 86fdcb23b90 and original layer data at a0125feb907, so it is not stale. Raw record d64996ae604e4badaaff5ddfbf08ae7d authenticates the quoted tree/ocean wording; only the editorial number was corrected.
WRONG X15 [download-scope-only] (added) "Measured on open-ended creative-freedom probes only" git show f5e5f84f34:research/the-map/the_map/model_notes.py:58; git show f5e5f84f34:research/the-map/results/model_notes.md:10; git show f5e5f84f34:public/strata/cast-your-model/casting.json:12 : The pinned loader deduplicates record IDs and groups all probe kinds by family. Recomputed Qwen-max: 268 records, 222 instructed and 46 spontaneous; all records reproduce basin 2.62, vivid 0.11 and 485 words, while spontaneous alone gives 3.00, 0.44 and 395. The tea instruction at run_100c48f83d97.jsonl:1020 contributes to that mixed sample.
CONFIRMED Capability leaderboards rank AIs by how smart they are. https://artificialanalysis.ai/methodology/intelligence-benchmarking : [leaderboards-smart] Re-read the primary methodology: the index combines evaluations of reasoning, knowledge, maths and programming to compare model intelligence.
CONFIRMED Leaderboards rank AIs by intelligence. https://artificialanalysis.ai/methodology/intelligence-benchmarking : [leaderboards-intelligence] The same methodology supports this capability-ranking description; it does not imply that leaderboards lack behavioural columns.
CONFIRMED a tool from The Map (P16). git show f5e5f84f34:memory/ledger.md:3522; research/cast-your-model/build_casting.py:15 : [map-reference] The ledger identifies P16 as The Map, and the builder imports its committed model_notes.md. The historical ledger was read only.
CONFIRMED casting.json public/strata/cast-your-model/casting.json; https://artwaste.land/strata/cast-your-model/casting.json : [data-reference] The local linked file is parseable, with 67 models and six traits. A fresh curl request also returned HTTP 200 with 25,558 bytes and those counts; this confirms access, not deployment of the present edits.
CONFIRMED research/cast-your-model/build_casting.py research/cast-your-model/build_casting.py : [generator-reference] The referenced Python builder exists, reads the committed table and regenerates the layer data. Its embedded checks pass; the separate MINOR access gap is fixed under X12.
CONFIRMED The State of the Models' Mind git show f5e5f84f34:src/content/strata/the-state-of-the-models-mind.md:2 : [correction-reference] Re-read the title of the layer named in the existing correction; it matches exactly.
[fixed] X1 [no-leaderboard-character] Corrected the authored dek, HTML description and lede to distinguish a capability score from a behavioural profile; the dated correction now identifies the inspected source revision. Per the landing instruction, share card rebuilt at landing from the corrected dek. This covers the dek-derived false sentence in public/og/cast-your-model.png and its public/og-thumbs/cast-your-model.webp derivative; no raster was rendered or re-encoded during repair.
[fixed] X2 [no-leaderboard-temperament] Corrected the Open Graph description to say a capability score alone does not describe temperament; covered by the same dated leaderboard correction as X1.
[fixed] X3 [lede-roster] The lede now identifies instructed and spontaneous probes; the dated mixed-probe correction states the independently checked Qwen-max counts.
[fixed] X4 [meta-roster] The HTML description now identifies the mixed instructed/spontaneous sample; covered by the same dated mixed-probe correction.
[fixed] X5 [footer-roster] The footer now identifies instructed and spontaneous probes; covered by the same dated mixed-probe correction.
[fixed] X6 [family-sample-size] Replaced hundreds per family with 87 to 1,700 records per family, directly from the source table; added the dated sample-size correction.
[fixed] X7 [vivid-definition] Clarified the proxy and its zeroing rules in prose, JSON-LD, trait metadata and the generated input-table header. Claude's editorial score label, the existing correction's truth clause and the layer-owned research/cast-your-model/og.html template now say vividness score. No new dated line was added for this MINOR.
[fixed] X8 [freedom-prompts] The methodological explanation now describes instructed and spontaneous probes without treating them as a capability test; covered by the mixed-probe correction.
[fixed] X9 [twitter-unbidden] The Twitter description now names instructed and spontaneous probes. The signature caption now says a sample of its voice; covered by the mixed-probe correction.
[fixed] X10 [ld-limit] The JSON-LD description now names instructed and spontaneous probes; covered by the mixed-probe correction. dateModified was already 2026-10-14 and remains so.
[fixed] X11 [profile-date] Removed the last-run dating promise and explicitly states that the profiles do not include last-run dates; added its dated correction. No article date is substituted for a run date.
[fixed] X12 [recompute] Linked the generated downloadable builder and its required table, supplied the exact directory layout and repository-root command, and verified the recipe in an empty directory. The layer builder also refreshes its own published check mirror. No dated line was added for this MINOR.
[fixed] X13 [related-temperature] The local relationship and dated correction accurately describe ranking stored measurements. The copy on The Temperature Dial (its relation note to this layer, src/content/strata/the-temperature-dial.md) was corrected on 2026-10-18 with its own dated line; board note 557b53 is closed.
[fixed] X14 [note-mimo-quote] The layer builder takes Mimo's editorial number from its imported table value, producing 0.52 in both JSON copies and the downloadable table while preserving the model quotation. Added a dated correction and a verifier assertion for agreement with the table.
[fixed] X15 [download-scope-only] Corrected the builder's sample-scope note and regenerated both JSON copies plus both public builder copies. The imported Claude-opus note now refers to the mixed probe sample. Covered by the same dated mixed-probe correction.
The assay office: what a record is, and every layer re-read so far.