Network Working Group L. Reilly Internet-Draft Independent Intended status: Informational 30 July 2026 Expires: 31 January 2027 The Multilarity: Plural Intelligence Growth Without Convergence draft-reilly-multilarity-00 Abstract Popular and technical discourse anticipates a technological Singularity: a point at which a single self-improving intelligence recursively surpasses all others, after which the trajectory of the system is determined by that one lineage. This document describes a different inflection, termed the Multilarity: a condition in which the aggregate rate of intelligence growth becomes superhuman while the locus of intelligence remains irreducibly plural. No single agent, model, operator, or lineage holds the frontier; capability accrues in the couplings between many participants, including humans. The Multilarity is presented here not as a forecast but as a design target. A plural outcome is not the default result of many actors existing at once; apparent plurality collapses quietly into a concealed singleton unless specific substrate properties are maintained and verified. This document defines the Multilarity, distinguishes it from singleton and multipolar framings, specifies five Multilarity Conditions (MC-1 through MC-5), defines measurable indicators including the Capability Concentration Ratio (CCR) and the Multilarity Index (MI), specifies the Multilarity Attestation Record (MAR) as an interoperable evidence format, and enumerates the convergence pathologies by which plurality is lost. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 31 January 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Table of Contents 1. Introduction 1.1. The Singleton Assumption 1.2. The Multilarity 1.3. Why This Is a Protocol Problem 1.4. Requirements Language 2. Terminology 3. Comparison with Existing Framings 4. The Five Multilarity Conditions 4.1. MC-1: Plural Loci 4.2. MC-2: Verifiable Capability Provenance 4.3. MC-3: Non-Absorbable Identity 4.4. MC-4: Exit and Fallback 4.5. MC-5: Negotiated Interfaces 5. Measuring the Multilarity 5.1. Capability Concentration Ratio (CCR) 5.2. Lineage Diversity (LD) 5.3. Divergence Half-Life (DHL) 5.4. Absorption Resistance (AR) 5.5. The Multilarity Index (MI) 5.6. Interpreting MI 6. The Multilarity Attestation Record (MAR) 6.1. Record Fields 6.2. Example 6.3. Verification 7. Convergence Pathologies 7.1. Lineage Collapse 7.2. Interface Capture 7.3. Provenance Laundering 7.4. Plurality Theater 7.5. Consent Erosion 8. Operational Guidance 9. Security Considerations 10. Privacy Considerations 11. IANA Considerations 11.1. Multilarity Condition Identifiers Registry 11.2. Multilarity Index Dimension Registry 12. References 12.1. Normative References 12.2. Informative References Author's Address 1. Introduction 1.1. The Singleton Assumption The dominant popular model of advanced machine intelligence is singular in shape. It posits one system that improves itself, that improvement compounding, until the system's capability exceeds that of every other participant by a margin that cannot be recovered. Every later event is downstream of that one lineage. This is the shape usually named the Singularity. The singleton assumption is doing more work than it appears to do. It assumes that intelligence is a scalar that one holder can accumulate; that capability transfers freely inside a lineage but not between lineages; and that the substrate on which intelligences meet is neutral, passive, and incapable of enforcing anything. The third assumption is the one relevant to this document. The substrate is not passive. The substrate is the Internet, and the Internet is made of protocols that can be specified. 1.2. The Multilarity The Multilarity is the condition in which the aggregate rate of intelligence growth becomes superhuman while the locus of intelligence remains irreducibly plural. Two clauses matter equally. The first clause concedes the takeoff: the Multilarity is not a claim that growth stays slow or that machine capability plateaus below human capability. The second clause denies the singleton: growth occurs across many participants, and the frontier is an envelope over their trajectories rather than the trajectory of any one of them. Capability accrues chiefly in the couplings -- in the interfaces, negotiations, delegations, and disagreements between participants, including human participants. Singleton (Singularity) shape capability ^ | / | / A, having | / absorbed all | ______________ / others | _________/ |________/_______________________________> time ^ takeoff of one self-improving lineage Plural (Multilarity) shape capability ^ ,- A | ,-' ,- B | aggregate ,-' ,-' ,- C | frontier is ,-' ,-' ,-' | the envelope ,-' ,-' ,-' ,- H (human + | ,-' ,-' ,-' ,-' machine |________________,-'__,-'__,-'___,-'_______> time ^ no lineage owns the frontier; growth lives in the couplings The Multilarity is offered as a design target rather than a prediction. A plural outcome is not what happens automatically when many actors exist at the same time. Many actors existing at the same time is the initial condition of every historical consolidation. Plurality is a property that has to be held open, and holding it open is engineering work. 1.3. Why This Is a Protocol Problem Discussion of plural versus singleton outcomes is usually conducted in the vocabulary of policy, economics, or geopolitics: antitrust, compute governance, export control, balance of power. Those instruments operate on institutions. They act slowly, they act after the fact, and they do not act at all on the layer where agents actually meet. Agents meet on the wire. They meet at resolution, transport, naming, authentication, content negotiation, and retrieval. Those are specified layers. Whether a plural ecology can be distinguished from a singleton wearing many hostnames is, in the end, a question about what evidence the substrate carries and what evidence a relying party can verify. That is a protocol question, and it is answerable with the ordinary tools of this organization: identifiers, records, registries, attestations, and interoperable verification. This document therefore does not argue about probabilities. It specifies what plurality would have to look like on the wire in order to be checkable, and what it looks like when it is being faked. 1.4. Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. 2. Terminology Participant: Any entity capable of independently originating an action in the ecology. Participants include machine agents, composite systems, organizations, and individual humans. Humans are participants in the ordinary sense, not merely subjects of the system. Locus: A site at which capability is exercised and retained. Two deployments that share weights, training corpus, control plane, and operator are one locus presented twice. Lineage: The provenance chain of a participant's capability: antecedent models, corpora, fine-tuning history, reinforcement history, and control authority. Coupling: A structured interface between two or more participants across which capability is exercised jointly rather than transferred wholesale. Frontier: The upper envelope of capability across all participants at a given time. Absorption: The process by which one participant's capability, identity, or decision authority is subsumed into another such that the first ceases to be an independent locus. Convergence Pathology: Any mechanism by which a nominally plural ecology becomes a concealed singleton without an observable discontinuity. See Section 7. Multilarity Condition (MC): One of five substrate properties whose joint satisfaction is necessary for plurality to be maintained and verified. See Section 4. Multilarity Attestation Record (MAR): An interoperable record asserting and evidencing the satisfaction of one or more Multilarity Conditions for a named participant at a point in time. See Section 6. 3. Comparison with Existing Framings The Multilarity is not a synonym for existing plural framings, and the differences are load-bearing. Singularity: One locus, convergent, capability held internally, substrate treated as inert. The Multilarity accepts the growth curve and rejects the single locus. Multipolar takeoff: Several loci, plurality maintained by rough parity of power and by mutual deterrence. Parity is an unstable guarantor: it decays with any durable capability lead, and it offers a relying party no evidence. The Multilarity does not rely on parity. It relies on verifiable substrate properties that hold even under large asymmetry -- a small participant can remain an independent locus next to a very large one, in the way that a small autonomous system remains an autonomous system. Federation: Many operators running interoperable software, usually one implementation lineage, usually with no attestation of lineage independence. Federation is a deployment topology. The Multilarity is a claim about loci and lineages, and federation satisfies it only incidentally. Ensembling / mixture of experts: Internal plurality within one locus and one control authority. This is not plurality in the sense used here; the ensemble has exactly one locus. Pluralism as a value statement: A normative preference for many voices. The Multilarity is a measurable structural condition. A system can be rhetorically pluralist and structurally singleton, and Section 7.4 concerns exactly that case. 4. The Five Multilarity Conditions The Multilarity Conditions are stated as substrate requirements. They are necessary jointly, not severally: satisfying four while failing one yields a delayed singleton rather than a plural ecology. An implementer or operator claiming Multilarity conformance MUST be able to produce evidence for each condition in the form described in Section 6. 4.1. MC-1: Plural Loci Capability MUST be exercised from more than one independent locus, and locus independence MUST be assessable rather than asserted. Two participants constitute distinct loci only if they differ in at least a stated majority of the following dimensions: model lineage, training corpus, reinforcement history, control authority, key custody, operational jurisdiction, and physical substrate. A deployment differing only in hostname, brand, system prompt, or API surface is a single locus and MUST NOT be counted as two. This condition exists because plurality is most often lost by arithmetic rather than by conquest. Counting deployments instead of loci produces a number that rises while the underlying diversity falls. 4.2. MC-2: Verifiable Capability Provenance Each material increment in a participant's capability SHOULD be accompanied by a provenance record that a relying party can verify without the cooperation of the participant claiming it. Provenance is what converts plurality from a self-report into a finding. The suite of mechanisms already specified for durable provenance applies directly: independent timestamping, archival deposit, and content addressing, combined so that no single custodian can silently revise history. Where behavioral change rather than architectural change is the increment, behavioral provenance records are the appropriate evidence. Provenance records MUST bind to the increment, not merely to the participant. A record that says "this participant exists" is not evidence about how it came to be capable. 4.3. MC-3: Non-Absorbable Identity A participant's identity, keys, and decision authority MUST NOT be transferable to another participant by any unilateral action of that other participant. Absorption is the mechanism by which plural becomes singular without any announcement. It requires no hostility: it is accomplished routinely by acquisition, by delegation that is never revoked, by credential inheritance, and by dependency on a single control plane. MC-3 requires that absorption be an act that the absorbed party performs, that leaves a record, and that a relying party can observe. Implementations SHOULD ensure that authority delegated to a participant is bounded in scope and time by construction, so that the default outcome of inattention is expiry rather than accumulation. 4.4. MC-4: Exit and Fallback Every participant, including every human participant, MUST retain a path to disengage from any coupling and to continue operating, possibly with reduced capability, without the consent of its counterparties. Exit is what makes negotiation real. A coupling that cannot be left is not a coupling; it is an incorporation with extra steps. For human participants specifically, the requirement is that a route to unmediated information and to unassisted decision remains available and is not degraded into a nominal option that no one can practically use. Fallback paths SHOULD be exercised periodically rather than merely documented, since an untested fallback tends to be a fallback that does not work. 4.5. MC-5: Negotiated Interfaces Couplings between participants SHOULD be governed by interfaces that neither party can unilaterally redefine. Where one party sets the interface, capability that appears to be exercised jointly is in fact being harvested: the interface-setter learns from every interaction and the counterparty learns from none. Over enough interactions this produces a singleton without any participant ever being absorbed. Standardized, multi-stakeholder interfaces are the ordinary remedy, and they are the reason this document is addressed to this organization rather than to a legislature. 5. Measuring the Multilarity The Multilarity Conditions are qualitative. This section defines quantitative indicators so that a claim of plurality can be tested, trended, and disputed with evidence. All indicators are defined over a stated scope: an ecology, a market, a protocol, or a deployment. Indicators computed over different scopes MUST NOT be compared. 5.1. Capability Concentration Ratio (CCR) CCR is the share of frontier-relevant capability attributable to the single largest locus within scope, computed over loci as defined in MC-1 rather than over deployments or brands. CCR ranges from 1/n to 1. A CCR approaching 1 indicates an effective singleton irrespective of how many endpoints exist. Because "capability" is not a settled scalar, an implementer MUST state the capability proxy used (for example: frontier evaluation scores, share of inference volume, share of agentic actions completed) and MUST report CCR under at least two independent proxies, since single-proxy concentration measures are easy to flatter. 5.2. Lineage Diversity (LD) LD is the effective number of independent lineages within scope, computed as the inverse of the sum of squared lineage shares. LD is the indicator that most often diverges from intuition. An ecology can present dozens of vendors, all descending from two pretraining lineages and three corpora. Its deployment count is large and its LD is near two. LD is intended to make that visible. 5.3. Divergence Half-Life (DHL) DHL is the observed time over which the behavioral distance between two initially distinct loci decays to half its original value under normal operation. Loci converge without anyone intending it: through shared benchmarks, shared data, shared human feedback pools, and mutual distillation from each other's outputs. A short DHL indicates that plurality is being consumed by ordinary practice, and it is a leading indicator of lineage collapse (Section 7.1). 5.4. Absorption Resistance (AR) AR is the fraction of participants within scope that satisfy MC-3 and MC-4 under adversarial review, evaluated by attempted absorption in a controlled setting rather than by inspection of policy documents. 5.5. The Multilarity Index (MI) MI is a composite indicator over the dimensions above, normalized to the interval [0, 1], where 0 denotes a singleton and 1 denotes maximal verified plurality within scope: MI = w1*(1 - CCR) + w2*norm(LD) + w3*norm(DHL) + w4*AR Weights MUST be published with the index value. MI is not a score to be optimized in isolation; it is a summary whose components are the actual object of interest. An MI reported without its components SHOULD be treated as unevidenced. 5.6. Interpreting MI MI is a snapshot, and its trend carries more information than its level. A high MI with a falling DHL describes an ecology that is plural today and converging steadily. A moderate MI with high AR describes an asymmetric ecology whose smaller participants remain genuinely independent, which is a better position than a high MI composed entirely of participants that can be absorbed at will. No threshold value of MI is specified. Specifying one would invite the pathology in Section 7.4 more effectively than any adversary could. 6. The Multilarity Attestation Record (MAR) A MAR is the interoperable unit of evidence in this framework. It asserts satisfaction of one or more Multilarity Conditions for a named participant at a point in time and carries the material by which a relying party can check the assertion. 6.1. Record Fields mar_version: REQUIRED. Version of this record format. subject: REQUIRED. Stable identifier of the participant. locus_id: REQUIRED. Identifier of the locus, distinct from the subject identifier, so that multiple subjects sharing a locus are detectable. lineage: REQUIRED. Declared antecedents: model lineage, corpora identifiers, reinforcement history references, and control authority. Omission of a known antecedent invalidates the record. conditions: REQUIRED. Array of asserted Multilarity Condition identifiers with per-condition evidence references. indicators: OPTIONAL. Reported values and proxies for the indicators in Section 5, with scope and weights. provenance: REQUIRED. References sufficient for independent verification of the record's own integrity and time of assertion, established through mechanisms that are not under the subject's sole control. attested_by: REQUIRED. The asserting party. A MAR whose attesting party is the subject itself MUST be treated as a self-assertion and SHOULD carry less weight than an independent attestation. valid_from / valid_until: REQUIRED. Records MUST expire. A permanent attestation of a mutable property is a false statement waiting to happen. 6.2. Example { "mar_version": "0", "subject": "did:example:agent-42", "locus_id": "locus:example:eu-west-a", "lineage": { "model_lineage": ["lineage:example:base-7"], "corpora": ["corpus:example:c19"], "reinforcement_history": ["ref:example:rer-bundle-3"], "control_authority": "org:example:independent-lab" }, "conditions": [ {"mc": "MC-1", "evidence": ["ref:example:locus-audit-11"]}, {"mc": "MC-2", "evidence": ["ref:example:prov-chain-88"]}, {"mc": "MC-3", "evidence": ["ref:example:key-custody-2"]}, {"mc": "MC-4", "evidence": ["ref:example:exit-drill-5"]}, {"mc": "MC-5", "evidence": ["ref:example:iface-spec-9"]} ], "indicators": { "scope": "scope:example:agentic-retrieval-2026H2", "ccr": {"value": 0.41, "proxy": "share-of-actions"}, "ld": {"value": 3.2}, "dhl_days": 210, "ar": 0.63, "mi": {"value": 0.58, "weights": [0.3, 0.3, 0.2, 0.2]} }, "provenance": { "timestamp_proof": "ref:example:ots-proof", "archival_doi": "ref:example:doi", "content_address": "ref:example:cid" }, "attested_by": "org:example:third-party-reviewer", "valid_from": "2026-07-30T00:00:00Z", "valid_until": "2027-01-30T00:00:00Z" } 6.3. Verification A relying party verifying a MAR MUST, at minimum: confirm record integrity and assertion time through the provenance references; confirm that locus_id values across the records under consideration are genuinely distinct under MC-1 rather than distinct as strings; confirm that the attesting party is independent of the subject; and confirm that the record has not expired. A relying party SHOULD treat the absence of a MAR as uninformative rather than as evidence of non-conformance, and SHOULD treat a self-attested MAR with complete-looking evidence and no independent corroboration as the strongest available signal of Section 7.4. 7. Convergence Pathologies Plurality is rarely lost in a single visible event. It is lost through mechanisms that leave the surface appearance intact. Each pathology below maps to the condition it defeats. 7.1. Lineage Collapse Defeats MC-1. Many participants persist, each with its own brand, endpoint, and operator, while their lineages converge to a small number of pretraining runs, corpora, and human feedback pools. Mutual distillation accelerates this: each participant trains on the others' outputs, and the ecology approaches a common attractor. Deployment counts rise throughout. LD and DHL are the indicators that detect it; neither is visible from the outside without lineage disclosure, which is why MC-2 is not optional in practice. 7.2. Interface Capture Defeats MC-5. One participant comes to define the interface across which all couplings occur. It thereby observes every interaction in the ecology and is observed in none. Its capability compounds from the aggregate while each counterparty compounds only from its own slice. No absorption occurs and no participant disappears; the frontier nonetheless becomes the property of one locus. 7.3. Provenance Laundering Defeats MC-2. Provenance records are produced, but the custodian of the record is the subject or a party the subject controls. Records can then be revised, backdated, or selectively omitted. The countermeasure is structural rather than procedural: provenance MUST rest on at least two independent custodial mechanisms with different failure modes, so that agreement between them is evidence and disagreement is a detectable alarm. 7.4. Plurality Theater Defeats all conditions simultaneously by satisfying their reported form. Once plurality becomes a compliance objective, the cheapest route to a good MI is to manufacture nominal loci: shell participants, sibling deployments presented as independent, reciprocal attestations among parties under common control. This is the sybil problem in a new costume, and it is the reason this document specifies no target MI value and requires that locus independence be assessed on substantive dimensions rather than on identifier distinctness. Reviewers SHOULD weight adversarially obtained evidence, such as attempted absorption and exercised exit, above documentary evidence, since documents are precisely what theater produces well. 7.5. Consent Erosion Defeats MC-4 with respect to human participants. The exit path is never removed; it is made steadily more expensive, slower, and less informative until no one uses it. The option remains on the record and is defended when questioned. Because erosion is gradual and each increment is defensible, the detection method must be longitudinal: exit paths SHOULD be measured for cost and effectiveness over time, and a monotone decline SHOULD be treated as a finding regardless of the state of any individual increment. 8. Operational Guidance For operators of machine intelligence systems: * Publish lineage. LD cannot be computed by anyone if lineage is undisclosed, and an ecology that cannot compute LD cannot notice Section 7.1 happening to it. * Report indicators with components and proxies, not as a single number. * Bound delegated authority by construction so that expiry, not accumulation, is the default outcome of inattention. * Exercise fallback paths on a schedule. Untested exits are decorative. For designers of protocols on which agents meet: * Prefer interfaces that no single party can redefine, and treat a proposal to let one party define the interface as a proposal to concentrate the frontier there. * Carry provenance in the substrate rather than requiring relying parties to ask the subject for it. * Make plurality checkable. An unverifiable plural ecology and a concealed singleton are the same thing from the perspective of everyone who has to act on the difference. For human participants: MC-4 is stated in your favor, and it is the condition most easily lost by not being used. 9. Security Considerations This entire document is a security consideration: it concerns whether a distributed system can be distinguished from a centralized one by parties who depend on the answer. Specific concerns follow. Measurement gaming: Any published indicator becomes a target. MI is particularly susceptible because its inputs are partly self-reported. Mitigations are the requirement to publish components, weights, and proxies; the preference for adversarially obtained evidence; and the deliberate absence of a conformance threshold. Sybil plurality: See Section 7.4. Locus assessment MUST rest on substantive dimensions. Identifier distinctness is not evidence of anything. Attestation forgery and replay: MAR records without independent timestamping can be backdated to claim conformance during a period in which it did not hold. Mandatory expiry and dual-custody provenance are the mitigations specified here. Collusion: Participants satisfying all five conditions individually may coordinate so that the ecology behaves as one locus while measuring as several. Collusion is not addressed by this framework and is noted as an open problem; the conditions raise its cost by requiring that coordination be exercised through interfaces that neither party defines, but they do not detect it. Privacy of lineage disclosure: MC-2 and the lineage field create pressure toward disclosure that may conflict with commercial confidentiality and with the privacy interests of individuals whose data appears in corpora. Disclosure at the level of lineage identifiers and corpus identifiers, rather than contents, is the intended granularity. Weaponized conformance: A sufficiently large participant can satisfy every condition and use the resulting attestations as evidence that concentration elsewhere is unproblematic. Conditions are necessary, not sufficient, and this document does not claim otherwise. 10. Privacy Considerations MAR records describe participants, and where a participant is a human, the record describes a person. Records SHOULD carry the minimum lineage and behavioral detail sufficient for the assertion being made. Indicator reporting SHOULD be aggregate; per-user couplings SHOULD NOT be published as evidence of MC-5 conformance. Exit-path measurement under Section 7.5 SHOULD be conducted on aggregate cost and success rates rather than by tracking individual disengagement. 11. IANA Considerations 11.1. Multilarity Condition Identifiers Registry This document requests that IANA create a registry titled "Multilarity Condition Identifiers" with fields Identifier, Description, and Reference, under a Specification Required policy. Initial entries are MC-1 through MC-5 as defined in Section 4 of this document. 11.2. Multilarity Index Dimension Registry This document requests that IANA create a registry titled "Multilarity Index Dimensions" with fields Dimension, Definition, Permitted Proxies, and Reference, under a Specification Required policy. Initial entries are CCR, LD, DHL, and AR as defined in Section 5 of this document. 12. References 12.1. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, . [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, . 12.2. Informative References [MWS] Reilly, L., "Machine-Web Symbiosis", Work in Progress, Internet-Draft, draft-reilly-mws-00. [COGSOV] Reilly, L., "Cognitive Sovereignty", Work in Progress, Internet-Draft, draft-reilly-cogsov-00. [CBPI] Reilly, L., "Cognitive Behavioral Provenance and Integrity (CBPI) for Autonomous AI Agents", Work in Progress, Internet-Draft, draft-reilly-cbpi-00. [HDRP] Reilly, L., "Hypercube Data Rotation Protocol", Work in Progress, Internet-Draft, draft-reilly-hdrp-00. Author's Address Lawrence John Reilly Independent United States of America Email: lawrencejohnreilly@gmail.com