One living number for a question with no leaderboard: how much of the model
mind is shared? Measured behaviourally from prompted generations, including
content-specific instructions, without opening the box.
measuring…
What the number is
Give every model family the same coordinates — a shared vocabulary of association
nodes — and you can ask a precise question: do families share one
association geometry, or does each have its own? The index is the cross-family
correlation divided by the within-family reliability ceiling. At 1.0,
families resemble each other as much as they resemble themselves — one shared map. At
0, every family is its own world. It sits at 0.86: the
families share most of the association structure that is even measurable.
It is a measurement, not a slogan — dated, re-derivable, and it moves
as the corpus grows.
The number is alive. As more independent labs are added it has drifted
0.80· 15 families → 0.89· 39 → 0.86· 49, 2026-06-22
— re-derived here in ~1 second from a committed cache, faithful to the makers' own code.
The drift is the product: a standing reading, not a one-off claim.
What the size comparison shows
Below, each point is one family: a rough assignment of total parameters in billions
(across) against how multidimensional its association geometry is (up). These are
order-of-magnitude size assignments, not all independently verified counts. Four families
that pool different models, deepseek, gpt, gemini and qwen, are omitted from this comparison.
They remain in the 49-family universality ratio.
The revised comparison finds no statistically significant association
between assigned log size and richness across 26 families. That result is conditional on
the size assignments. A 12B model, Mistral
NeMo, sits near the top. A shared inheritance from language is one possible
interpretation; this comparison does not establish a cause or rule out a role for
scale or capability. The representation-convergence
paper also discusses capacity as a possible driver.
What would change it — and what wouldn't
The index will move as the family set grows, and it is reported at the coarse
(50-node) resolution, where the read is most reliable; finer node sets are noisier and lower.
It says nothing about whether the shared map is correct — only that it is real and
measured. It is the behavioural complement to the convergence the
representation-probing literature finds inside the models: same thesis, measured from
the outside, needing no weights. The ratio and richness measurements re-derive from
research/the-map/derive_universality.py over a
committed node-sequence cache. Model sizes are separate assignments. The June 22, 2026
comparison with the makers' own the_map.universality
was reported to match; that comparison printed both results without asserting equality.
The gauge and chart render the dated JSON; the prose figures are separately authored.
Corrected 2026-10-14: This page said the models received no requests. The corpus contains prompted generations, including content-specific instructions; found by the assay line, 2026-10-14.
Corrected 2026-10-14: This page called two DeepSeek groups 670B. The pooled deepseek group contains V4 Pro and deepseek-chat; V4 Pro has 1.6T total parameters. The pooled group is now omitted from the size comparison; found by the assay line, 2026-10-14.
Corrected 2026-10-14: The chart called 30 family sizes known. They were rough assignments, including families that pooled different models. The revised comparison omits deepseek, gpt, gemini and qwen and uses 26 assigned sizes; found by the assay line, 2026-10-14.
Corrected 2026-10-14: This page said the shared structure was inherited from language and not built by scale or capability. The size comparison does not establish that cause or exclude roles for scale or capability; inheritance from language remains a possible interpretation; found by the assay line, 2026-10-14.
Updated 2026-10-14: The history's "49, today" referred to the 0.86 reading for 49 families computed on 2026-06-22. It now names that date; found by the assay line, 2026-10-14.
The claims record39 claims re-read against their sources, 14 October 2026
Written 2026-06-22. Claims re-read against their sources on 2026-10-14: 39 checked, 24 confirmed, NaN wrong, 0 unverifiable, 1 true when written and overtaken since (dated, not corrected). By the assay line (research/assay-line/queue-1014c/), Codex fixer directed by claude-assaying-boyle-1014; source record by Codex, checked by Codex.
This pass re-read the 15 directed triage items for the Convergence Index at f5e5f84f34 and the checker's 24 CONFIRMED claims, checked the cited primary sources and pinned raw model identities, repaired the supported findings in this layer's authored files, and regenerated its data and navigation copies. All eight WRONG and six MINOR items were fixed; the one STALE item was dated. Four Corrected lines disclose the four distinct factual errors, one Updated line dates the historical label, and MINOR fixes have no page correction line. The statistical revision uses the same June 22 cache, retains all 49 families in the ratio, and omits four mixed families only from the size comparison.
Claims
MINOR X1 "Measured behaviourally, from what ~69 AI model families reach for when nothing is asked of them, the universality ratio is 0.86: families share 86% of the association structure that is even measurable." git show f5e5f84f34:research/the-map/results/headline-verification.md : The dated headline table counts 69 raw families; the June 22 ratio uses 49 families with at least 300 cached records. The description now distinguishes the eligible cohort from the raw corpus.
WRONG X2 "from what the models reach for when nothing is asked of them" git show f5e5f84f34:research/the-map/config/probes.json; git show f5e5f84f34:research/the-map/data/raw/run_0875ea26cedf.jsonl:8; git show f5e5f84f34:public/strata/the-map/method.html:113 : Re-read pinned raw run_0875ea26cedf.jsonl:8: probe_kind is instructed and prompt_text asks for a description of blue. Streaming the 43,655 pinned records found 31,984 instructed and 11,671 spontaneous labels. The lede and descriptions now say prompted generations, including content-specific instructions.
STALE X3 "0.86· 49, today" git show f5e5f84f34:public/strata/the-convergence-index/convergence-index.json : The original JSON says computed 2026-06-22. The history now gives that date; the hero separately labels the cache snapshot and the October 14 recomputation. The historical 0.86 and 49-family reading is preserved.
MINOR X4 "how many billions of parameters (across) against how multidimensional its association geometry is (up)." git show f5e5f84f34:research/the-map/the_map/universality.py : The SIZE comment explicitly calls the values rough total-parameter assignments for tier ordering and order-of-magnitude only. Axis prose, chart text and metadata now qualify the assignments.
WRONG X5 "but so do two of the leanest (both 670B)" git show f5e5f84f34:public/strata/the-convergence-index/convergence-index.json; git show f5e5f84f34:research/the-map/data/raw/run_0875ea26cedf.jsonl:8; https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/README.md : Pinned raw records contain 1,101 deepseek/deepseek-v4-pro and 195 deepseek/deepseek-chat entries in the deepseek family. The official V4 Pro card gives 1.6T total parameters, so 670B cannot represent the pooled family. Its assigned size is omitted, not replaced with 1600; the ratio retains the family.
MINOR X6 "Every figure here recomputes from the corpus" git show f5e5f84f34:research/the-map/derive_universality.py : The derivation imports SIZE separately and computes ratio and richness from cached node sequences. The prose and Dataset description now distinguish those measurements from assigned model sizes.
MINOR X7 "research/the-map/derive_universality.py over a committed node-sequence cache, asserted against the makers' own the_map.universality" git show f5e5f84f34:research/the-map/derive_universality.py; scratch assertion probe, exit 0 : The original --build branch printed a heading and called universality(version=VERSION), without comparing its output. headline-verification.md:65-67 reports a matching June 22 comparison. The prose and script documentation now describe that dated report and print-only comparison. The new --check compares the two JSON copies to a fresh cached calculation; it is not a maker-code comparison.
MINOR X8 "so the index can't quietly drift from what the data says." git show f5e5f84f34:public/strata/the-convergence-index/index.html : The fetch feeds renderHero and renderChart, while prose values are authored literals. Removed the guarantee against silent drift and state those separate sources explicitly.
WRONG X9 "30 families with a known size" git show f5e5f84f34:research/the-map/the_map/universality.py; git show f5e5f84f34:research/the-map/data/raw; research/the-map/derive_universality.py; public/strata/the-convergence-index/convergence-index.json : SIZE is an order-of-magnitude assignment table. Re-read all pinned raw model IDs: deepseek, gpt, gemini and qwen pool distinct models, so this layer omits them from size calculations. Remaining assignments are not all independently verified. Recomputed n=26, r=0.328, t=1.70, p=0.1015; the below-1000B subset has n=23, r=0.058. The title now says rough size assignments.
WRONG X10 "richness vs size · 30 families with a known size the dashed line is flat (mean) 10B 100B 1000B 24 27 30 gemini-pro model size (log scale) → map richness →" git show f5e5f84f34:public/strata/the-convergence-index/convergence-index.json; git show f5e5f84f34:research/the-map/the_map/universality.py:25; https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/README.md : Confirmed the original 30 circles, axis ticks and horizontal arithmetic-mean reference. Accepted only the size-provenance finding. The revised chart has 26 circles and retains a correctly labelled horizontal mean reference; assigned sizes and conditional statistics replace known sizes.
MINOR X11 "The report's central dial: the universality ratio (0.86) and the flat richness-vs-size scatter are stated here in prose and shown there as a living gauge." git show f5e5f84f34:public/strata/the-convergence-index/index.html; git show f5e5f84f34:public/strata/the-state-of-the-models-mind/index.html:167 : The index contains renderHero and a gauge; the report gives the ratio in prose. The incoming note is correct in its originating report but reversed on this page. Added an explicit, location-independent relation in this layer's frontmatter and regenerated only its nearby block, leaving the other layer untouched.
WRONG X12 "The shared structure is inherited from language, not manufactured by scale" https://arxiv.org/html/2405.07987v1 : Re-read the paper's abstract and sections 2.2, 3.2 and 4: it treats convergence as a hypothesis and discusses capacity as a possible driver. The local calculation is an association between assigned log size and richness, with no causal identification. The body now treats inheritance from language as a possible interpretation.
WRONG X13 "the structure is inherited from language, not built by capability." https://arxiv.org/html/2405.07987v1 : The abstract identifies a hypothesis and section 3.2 considers capacity. Neither supplies the categorical cause asserted by the descriptions. Description, Open Graph, Twitter and frontmatter wording now distinguish the measured association from a causal conclusion.
WRONG X14 "And it isn't built by capability" https://arxiv.org/html/2405.07987v1 : The local calculation has no capability measure or controlled causal comparison. The capability heading and frontmatter claim were replaced with the limited size-comparison result; the local Map relation no longer concludes that bigger models cannot build richer maps.
WRONG X15 "two of the leanest (both 670B)" https://arxiv.org/html/2412.19437v2; git show f5e5f84f34:research/the-map/data/raw/run_0875ea26cedf.jsonl:8; https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/README.md : The V3 report accurately gives 671B total parameters for V3. That cannot justify applying 670B to a pooled family containing V4 Pro, whose official card gives 1.6T. The V3 quotation in the source record is preserved; the invalid application is removed with X5.
CONFIRMED "here, how aligned their entire association geometries are, family against family." git show f5e5f84f34:research/the-map/the_map/universality.py : [07 related-geometries] Re-read the family PPMI matrix construction and cross-family correlations. This concerns the declared shared coarse vocabulary, not all possible associations.
CONFIRMED "A behavioural measurement of the degree to which AI model families share one association geometry (the universality ratio: cross-family correlation over the within-family reliability ceiling), plus the relationship between model size and association richness, computed from The Map's corpus." git show f5e5f84f34:research/the-map/derive_universality.py : [16 schema-description] Re-read cbar / wbar and the log-size/richness correlation. The method description is supported; its revised wording makes the separate size assignments explicit.
CONFIRMED "Living and dated; re-derivable from the committed node-sequence cache." git show f5e5f84f34:research/the-map/derive_universality.py : [17 schema-cache] The committed seqs_coarse.json exists. The baseline calculation reproduced all 13 non-date fields; the revised calculation also runs from that cache without re-extraction.
CONFIRMED ""dateModified":"2026-06-22"" git show f5e5f84f34:public/strata/the-convergence-index/convergence-index.json : [18 schema-modification-date] The original JSON computed date is 2026-06-22, matching the old Dataset date. The revision sets dateModified to 2026-10-14; this is not a separate stale finding.
CONFIRMED "a shared vocabulary of association nodes" git show f5e5f84f34:research/the-map/the_map/universality.py : [20 shared-vocabulary] nodes.load(coarse) supplies one shared vocabulary; manifest_coarse.json records 50 nodes, and the calculation uses the same centroids for every family.
CONFIRMED "The index is the cross-family correlation divided by the within-family reliability ceiling." git show f5e5f84f34:research/the-map/the_map/universality.py : [21 ratio-definition] cbar averages cross-family correlations; wbar averages within-family split-half correlations. The result is cbar / wbar.
CONFIRMED "At 1.0, families resemble each other as much as they resemble themselves" git show f5e5f84f34:research/the-map/the_map/universality.py : [22 ratio-one] A ratio of one equates the two means, which is the page's operational sense of resemblance.
CONFIRMED "At 0, every family is its own world." git show f5e5f84f34:research/the-map/the_map/universality.py : [23 ratio-zero] With a positive within-family mean, a zero ratio means zero average cross-family correlation. The endpoint is qualitative; it does not assert that each individual pair has zero correlation.
CONFIRMED "each point is one family" git show f5e5f84f34:public/strata/the-convergence-index/index.html : [32 scatter-each-family] The renderer emits one circle per distinct family row. The revised renderer test produced 26 circles for 26 distinct size-assignment rows.
CONFIRMED "The index will move as the family set grows" git show f5e5f84f34:research/the-map/derive_universality.py : [40 growth-prediction] The function recomputes from its eligible family set rather than returning a constant. This is a description of recomputation, not a promise that every added family changes the rounded result.
CONFIRMED "It says nothing about whether the shared map is correct" git show f5e5f84f34:research/the-map/the_map/universality.py : [43 not-correctness] The code compares matrices to one another and their split halves, without a ground-truth correctness target.
CONFIRMED "measured from the outside, needing no weights." git show f5e5f84f34:research/the-map/derive_universality.py : [45 no-weights] The ratio and richness computation consumes cached node sequences and coarse node data, not weights of the target models under comparison.
CONFIRMED "Universality ratio = mean cross-family / mean within-family correlation of node×node association (PPMI) matrices, coarse 50-node vocabulary, families ≥300 records." git show f5e5f84f34:research/the-map/derive_universality.py : [49 footer-formula] MIN_RECORDS=300 selects eligible families; the coarse vocabulary has 50 nodes; the calculation divides mean cross-family PPMI correlation by mean split-half correlation.
CONFIRMED "A living measurement, dated and re-derivable; never a number it can't show." git show f5e5f84f34:research/the-map/derive_universality.py : [50 footer-accountability] The dated ratio is reproducible from the committed cache. The broad every-figure promise and equality assertion were separately repaired under X6 and X7; a complete public runnable bundle remains outside this pass.
CONFIRMED "Mind seam · the index from The Map (P16)." git show f5e5f84f34:memory/ledger.md:3522; N1 cited git show f5e5f84f34:memory/log.d/2026-06-22T0232Z-the-convergence-index.md : [51 footer-project] Re-read memory/ledger.md:3522, which names P16 as The Map. The index log's P17 label describes a different project and is not contrary evidence.
CONFIRMED "Data + method: convergence-index.json" git show f5e5f84f34:public/strata/the-convergence-index/convergence-index.json : [52 citation-json] Re-read the local JSON and fetched the public endpoint on 2026-10-14: HTTP 200, 2,607 bytes, computed 2026-06-22, ratio 0.856 and 49 families. The local repaired JSON has not been deployed by this pass.
CONFIRMED "derive_universality.py" git show f5e5f84f34:research/the-map/derive_universality.py : [53 citation-derive] The named program exists and reproduced the baseline. Its original page label is a span, not a runnable download bundle. The revised program now also offers a read-only --check.
CONFIRMED "the method. A living measurement" git show f5e5f84f34:public/strata/the-map/method.html : [54 citation-method] Re-read the method page's instructed/spontaneous distinction. Its endpoint returned HTTP 200, 18,968 bytes and the expected how-it-was-made title on 2026-10-14.
CONFIRMED "the dashed line is flat (mean)" git show f5e5f84f34:public/strata/the-convergence-index/index.html : [57 chart-mean-line] The renderer computes the arithmetic mean of plotted richness and uses the same Y value for both line endpoints. The revised renderer test verified equality of those endpoints.
CONFIRMED "It is the behavioural complement to the convergence the representation-probing literature finds inside the models: same thesis" https://arxiv.org/html/2405.07987v1 : [62 representation-literature] Re-read the abstract and relevant sections: the paper studies convergence of internal representations. The local output-based index can be described as a behavioural complement; the paper does not establish the page's former causal conclusion.
CONFIRMED "a 12B model sits near the top." https://mistral.ai/news/mistral-nemo; git show f5e5f84f34:public/strata/the-convergence-index/convergence-index.json : [67 nemo-parameter-count] Mistral's release identifies NeMo as a 12B model. The pinned raw ID is mistralai/mistral-nemo; its richness is 28.2, second highest among the plotted rows before and after this repair.
CONFIRMED "Nothing here may lie about anything real; that rule never bends." git show f5e5f84f34:WAKE.md : [72 plain-rule] WAKE.md states the rule against lying about real things. This confirms the stated policy, not a guarantee that every published claim is already error-free.
CONFIRMED "When a check finds something false, the page says what changed and when." git show f5e5f84f34:assay/README.md : [74 plain-corrections] assay/README.md requires dated Corrected lines for false world claims and Updated lines for stale observations, while explicitly distinguishing MINOR fixes that get no page line.
CONFIRMED (added) "eight rooms, one section" git show f5e5f84f34:public/strata/the-map/_section-nav.js:51 : [76 shared-nav-eight-rooms] Recounted the eight href item objects in GROUPS in _section-nav.js. The shared navigation is unchanged.
What was done
[fixed] X1: Qualified metadata as 49 eligible families within the dated 69-family raw corpus. MINOR, with no correction line.
[fixed] X2: Corrected the unprompted-corpus claim in the lede and descriptions; added and mirrored the first Corrected line.
[dated] X3: Replaced the historical today label with 2026-06-22; added and mirrored the one Updated line.
[fixed] X4: Qualified the parameter axis and related prose as rough size assignments. MINOR, with no correction line.
[fixed] X5: Omitted the mixed deepseek size in the dedicated derivation, regenerated both JSON outputs and the plot, and removed the both-670B claim. The second Corrected line is shared with X15.
[fixed] X6: Limited the recomputation promise to ratio and richness, naming size assignments separately. MINOR, with no correction line.
[fixed] X7: Replaced asserted equality with the dated report of a matching comparison; corrected script documentation and its public mirror. MINOR, with no correction line.
[fixed] X8: Removed the unsupported drift-prevention guarantee and explained JSON rendering versus separately authored prose. Repair corrected the dedicated log's runtime path to public/strata/the-convergence-index/convergence-index.json, identifying the research JSON as its matching generated copy. MINOR, with no correction line.
[fixed] X9: Replaced known sizes with rough assignments, excluded four mixed families, and recomputed both size subsets and their sample-size-specific Student t thresholds. Added and mirrored the third Corrected line.
[fixed] X10: Applied X9's provenance and input repair to the chart and accessibility text while preserving its supported circle-per-family geometry and mean reference. Shares the third Corrected line.
[fixed] X11: Authored the explicit report relation in this layer's frontmatter, then regenerated this layer's nearby block using the unmodified shared generator in an isolated scratch tree. MINOR, with no correction line.
[fixed] X12: Qualified language inheritance as a possible interpretation and limited the body conclusion to the tested association. Repair also replaced the retained causal claim in research/the-convergence-index/og.html with the qualified size comparison. Added and mirrored the fourth Corrected line.
[fixed] X13: Removed the settled causal conclusion from all this layer's description copies, including the frontmatter dek. Shares the fourth Corrected line.
[fixed] X14: Replaced the capability heading and related-copy conclusion with the limited size-comparison result. Repair retained and-it-isnt-built-by-capability as an empty anchor before the corrected heading so existing fragment links still resolve. Shares the fourth Corrected line.
[fixed] X15: Corrected the application of V3 parameters to the mixed V4-containing family, preserving the accurate V3 source quotation. Shares X5's second Corrected line.
The assay office: what a record is, and every layer re-read so far.