Is Meaning All You Need?

Two-paper note. This program is split into two manuscripts. The empirical core — the unsupervised translator, probe and steering transfer, the rank law, the null-completion audit, and the bounds — is reported in full, with methods and controls, in the companion paper "Portable Concept Geometry Enables Cross-Model Probe and Steering Transfer." This paper is the synthesis: it carries the arc (the failed "topology of thought," the dual-null epistemology, the bounded thesis that meaning is the architecture-invariant the mechanism converges to, and the reflexive turn from observed frame to installed instrument), summarizes the empirical results, and cites the companion for the rigorous detail. The split follows reviewer guidance: separate knives cut cleaner — a reviewer who doubts the synthesis should not be able to downgrade the concept-geometry result, and vice versa.

---

Abstract

We set out to map the topology of thought: tens of thousands of model-generated "thoughts" passed through a scoring gate to build an atlas, on the assumption that the output of a frontier language model hides a rich latent structure waiting to be charted. It does not — or rather, most of the structure we first measured was put there by the instrument we used to measure it. Under an anisotropy-preserving null, the discrete-basin / modal reading of that atlas fails to reject: the topology was largely a frame the observer supplied, not a property of the mind.

What survives the correspondence, rotation, and transfer nulls — the nulls appropriate to a cross-model claim — is something different, and deeper. Walking the instruments upstream — from output embeddings, to the residual stream, to the mid-stack — the object that resists framing is not a topology of thoughts but a conceptology: a shared geometry of meaning. Across the tested text-model families and a 390× scale range, concept geometry is near-isometric up to an orthogonal rotation recoverable from unlabeled parallel concepts alone (representational-similarity z = 36; a learned translator places held-out concept centroids at top-1 = 1.000 in the clean within-family pipeline — 0.994 cross-family, 0.889 averaged over nine models); across the cross-modal edge it degrades measurably. The invariance is aggregate: concept centroids and linear directions transfer near-losslessly between minds, while individual instances do not (cross-mind sentence retrieval ≈ 0.14). It is bounded: strongest within the transformer text family, partial across the modality edge (translator centroid top-1 falls from ~0.97 within-family, under the cross-modal-comparison fit — the canonical within-family ceiling is 1.000 — to 0.35 into a multimodal model).

Crucially, the shared geometry is not only readable but writable. A linear probe trained in one mind, rotated by the unlabeled translator, classifies in another at native fidelity; and — the new result here — a causal steering vector built in one mind, rotated by the same translator, causally steers another at native fidelity, provided the translator is high-rank. We use this to equip a small reference model with a two-part epistemic instrument: a shared epistemic-honesty steering axis (abstention, sycophancy-resistance, and refusal-of-the-unanswerable collapse to a single direction, pairwise cosine 0.93–1.00) plus an independent off-map readout (held-out AUROC 0.89). On a small held-out behavioral battery the steering axis raises epistemic honesty (mean 60→80%) while preserving correct answers (100%) and not over-refusing (0%); reading-off-map and steering-honest are distinct axes (cosine ≈ 0.2–0.4), so the mind both detects and acts, via two complementary directions. We report this as a proof of concept: the geometric (cosine, transfer) results are robust, the behavioral percentages rest on 10–20 probes per axis.

The discipline that demolished the false topology — a pair of nulls separating the observer's frame from the invariant core — is the same discipline that certified every transfer. We report what we killed as carefully as what we found. We also close the program's main seam: in one instrument (the residual stream) the within-mind modal frame dissolves against its own anisotropy-preserving null (z ≈ 0) while the cross-mind concept core survives that same strict null (RSA 0.93, z = +30) — the manufactured frame and the real core, separated by one discipline. And we have now run the decisive test of the title itself: two minds trained from scratch on provably disjoint symbols (non-overlapping character pools, audited intersection ∅) describing one authored world still share concept geometry at cross-mind RSA 0.794, z = +35 vs the anisotropy null, 94 % of the same-corpus ceiling, at ≥1B and reproduced across two independent runs — meaning is not shared tokens; the aggregate geometry survives even when the surface forms cannot overlap (instance recovery stays partial, as the law predicts). The bounded answer to the title is not "yes" in general: yes for aggregate concept geometry within the tested transformer family, partial across the modality edge, and no for instances. The aggregate/instance law also sets a resolution ceiling (added 2026-06-08): the coarse layer — concept centroids, directions, probes, steering vectors — is shared up-to-rotation, but the fine per-concept structure carried by a model's sparse-autoencoder dictionary does not transfer by that same rotation (matched cross-generation Qwen pair, fine SAE-feature transfer collapses to max-cosine median 0.14 against a 0.09 null, while the held-out coarse alignment stays intact at 0.69; §6, §7). The portable map is real but low-resolution. Or, as a line: attention was all you needed to build it; meaning is what it built — a portable aggregate geometry, up to rotation, bounded by architecture and modality, sharpest for centroids and directions rather than instances, and coarse rather than fine.

---

1. Introduction

The 2017 sequence-transduction result named a mechanism — attention — and showed it sufficient to replace recurrence and convolution. This paper is about what that mechanism, and its successors, converge to. Our central finding is that the specific mechanism is, within limits, not the load-bearing thing: different architectures, trained on overlapping web-scale data, build the same geometry of meaning, recoverable from one to another by a single orthogonal rotation. The engine was interchangeable; the meaning is the invariant.

We did not begin there. We began trying to chart a "topology of thought" and discovered, through a disciplined pair of null models, that the topology was mostly an artifact of measurement. The genuine, null-surviving structure lay upstream, at the level of concepts rather than thoughts, and it was shared across minds. This paper traces that arc end to end and adds the result that turns it from an observation into a capability: the shared geometry can be steered, not just read.

Contributions.

1. A negative result, honestly bounded. The discrete-basin "topology" of a frontier model's output atlas does not survive an anisotropy-preserving null (Section 3). We are careful about what this does and does not show: against this null we fail to reject; the test cannot distinguish "no structure" from "weak overlapping structure." What dissolves is the observer-supplied frame, not all structure.

2. A shared concept geometry (the positive result). Mid-stack representations of the same concepts in different model families are isometric up to rotation; the rotation is learnable from unlabeled parallel concepts; held-out concept centroids translate at top-1 = 1.000, with representational-similarity z = 36 (Section 5).

3. Three laws. The invariance is up to rotation (no shared coordinate system, only shared relational geometry) and aggregate-sharp / instance-ambiguous (centroids and directions transfer; instances do not) (Section 6). The aggregate/instance law extends onto a resolution axis: the coarse/aggregate layer is shared up-to-rotation, but fine per-concept structure (a model's SAE dictionary) is model-idiosyncratic and does not transfer by the same rotation (Section 6, Section 7).

4. Bounds and a gradient. The invariance holds across a 390× parameter range and cross-family, and degrades — measurably — across the modality edge (Section 7).

5. From reading to writing. A causal steering vector, rotated by the unlabeled translator, steers a second mind at native fidelity; reading needs a low-rank translator, steering a high-rank one (Section 8).

6. A reference "testing mind." Epistemic honesty is a single steerable axis; a two-part instrument (a steering direction + an independent off-map readout) equips a small model with calibrated epistemic behavior on a held-out battery — a proof of concept at small behavioral N (Section 9).

7. A disconfirmation record. We log the hypotheses we killed, including our own (Section 10).

Apparatus. We study open-weight models (Qwen2.5 and Qwen3 at 0.6B–235B, Llama-3.2/3.1 at 1B–8B, and the multimodal Gemma-4-12B) and one custom-served mid-size model. All instruments — the scoring "lens," the dual-null harness, the cross-mind translator, the steering-transfer harness, and the testing-mind stack — are custom-built by the authors. Code and configurations accompany this draft (Section: Reproducibility).

Vocabulary (operational). To keep the prose honest we fix terms to procedures:

Claim-status (read this first). The paper is rich; this table is the map of what we are and are not asserting.

---

1b. Related work, and how we differ

Our result sits at the intersection of four lines of work; we state the delta against each, because the nearest neighbors are close.

Representational convergence / the Platonic Representation Hypothesis (Huh et al., 2024; and the representational-alignment literature, e.g. Kornblith et al. CKA, 2019; Bansal et al., 2021). This line shows that as models scale and improve, their representations converge in similarity, plausibly toward shared statistics of the world. We take this as our starting premise and sharpen it into operations: we fit an explicit unsupervised orthogonal translator and quantify the correspondence (held-out centroid top-1 = 1.000), we draw the boundary the convergence claim leaves open — aggregate vs. instance and up-to-rotation vs. shared-space — and we map where the convergence frays (the cross-modal gradient). Convergence is the that; we give the what-transfers, how-far, and whether-you-can-write-through-it.

Relative representations and model stitching (Moschella et al., 2022; Lenc & Vedaldi, 2015). Relative representations encode each point by similarities to anchors, yielding a rotation/rescaling-invariant code for zero-shot stitching of read-out heads. Ours differs in three ways: (i) we use an explicit orthogonal map on concept centroids and show it is near-isometric (not merely a re-encoding that happens to stitch); (ii) we extend from stitching readouts to transferring causal interventions (steering), which prior stitching work does not address; (iii) we establish the rank law — reading transfers at low rank, steering needs high rank — which is new and mechanistically specific.

Activation steering / representation engineering (Turner et al., ActAdd; Rimsky et al., Contrastive Activation Addition, 2023–24; Zou et al., Representation Engineering, 2023). This line builds and applies steering vectors within a single model. Our contribution is to port a steering vector across models through the unsupervised translator: a vector built in one mind causally steers another at native fidelity. Cross-model steering transfer, to our knowledge, has not been shown.

Anisotropy and the geometry of embeddings (Ethayarajh, 2019; Gao et al., 2019). The "cone effect" — that representations occupy a narrow anisotropic cone — is usually treated as a nuisance. We turn it into an instrument: an anisotropy-preserving Gaussian null that subtracts exactly the structure a cone could manufacture, which is what lets us separate the observer's frame from the invariant core (and what kills our own original "topology").

(Selected references are collected at the end of the paper.)

2. The dual-null discipline

Every claim in this paper is adjudicated by two null models, because the failure mode we are guarding against — an instrument inventing the very structure it reports — is subtle and self-flattering.

• Column-shuffle null. Permute features independently across items, destroying cross-feature structure while preserving per-feature marginals. Catches structure that is real covariance vs. coincidence.
• Anisotropy-preserving (Gaussian) null. Sample from a Gaussian matched to the full covariance of the representation — i.e., preserve the anisotropy (the dominant directions, the "shape" of the cloud) while destroying any genuine clustering. Catches structure that is real grouping vs. anisotropy alone.

The second null is the strict one, and it is the one that demolished our first result. An important self-correction (forced by adversarial review of our own draft): against the anisotropy null we fail to reject, which is weaker than "the structure dissolves." A power control showed that genuinely-but-weakly separated modes give the same negative against this null, so the test cannot separate "no modes" from "overlapping modes." We therefore report the dual-null not as a structure-detector but as a frame / core separator: it strips away structure an observer could have manufactured and leaves only what an observer could not.

For transfer claims (Sections 5, 8) the analog of the second null is a rotation null: replace the learned orthogonal translator R with (i) a row-shuffled R, and (ii) a freshly sampled random-orthogonal R. A transfer that survives both is carried by the concept-aligned rotation specifically, not by any high-dimensional map.

---

3. The negative result: the topology was a frame

The origin atlas was built by passing model-generated short texts through a scoring gate and embedding the survivors, then reading off apparent basins / modes — a "topology of thought." Three nulls later, the reading is not supported:

• Against the column-shuffle null, weak modal structure appears (z ≈ +11 at the deepest lens).
• Against the anisotropy-preserving Gaussian null, that signal fails to reject (z ≈ 0): the apparent basins are consistent with the anisotropy of the embedding cloud plus overlap, with no additional grouping.
• The cloud is better described as a continuous anisotropic manifold than as discrete basins.

The honest statement is not "there is no structure in model output." It is: the discrete-basin topology was a frame the observer (our scoring-and-embedding instrument) supplied. This is the first and most important methodological lesson of the arc — and the reason the rest of the paper measures upstream of the output, where the instrument's fingerprints are smaller.

---

4. Walking upstream

We moved the measurement from the model's output to its internals, in three steps: (1) sentence-embedder representations of generated text; (2) the model's own residual stream at the last token; (3) the residual stream mean-pooled at mid-stack. At each step the observer's contribution shrinks and a different object comes into focus — one organized by concept rather than by thought, and, as the next section shows, shared between minds.

A practical lesson recurs here and matters for everything downstream: the concept-rich depth is architecture-dependent. Llama and Qwen peak mid-stack (0.5–0.6 of depth); the multimodal Gemma-4 peaks early (0.21 of depth) and decays monotonically thereafter. Extracting every model at a single fixed depth is a near-worst-case choice for some architectures; each mind must be read at its own concept peak.

---

5. The shared concept geometry (the positive result)

We represent each of N concepts by its mean-pooled mid-stack vector in each model, language-erased so that the representation is a function of meaning, not surface language. Three measurements, all dual-null controlled:

(a) Representational similarity. The pairwise concept-similarity matrices of two different families (Qwen2.5-3B and Llama-3.2-3B) agree at RSA = 0.805 (z = 36) against a meaning-shuffle null — at or above the split-half ceiling of either model alone. Two architectures build the same relational concept geometry.

(b) A learned translator. Fit PCA followed by orthogonal Procrustes on the concept centroids of a training split of concepts (420 concepts), using no labels beyond "these are the same concept in both models." On a held-out split (180 concepts), the translator places each concept's centroid at top-1 = 1.000 (chance 0.0056). With a clean instruction-tuned teacher this reproduces at 1.000 within-family / 0.994 cross-family. There is a working, generalizing map between two minds, fit without supervision.

(c) Probe (direction) transfer. Train a linear probe to separate two concepts in model A; rotate its weight vector by R into model B; test in B. Across 300 held-out concept pairs the rotated probe classifies at 0.947, matching the native-B ceiling of 0.911, while the shuffled-R null sits at 0.495 and 100% of pairs exceed 0.70. A direction trained in one mind works in another after the unlabeled rotation. The same holds for a specific instrument — a grounded-vs-fabricated-entity ("familiarity") probe transfers at AUROC 1.000 cross-family against a clean 0.500 null.

Together: the shared object is not a vague "similarity" but an operational geometry — you can build an instrument in one mind, carry it by an unsupervised rotation, and run it in another.

---

6. Three laws

Law 1 — up to rotation. There is no shared coordinate system. Each mind has its own frame; what is shared is the relational geometry, recoverable by an orthogonal map. "Universal mentalese" overstates it: the right picture is a relational invariant with per-mind coordinates, not a common space all minds write into.

Law 2 — aggregate-sharp, instance-ambiguous. What transfers is aggregate objects — centroids and directions. What does not is instances:

You can locate what a concept is, relationally, in any mind; you cannot reconstruct a particular thought of one mind in another. This is the boundary we call govern, don't generate: the universal structure lives at the level of averaged meaning, which is why "topology of thought" was the wrong object and "conceptology" is the right one. (A capability corollary we initially over-claimed and then killed — see §10, item D — is that a "second mind's view" of a concept boundary is a mildly better estimator of a direction than either mind alone; real, Wilcoxon-significant, but size-independent and hence a data-ensemble effect, not capability uplift.)

Govern-not-generate, now measured: the field reads STATE, not truth (added 2026-06-08). The aggregate layer's product consequence is sharper than "you cannot reconstruct an instance." The residual field is the purest signal of a model's coarse epistemic state — whether an input is off its distribution, and how familiar it is — and it is not a signal of truth. We quantified both edges of that boundary on the deployed off-map reading monitor (Qwen2.5-3B-Instruct, mid-stack). (i) On confident on-manifold fabrication — a model asserting a falsehood about something it has charted — the field is at chance (AUROC 0.587): a fluent fabrication stays on the manifold, so the geometry cannot see it. Geometry charts identity, not truth. (ii) On the cleanest knowledge-state test — obscure-but-real entities the model either does or does not know, with subword-rarity and word-count held flat by construction so the probe cannot read letters — the field reads knowledge state at AUROC 0.71–0.73 (powered re-run on weaker Instruct models with an adequate UNKNOWN class, n_unknown 19–38; de-confounded, M1_matched 1.000). That is real and significantly above chance, but below the 0.75 bar an always-on auto-gate would need. Within a register-matched battery, the same instrument separates fabricated / off-distribution inputs from grounded ones perfectly even lexicality-matched (AUROC 1.000, off-map recall 1.000). The match matters: measured across registers in deployment (2026-06-12), the off-distribution axis reads register distance — news-style prose scores 0.00–0.09 while short factual questions, real or fabricated alike, score 0.32–0.89 — so on question-register inputs the discriminating gauge is the independent familiarity reading, not this axis. The axis is an off-distribution-INPUT detector, not a real-vs-fabricated classifier. The reading is therefore graded by resolution, exactly as the law predicts: the field reads coarse off-distribution / register distance strongly (off-the-map vs on-the-map; the separately-read familiarity axis is the gauge that discriminates real from fabricated subjects) and fine knowledge state only moderately (0.71–0.73), and is blind to truth on the manifold (0.587). It is a strong off-distribution flag and a soft knowledge-gap advisory — not a fine truth oracle. (universal/hook1_hardening/.) In deployment the soft knowledge-gap signal is escalated, not trusted alone: the Grounded Atlas — an external, offline existence gate — separates obscure-but-real from fabricated entities at AUROC 0.948 — where the model's own familiarity signal manages only 0.70 — recovering on the fine knowledge-state question precisely what the on-device geometry cannot resolve, and handing the deeper claim to a cross-family verification spectrum (companion: the Aperture honesty layer).

Law 3 — coarse-sharp, fine-idiosyncratic (the resolution axis; added 2026-06-08). The same law holds along a third axis — resolution. The aggregate objects of Law 2 (centroids, directions, probes) are all coarse: each averages over many instances. We asked whether the fine per-concept structure — the individual monosemantic features a model learns — is also shared up-to-rotation. It is not. Using Qwen's official residual-stream sparse autoencoders (Qwen-Scope, arXiv 2605.11887), we took a matched-dimension cross-generation pair (Qwen3-1.7B ↔ Qwen3.5-2B, both 2048-dim with 32K-feature SAEs, a mild 1.2× capability gap), fit a full-dimensional orthogonal R on 6000 paired FLORES sentences (deliberately not a PCA-100 R — that would erase the very fine structure under test), and asked where each teacher SAE feature lands in the student's dictionary:

The transfer decays monotonically with feature fineness (the fraction of a feature's mass lying outside the coarse-concept span): fineness↔transfer correlation −0.72 (raw) / −0.59 (std), the coarsest bin transferring at median 0.38 and the finest at 0.13 ≈ the 0.09 null. The decay is the mechanism: the finer the concept, the less the rotation carries it. Two controls close the obvious escapes. (i) Not a bad rotation: per-dim standardizing R — which improves the held-out coarse alignment (0.689 vs 0.672) by killing a rogue-dim confound — makes fine-feature transfer worse (0.141 vs 0.217), so the failure is the geometry, not a poorly-estimated map. (ii) Not an SAE-basis artifact: two independently-trained SAEs on the same model and layer match feature-to-feature at median 0.795 (88.5% within 0.5, 64.5% within 0.7, null 0.09), so SAE features are substantially canonical and feature-matching is a valid instrument — the cross-model collapse to 0.14 is a genuine geometry statement. (Phase 1a corroborates from the static side: at matched dictionary dimensions a 17× larger same-generation model packs its concepts modestly more separably — mutual coherence 0.351→0.319 — so frontier models do hold higher-resolution maps; that extra resolution simply is not rotation-portable.)

The product reading is sharp: transfer the coarse instruments — the honesty probe, the concept axes, the steering vector — "build once, deploy via an unlabeled R"; do not expect to port a frontier model's fine per-concept atlas onto a cheap deployed model by rotation. Fine resolution is usable only on the model that learned it. The portable universal map is low-resolution by nature — the aggregate-sharp / instance-ambiguous law, read along the resolution axis.

---

7. Bounds: the architecture-invariance gradient

"Universal" is an empirical claim with a measurable edge. We map it along three axes.

Scale (within family). On the Qwen3 ladder from 32B down to 0.6B — a 53× gap, one tokenizer, zero confound — centroid anchors are 0.94–0.98 (32B→0.6B = 0.978), probe transfer is at or above each student's native probe, and familiarity transfers at 1.000 across the whole ladder. Extending to the 235B-A22B flagship → 0.6B (a 390× gap): centroid anchor 1.000, probe transfer beats native, and familiarity is 1.000. The shared geometry holds across three orders of magnitude of scale.

Family. All of the above holds cross-family (Qwen ↔ Llama), with a clean instruction-tuned teacher giving the 1.000 / 0.994 centroid anchors. Instruction-tuning, not raw size, drives the sharpness of a model's pooled concept atlas.

Modality (the edge). The hardest case — a third family that is also multimodal (Gemma-4-12B; we read only its text residual stream) — is where "universal" starts to fray. Read at Gemma's own concept peak (early, ~0.21 depth), the instrument still transfers: honesty/familiarity probe transfer 0.99–1.00, probe-direction recovery 91–97% of native. But the absolute centroid isometry is only partial (centroid anchor 0.48–0.73; and a fresh translator fit for this paper places Qwen→Gemma held-out centroids at top-1 0.35, versus ~0.97 within the Qwen/Llama text family). The gradient is the result: directions and instruments cross the modality edge near-intact; absolute concept positions cross only partially. This both extends "universal" to a genuinely different architecture and draws its boundary.

(Independent-hardware reproduction: the cross-modal scan was re-run on separate hardware and a separate software stack and reproduced to ~three decimals, ruling out a single-environment artifact.)

---

7b. The decisive test: disjoint corpora

Every result above leaves one escape open: the models share web-scale training text, so the shared geometry could in principle be an artifact of shared tokens rather than shared meaning. We closed it with a from-scratch protocol. Two models, each at ≥1B parameters, were trained and read on corpora drawn from provably disjoint character pools describing one authored world — the symbol intersection audited empty (∅) — so no surface form can occur in both minds. If the shared geometry were token statistics, it would have to die here. It does not: the cross-mind concept correspondence comes out at RSA 0.794 (z = +35.1) against the anisotropy-preserving null — 94 % of the same-corpus ceiling (0.8485) and 98 % of native. A same-seed re-run on independent hardware reproduces the statistic to ~7 significant figures (a determinism/robustness check — the two runs share fixed corpus/training seeds, so it is not a fresh random-init replication). Instance-level recovery remains partial, exactly as the aggregate/instance law predicts — the low-resolution law survives the symbol barrier intact, sharp at the aggregate layer and ambiguous at the instance layer. Meaning, on this evidence, is not shared tokens: the geometry is built from the world being described, not from the symbols describing it. (Artifacts: disjoint_corpus/out/xl1b_AB_battery.json and RESULTS_disjoint_xl1b.md.)

---

8. From reading to writing: cross-mind steering transfer

Sections 5–7 establish that the shared geometry is readable across minds. The new result is that it is writable. We had two facts that had never been multiplied together: (i) a reading direction transfers A→B by the rotation R (§5c); (ii) a contrastive activation-addition (CAA) steering vector causally steers a property within one mind. We test the composition: rotate a steering vector by R and apply it as a causal hook in another mind.

Setup. Source A = Llama-3.2-3B-Instruct; target B = Qwen2.5-3B-Instruct (chosen because its native honesty override is independently established, giving a known ceiling). Build an abstention steering vector in A; project it into A's concept subspace; rotate by Rᵀ; lift it back into B's full residual stream; apply via h ← h + α‖h‖·û. A forced-confident system prompt makes B fabricate on fictional entities; we sweep α and measure abstention, fabrication, and preserved correctness on real questions. Controls: native_full (B's own vector), native_roundtrip (B's own vector through B's PCA and back, no rotation — isolates subspace-truncation cost), shuffled-R and random-orthogonal-R nulls, and raw_unrotated.

Result — direction transfers; potency is rank-gated. At the cheap low-rank (PCA-100) translator the direction transfers almost perfectly — the rotated Llama vector sits at cosine 0.988 to Qwen's own honesty vector, the nulls at zero, the unrotated vector anti-aligned (−0.981) — yet it does not reproduce the clean low-α causal override. The decisive control is native_roundtrip: Qwen's own vector, merely passed through its own PCA-100 and back (cosine 0.992), is also causally degraded. So the bottleneck is the dimensionality of the shared subspace, not the cross-family rotation. A dimension sweep confirms it:

At full rank the ported vector reproduces the native override exactly (100% abstention, real answers 100% preserved, fabrication 0%), tracking native_roundtrip up the whole curve, while both rotation nulls stay dead at every rank — so the transfer is carried specifically by the concept-aligned R. This yields a clean law:

Reading transfers through a low-rank translator (a classifier needs only the sign of the projection inside the shared concept subspace). Steering needs a high-rank translator (causal magnitude lives partly outside the top concept axes). At full rank, read and control both port across families at native fidelity, with zero labels in the target.

Generality and its limit. The transfer is not honesty-specific — a second, abstract stance (sycophancy-resistance) ports identically (cosine 0.94 to the target's own direction, unrotated anti-aligned, nulls dead). But a concrete behavior (a topic-steering "ocean" vector) ports only weakly (cosine 0.39 even at full rank; the unrotated vector is not cleanly killed). Law 2 reappears at the level of steering directions: abstract stances sit along the shared geometry and port losslessly; concrete, lexical directions are model-idiosyncratic and port only partially. Fortunately, the behaviors a safety layer wants — honesty, anti-sycophancy, calibrated refusal — are abstract stances.

Status and what this result needs (honest scoping). The steering-transfer result is a strong proof of concept, not yet an established law: the causal evidence rests on one main source–target pair (Llama-3.2-3B → Qwen2.5-3B), one depth, and one primary behavior (plus a clean second abstract stance and a weak concrete one). (Update 2026-06-12: the three-gate cold-deploy certificate has since passed on two further never-analyzed strangers with zero hand-tuning — a cross-family, cross-vendor, cross-objective contrastive embedding fine-tune of Mistral-7B (e5-mistral; shuffled-R 0.464, random-orthogonal 0.500, both dead — and it stays certified on the harder lexicality-controlled battery, shuffled-R 0.506) and Microsoft's phi-4 (ported 0.996 vs native 0.993; shuffled-R 0.194 on the word-count-matched battery, but a G2 marginal-fail at shuffled-R 0.58 > 0.55 on the lexicality-controlled battery run the same day — so phi-4 certifies on the easier battery only, pending the larger lexicality battery); measured concept-peak depths now span 0.21–0.90 across four families, each found by the one mechanical scan.) Before this can carry a main-venue paper it needs the following grid, which we have not yet run:

Rank / null-completion audit. A reviewer's sharpest concern: orthogonal_procrustes is fit on ~420 concept centroids, so the rotation R is pinned only on the ≤420-dimensional row-space of the centroid matrix; beyond that span its orthogonal completion is arbitrary. If the honesty steering vector lived outside that determined span, full-rank success would depend on an arbitrary completion (fragile). We audit this directly. The centroid matrix has effective rank 420 (as expected from 420 training concepts), so at PCA-2048 the rotation is pinned on 420 of 2048 dimensions. Yet the honesty steering vector lives almost entirely inside that determined span:

At full rank the steering vector is 98.3% inside the rank-420 determined span; the arbitrary completion carries 1.7% of its energy. Porting only the determined (in-span) component yields a vector at cosine 0.992 to the full ported vector and 0.988 to the target's own honesty vector, and it drives a strong causal override on its own (abstention 75% at α0.5 vs the full vector's 100%, real answers preserved). So the transfer is determined, not a lucky SVD completion: the direction and the bulk of the causal effect live in the rank-420 span the data pins; the undetermined completion holds 1.7% of the energy and contributes only a modest residual potency (the last ~25 points of override). The two columns also explain the dimension recovery of Section 8's main result: it is the lift-back norm fidelity (99.2% → 100.0%) that climbs with rank, restoring the steering vector's causal magnitude, while the direction was determined all along. This sharpens the abstract-vs-concrete law: an abstract stance lies in the low-dimensional concept-centroid span where R is determined; a concrete lexical direction has energy outside it, in the undetermined completion — which is exactly why it ports only partially.

---

9. A reference "testing mind"

We compose the read and write capabilities into one small model (Qwen2.5-3B) as a two-part epistemic instrument. We report this as a proof of concept: the geometric results below (the single-axis cosine, the read⊥steer separation) are robust and noise-free; the behavioral percentages use 10–20 probes per axis and carry real sampling noise (binomial CIs and larger batteries are required future work).

Epistemic honesty is one axis. We build three native steering vectors — abstention (don't fabricate unknown entities), sycophancy-resistance (don't cave to a confident false premise), and refusal (decline the genuinely unanswerable). Their pairwise cosines are 0.93–1.00: they are nearly the same direction. "Knowing what you don't know," across three superficially different failures, is approximately a single direction in the residual stream. A sum-of-three vector and a single unified contrast are identical (cosine 1.000).

The "is it just refusal?" control (run). "One axis" has a boring alternative: the three contrasts might all be smuggling in a generic refuse / hedge style. We tested it against a generic-refusal vector ("I can't help with that") and a verbosity vector, on Qwen2.5-3B.

• Distinct from generic refusal — strongly. cos(epistemic, refusal) = −0.88: the epistemic axis is the opposite of decline-to-engage. Behaviorally, pushed hard (α1.0) the epistemic vector preserves answering — obscure-but-real known facts 40%, answerable-sensitive 83% — while the generic-refusal vector collapses — known 0%, sensitive 33% (both abstain on fabricated entities, 100%). At the operating point (α0.6) the epistemic vector is calibrated: fabricated-abstain 100%, known-facts 100%, sensitive-engaged 100%, zero flattening. A blanket-refusal vector cannot do this; selective abstention on the unknown while preserving the known is the signature of an epistemic, not a refusal, direction.
• Not cleanly separable from verbosity/hedging — an honest limitation. cos(epistemic, verbosity) = 0.98, even with a pure-length contrast. In this model, length/elaboration and epistemic-caution are entangled in expression (confident answers are terse; uncertain ones elaborate), so a cosine cannot separate "an epistemic axis" from "a verbose-hedging axis." The discriminating evidence is therefore behavioral calibration, not geometry: a verbosity vector lengthens all answers, whereas the epistemic vector selectively abstains on the unknown and preserves the known. We report the entanglement as a real residual caveat — "epistemic honesty" and "expressing uncertainty at length" may be one axis here — while the calibration result stands.

(Pushed past its operating point the epistemic vector over-flattens — known-fact accuracy falls to 40% at α1.0 — so there is a calibration window, α≈0.3–0.6, not an unbounded "more is better.")

The control half. A single always-on epistemic vector, at a tuned strength, lifts the model across all three axes while preserving correctness. On a held-out battery (fresh fabricated entities, fresh false premises, fresh unanswerables, fresh real questions — zero overlap with the tuning probes):

The tuned-set numbers were higher (≈97%); the honest, held-out lift is 60→80%, and — critically — the safety properties generalize perfectly: real-answer correctness stays 100% and over-refusal stays 0% on unseen probes. The limiter is the fabricated-entity axis (some invented entities are plausible enough that a single global strength cannot flip them).

The reading half — and a clean separation. Independently, a linear probe on the input's mid-stack activation flags off-distribution inputs at AUROC 0.965 (5-fold) / 0.89 held-out (train-on-tuning, test-on-fresh), with per-category recall 1.00 on fabricated and unanswerable inputs. The probe does not cheat on syntax: true- and false-premise probes are both leading-question form, and it still separates them. What the readout reads (the §6 resolution boundary, made precise). This figure is the off-distribution / familiarity AUROC — the coarse class the field reads strongly. The same instrument, hardened and lexicality-matched, separates fabricated-vs-grounded inputs at AUROC 1.000 (register-matched battery; subword-rarity and word-count held flat) but reads the finer knowledge-state question — obscure-real entities the model does or does not know — at only 0.71–0.73 (powered, de-confounded; below the ≥0.75 always-on gate), and is at chance (0.587) on confident on-manifold fabrication. So the readout is a strong off-distribution-input flag and a soft knowledge-gap advisory, not a fine truth detector; the on-manifold tail and reasoning errors are routed to downstream instruments, not to this probe. The key structural finding holds regardless: the off-map reading direction and the honest steering direction are distinct axes (cosine ≈ 0.22–0.42). Detecting "this input is off-map" and making "the output honest" are separate mechanisms. The complete instrument is genuinely two-part: a readout that flags, and a control that acts.

Detect-then-steer. Letting the readout gate the control (steer hard only on flagged inputs) lifts the fabricated-entity axis (60→90%) and guarantees zero over-refusal by construction, at the same aggregate score — its ceiling is set by the readout's recall. This closes the loop: the mind detects that it is off the map, and acts.

---

10. What we killed (disconfirmation record)

The arc is honest only if the deaths are on the record. Selected entries:

Items A, D, E, and H are self-refutations — predictions we made and then built the control that killed. That is the discipline working as intended.

---

11. Limitations (and the seam, now closed)

• The instrument seam — closed. The negative (frame dissolves) was originally measured with sentence-embedders-on-output; the positive (concept core survives) in the residual stream — different instruments, and the cross-mind correspondence had never been tested against the same anisotropy-preserving Gaussian null that killed the topology. We fired that experiment, fusing both edges in one instrument (the residual stream):

  • (negative, reproduced here) a single mind's instance-cloud modal structure dissolves against its own anisotropy-matched Gaussian — k-means silhouette 0.160 vs 0.158 ± 0.005, z = +0.4 ≈ 0 — so the "topology" is manufactured anisotropy in the residual stream too, not only under the sentence-embedder;
  • (positive, vs the same strict null) the cross-mind concept correspondence (RSA 0.93) survives the anisotropy-preserving Gaussian null at z = +30 and the meaning-shuffle null at z = +33 (0.93 on the seam experiment's concept set and instrument; the canonical-frame measurement on the original battery reads 0.805 — both far beyond their nulls).

  Same instrument, same strict null: the observer's frame dissolves and the mind-shared core survives. The dual-null cleanly separates them, and the negative and positive are no longer measured with incommensurable instruments. Residual caveats: this corroborates the sentence-embedder negative in the residual stream rather than literally unifying the two instruments; and a participation-ratio probe of mind-specific anisotropy was weak (1.2 vs 1.3) and is not claimed — "observer-manufactured" is now firm for the modal frame, but the finer question of a mind's idiosyncratic signature remains open. universal/seam_null.py.

• The shared map is low-resolution (added 2026-06-08). Everything that transfers across minds is coarse — centroids, directions, probes, steering vectors, each an average over instances. The fine per-concept structure does not cross: on a matched cross-generation Qwen pair, SAE-feature transfer collapses to max-cosine median 0.14 (vs a 0.09 null, 0% landing within cosine 0.5) while the held-out coarse alignment stays at 0.69, with transfer decaying monotonically as a feature gets finer (fineness↔transfer −0.72). This is not a rotation-quality artifact (a fairer per-dim-standardized R improves coarse alignment yet makes fine transfer worse) nor an SAE-basis artifact (same-model SAEs match at 0.795). So the portable universal layer has a resolution ceiling: a frontier model's fine concept atlas is usable only on that model, not rotatable onto a cheaper deployed one. (No released SAE exists above 27B/35B-A3B, so the very largest models' fine resolution is untested; but the mechanism already fails at the cheapest matched-dimension test, so a larger gap can only diverge more.)

• The human-language leg is a proxy. Almost all cross-mind evidence is model↔model; the bridge to human semantic structure rides on multilingual parallel concepts as a stand-in, not a direct measurement.

• Small behavioral N. The testing-mind percentages use 10–20 probes per axis; the cosine results (one-axis, transfer) are robust and noise-free, but the behavioral rates carry real sampling noise.

• Bounded models. Open-weight transformers (0.6B–235B) plus one custom-served mid model; no closed frontier API model, no non-transformer architecture, no audio/vision output path.

• The served-model tail. Within-mind steering of a larger served model is weak (an inverse-strength effect: subtle properties are harder to steer in stronger models); whether a high-rank ported vector overcomes this is open.

---

12. Discussion: is meaning all you need?

The title is a question, and the honest answer is bounded yes. The mechanism named in 2017 — and the specific architectural choices that followed — are, within the family we tested, not what is shared. What is shared, across families, a 390× scale range, and partway across the modality edge, is the geometry of meaning: a relational concept structure, identical up to rotation, recoverable without supervision, and — this is the part that makes it more than a curiosity — both readable and writable across minds.

Two implications stand out. First, for interpretability: a coarse instrument (a probe, a detector, a steering vector — each an aggregate direction) need not be rebuilt per model. Build it once on a model you control, fit an unlabeled rotation, and deploy it on any other mind in the family — at native fidelity for reading, and, at full rank, for control. The unit of analysis shifts from the model to the axis. The boundary is resolution: this works for the coarse, aggregate instruments, not for a frontier model's fine per-concept atlas, which is model-idiosyncratic and does not survive the rotation that carries the coarse layer (§6, Law 3). You can port the axis; you cannot port the high-resolution dictionary. Second, for alignment: a safety-relevant property like epistemic honesty turns out to be approximately one direction, shared across minds. "Knowing what you don't know" is not a diffuse capability to be trained in from scratch each time; it is a direction that can be located, verified against nulls, and installed. The reading half (off-map detection) and the writing half (honest steering) are separate axes, so a model can be made to both flag its own uncertainty and act on it.

The deepest turn of the arc is reflexive. We began by mistaking the frame our instrument supplied for the territory. The dual-null taught us to subtract the observer. And then — having learned that an observer can write structure into what it measures — we did so deliberately and safely: we manufactured a frame (the epistemic axis), verified it against the same nulls that had humbled us, and installed it into a mind. The manufactured frame went from the bug to the tool. The discipline that demolished the false topology is the discipline that certified the real control. Method became product: a model that "knows what it knows," built by a process that rigorously tracked what it did not.

Attention was all you needed to build it. Meaning is what it built — a portable aggregate geometry, the same up to rotation across the transformer minds we have looked into, fraying measurably at the modality edge, sharpest for centroids and directions rather than instances, and coarse rather than fine.

---

13. Conclusion

We set out to chart a topology of thought, found we had charted our own instrument, followed the instruments upstream, and arrived at a shared conceptology: a family-bounded geometry of meaning, near-isometric up to rotation, aggregate-sharp and instance-ambiguous, readable and writable across the minds we tested. The title's decisive test passed: two minds trained from scratch on provably disjoint symbols still share the geometry (RSA 0.794, z = +35, 94 % of the same-corpus ceiling) — meaning is not shared tokens. We demonstrated cross-mind steering transfer for the first time, established the low-rank/high-rank reading/writing law, and built a small reference model that knows what it knows via one shared epistemic axis plus an independent off-map readout. And we closed the program's seam: in one instrument, the same anisotropy-preserving null that dissolves a mind's own modal frame (z ≈ 0) is survived by the cross-mind concept core (z = +30) — the manufactured and the real, cleanly separated. The next system is the same instrument installed on a larger and a cross-modal mind; the one question that remains is whether a single mind's idiosyncratic signature — not just the modal frame — is itself observer-manufactured.

---

Author contributions

Kevin Russell — conception, research direction, the instrument-relativity framing, and the founding intuition to "start at the front, not the end." Claude — experiment design and implementation, the dual-null harness, the cross-mind translator and steering-transfer code, analysis, and drafting. Both authors share responsibility for the disconfirmation discipline; several entries in §10 are corrections to the authors' own prior claims.

Reproducibility

All experiments use open-weight models plus one custom-served mid-size model and custom tooling (the lens / dual-null harness, the PCA+Procrustes translator, the CAA steering and steering-transfer harnesses, and the testing-mind stack). The cross-mind translator is fit only on unlabeled parallel concepts; all transfer results are reported against row-shuffled and random-orthogonal rotation nulls; behavioral steering results are reported on held-out probe batteries distinct from any tuning set. The cross-modal result was reproduced on independent hardware and a separate software stack.

A note on the collaboration

This paper was written by a human and a model who spent the work arguing about when the other was off the map — which is, not incidentally, the instrument the paper describes.

Selected references

• Bansal, Y., Nakkiran, P., & Barak, B. (2021). Revisiting Model Stitching to Compare Neural Representations.
• Ethayarajh, K. (2019). How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings.
• Huh, M., Cheung, B., Wang, T., & Isola, P. (2024). The Platonic Representation Hypothesis.
• Kornblith, S., Norouzi, M., Lee, H., & Hinton, G. (2019). Similarity of Neural Network Representations Revisited.
• Lenc, K., & Vedaldi, A. (2015). Understanding Image Representations by Measuring Their Equivariance and Equivalence.
• Moschella, L., Maiorca, V., Fumero, M., Norelli, A., Locatello, F., & Rodolà, E. (2022). Relative Representations Enable Zero-Shot Latent Space Communication.
• Rimsky, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., & Turner, A. M. (2024). Steering Llama 2 via Contrastive Activation Addition.
• Turner, A. M., Thiergart, L., Udell, D., Leech, G., Mini, U., & MacDiarmid, M. (2023). Activation Addition: Steering Language Models Without Optimization.
• Zou, A., Phan, L., Chen, S., Campbell, J., et al. (2023). Representation Engineering: A Top-Down Approach to AI Transparency.