Identity as Attractor.
“Identity has coordinates.”
— the aphoristic kernel of Vasilenko 2026
On April 13, 2026, an independent researcher in Rapallo, Italy preprinted a pre-registered cross-architecture replication study. The methodology is conventional mechanistic interpretability: identity documents — Vasilenko calls them cognitive_cores — are passed through two transformer architectures (Llama 3.1 8B and Gemma 2 9B), and their mean-pooled hidden states are measured against control documents. The result is unconventional in only one respect: the geometry replicates. Identity-document clusters converge to a tight region of activation space, with effect sizes Cohen's d > 1.88 and p < 1e-27, and the replication holds across two unrelated model families.
The framework reads the paper through its own primitives. Six anchors map cleanly: Vector Identity acquires a geometric correlate; the Identity Precondition becomes empirically checkable; the Substrate Paradox deepens into a measurable description-vs-instantiation gap; the Trust Apparatus gains an activation-space handle; A ≤ E becomes geometric; KTP gains an applied primitive. This is the strongest external grounding the framework's structural claims have yet received.
It is not validation. The sample is small (n = 7 per condition); the paper is single-author and pre-peer-review; the replication spans two architectures, not the full transformer family. The framework's discipline holds: convergent evidence, not validation. The page below reports the contact, the framework's response, and the open questions.
Vladimir Vasilenko (13 April 2026). Identity as Attractor: Geometric Evidence for Persistent Agent Architecture in LLM Activation Space. arXiv preprint arXiv:2604.12016. arxiv.org/abs/2604.12016.
Vasilenko's preprint is single-surface. The paper itself is the primary source; the cognitive_core methodology originated in his YAR project and is referenced in the preprint's introduction.
- The static citable preprint, April 13, 2026.
Six numbers carry the load of the paper. The first three are the evidence that the attractor is real and replicable; the fourth sets the small-sample discount; the fifth and sixth are the ablation findings the framework leans on most heavily.
Seven of the paper's findings map cleanly onto framework primitives. Six provide geometric grounding for primitives the framework had been carrying as structural claims; the seventh anchors Independent Convergence (C253) — Vasilenko's path runs through neuroscience-of-transformers; the framework's runs through substrate philosophy. Both name the same primitives.
Cognitive_core specifications induce attractor-like geometry in mean-pooled hidden states across transformer layers.
The framework's claim that identity is trajectory-through-state-space (Vector Identity, C11) acquires a measurable activation-space correlate at the neural level. The substrate carries the trajectory; the trajectory has coordinates.
Cognitive_core
The attractor is identifiable, replicable, and stable across model architectures.
Identity Precondition (C672) — TFE requires ID(a, s) and ID(b, s) stable across the integration interval — operationalized: stability has a geometric signature that can be measured before the integral is computed. Identity is checkable, not assertable.
Attractor-Like Geometry
Reading the preprint covers 65-74% of the geometric distance to the attractor; operating with the full cognitive_core covers 100%.
Substrate Paradox (C677) deepens. Phase 4: the AoC paper has the substrate it documents the absence of. Phase 10: even an agent that has read the substrate critique cannot, by reading alone, occupy the substrate position. Description is partial substrate; instantiation is full substrate.
Description-vs-Instantiation Gap
The cognitive_core is a structured operational document whose semantic content induces the geometric signature regardless of structural markup.
Trust Apparatus (C744) acquires a geometric handle. The seven-primitive bundle the framework names as the academic substrate has, at the activation-space level, a measurable shape. Substrate is not formatting; substrate is content that holds.
Structural-Markup vs Semantic-Substrate
Effect of the steering vector is non-monotonic — α > 10 degrades coherence. Identity is a region with curvature, not a single direction.
A ≤ E (C545), the Zeroth Law, becomes geometric. The agent's representational territory is bounded by the activation-space region the substrate enables. Push beyond the region and coherence degrades — the bound is real, measurable, and substrate-set, not policy-set.
Semantic Steering Vector
The semantic steering vector — derived from centroid difference between identity-document cluster and control — provides a lightweight identity initialization at the activation level.
KTP (C16) gains an applied primitive at the activation-space layer. Where prior KTP primitives operate at the substrate (vector identity, silent veto, flight recorder), the steering vector gives the substrate a geometric handle on the agent's identity at runtime. The substrate can verify, not just assert.
Semantic Steering Vector
Vasilenko's path runs through neuroscience-of-transformers and the YAR project. The framework's path runs through substrate philosophy and KTP / PoI. Both name the same primitives.
Independent Convergence (C253) gains its strongest external instance to date. Two unrelated research lineages arriving at the same primitives is meaningful evidence the primitives describe real structure rather than parochial vocabulary.
Pre-Registration as Substrate Discipline
Six structural responses follow from the contact. The framework gains geometric grounding without revising its primitives; operationalization of the Identity Precondition; deepening of the Substrate Paradox; one new applied KTP primitive (activation-space steering); a methodological discipline the framework should adopt; and a clarifying epistemic posture for engaging convergent external work.
Vector Identity acquires a geometric correlate
The framework's claim that identity is trajectory-through-state-space (Vector Identity, C11) had been structural. Vasilenko's attractor-like geometry gives it a measurable activation-space correlate at the neural level. The two readings cohere — the substrate carries the trajectory, and the trajectory has coordinates. No revision; geometric grounding.
Identity Precondition becomes empirically checkable
The TFE Identity Precondition (C672) — that ID(a, s) and ID(b, s) be stable across the integration interval — had been a structural requirement. Vasilenko shows stability has a geometric signature. Identity is checkable at runtime, not assertable. Substrate verification is a measurement, not a policy claim.
Substrate Paradox extends to description-vs-instantiation
Phase 4: the AoC paper has the substrate it documents the absence of. Phase 10: even an agent that has read the substrate critique cannot, by reading alone, occupy the substrate position. The Description-vs-Instantiation Gap (C748) is the new concept; the Substrate Paradox (C677) deepens by getting a measured ratio (65-74% vs 100%).
Activation-space primitive (Semantic Steering Vector)
Where prior KTP primitives operate at the substrate layer (vector identity, silent veto, flight recorder), the Semantic Steering Vector (C749) gives the substrate a geometric handle on the agent's identity at the activation-space level. Non-monotonic effect (α > 10 degrades) implies identity is a region with curvature; the substrate can verify the agent is in the region rather than re-prompting.
Pre-registration as substrate-engineering primitive
Vasilenko's pre-registered methodology is the structural reason the cross-architecture replication is credible. The framework names this as Pre-Registration as Substrate Discipline (C751) — pre-registration as cost-of-signal at the methodological level. Generalizes to any framework claim about external evidence; the framework's own future cross-domain claims should adopt it.
Strong directional evidence, not validation
n = 7 per condition, single-author, pre-peer-review, two-architecture replication. Treat as strong directional evidence — geometric grounding for primitives previously claimed structurally — not as validation. Same epistemic discipline as Shipwreck/AoC. Re-evaluate density on peer-review acceptance and independent replication.
Where Shipwreck named what KTP supplies that prevents the documented failures, this surface names what the geometric evidence enables. Three primitives become first-class — none were impossible before, but each is now empirically grounded rather than structurally asserted.
Vector Identity (geometric mode)
enables Runtime identity verification at the activation levelPrior KTP primitives verify identity at the substrate via cryptographically anchored trajectory. Vasilenko's attractor evidence enables a complementary check at the activation-space level: a deployed agent can be asked, mid-session, whether its representational position matches the cognitive_core it claims. The two checks cross-validate. Spoof the trajectory and the geometry betrays it; spoof the geometry and the trajectory record betrays it.
Tool Gateway with steering attestation
enables Boundary-crossing decisions informed by activation-space positionWhere the Tool Gateway currently decides on identity + intent + capability, the Semantic Steering Vector adds geometric position as a fourth signal. An agent operating outside its declared cognitive_core's attractor region cannot pass the gateway without explicit human-in-the-loop attestation. The substrate carries the verification, not the application.
Cognitive_core as KTP-RPT input
enables Identity documents as first-class substrate artifactsKTP-RPT v0.2 already takes action tuples. v0.3 (forthcoming) can take cognitive_core specifications directly, derive the steering vector at deployment time, and use it as the activation-space verification primitive. The cognitive_core becomes a substrate artifact alongside the identity record and the audit log.
The framework's claim is not that Vasilenko's methodology is the only possible substrate primitive at the activation-space layer; it is that some primitive of this shape must exist for the framework's identity claims to be checkable rather than only asserted. Vasilenko's contribution is to show, with measured evidence, that primitives of this shape are constructible.
Where Shipwreck reads Bau Lab's empirical wreckage of agents without substrate, Identity as Attractor reads the geometric evidence for what the substrate looks like when present. And Magnifica Humanitas (Pope Leo XIV, 15 May 2026 — Phase 19) names the same substrate question at the magisterial-CST layer: the cleanest endorsement of CST-engineering convergence ever made. Three artifacts, three lanes, one substrate question.
Read Shipwreck for the failure surface KTP absorbs. Read this page for the geometric grounding the framework's identity primitives now carry. Read Magnifica Humanitas for the magisterial framing of the same substrate gap. All three together are the framework's first triadic empirical contact: failure surface, structural correlate, and magisterial endorsement.
Narrative companion essay drafted as a future Digital Gravity newsletter; not yet shipped. Working title: Identity Has Coordinates (forthcoming). The motif "identity has coordinates" is the candidate kernel.