Internet-Draft Cognitive YANG-TMF Reconciliation November 2026
Janz & Peters Expires 21 May 2027 [Page]
Workgroup:
ONSEN
Internet-Draft:
draft-janz-onsen-cognitive-reconciliation-00
Published:
Intended Status:
Informational
Expires:
Authors:
C. Janz
Huawei Technologies
B. Peters
NBN

Cognitive Reconciliation for YANG and TM Forum Model Interoperability

Abstract

Network operators must make YANG-modeled service and network interfaces interoperate with TM Forum (TMF) OSS/BSS APIs. Today this is done with hand-built, model-specific adapters, whose construction and maintenance is ongoing human work. The ONSEN problem statement records the absence of a standard alignment between TMF APIs and YANG models as an open problem. This document applies ad hoc semantic reconciliation by cognitive agents, as developed in draft-janz-nmrg-adhoc-semantic-reconciliation, to that problem. Two cognitive agents lift their respective models, bind through a thin shared reference, and each generates and operates its own inbound translator, with no model-specific adapter agreed in advance. This document reports results from a demonstrator whose first case reconciles the L3SM service model to a TMF representation. It runs the reconciliation in two forms, one anchored on the mapping-root annotations of draft-lls-onsen-yang-tmf-mapping, and one with those annotations withheld, in which the agents must locate the services themselves in the plain model, and the two forms together size what the annotations contribute. It shows the reconciled mapping operating on real instance data, shows the model-drift lifecycle by which the translators are kept current, reads the demonstrator as an empirical point on a cognition-placement spectrum, from over-arching cognition over two inert systems to both systems cognitive, and is explicit about what over-arching cognition achieves when neither system is itself cognitive. A second case reconciles the L3SM service model to the L3NM network model, a cross-abstraction mapping where a cognitive counterpart earns its place, so the two cases together give the two poles of the spectrum on two ONSEN problem areas. The document closes with recommendations for operational practice and for standardization.

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 21 May 2027.

▲

Table of Contents

1. Introduction

The ONSEN problem statement [I-D.kbf-onsen-problem-statement] groups the working group's initial work into a small number of problem areas. One of them, "OSS/BSS Interface and API Interoperability", records that there is no standard northbound API semantics and no alignment of TMF640 [TMF640] and TMF641 [TMF641] with YANG models. The mapping work in [I-D.lls-onsen-yang-tmf-mapping] addresses this area by defining how YANG schemas and instance data map to TMF data models and APIs, and by providing a reference implementation.

A standardized mapping of this kind is valuable, and its value rests on a specific premise: that building and maintaining the adapters between a YANG-modeled interface and a TMF consumer is costly human work, and that a single agreed mapping removes that cost. The premise is sound where the systems on each side are conventional software that must be told, in advance, how the two models line up.

This document examines what changes when the systems are supported by cognitive agents, in the sense of [I-D.janz-nmrg-adhoc-semantic-reconciliation]. That work shows that machine reasoning can replace much of the advance agreement that network management has needed: each side lifts its own data into a semantic model, and two such models are reconciled ad hoc, with no shared model agreed beforehand, anchored on a thin shared reference. Applied here, the cost that motivates a standardized full mapping largely collapses. What remains worth standardizing is not the full mapping but the thin reference that anchors the correspondence, and the boundaries of authority over contested cases. This document describes the approach, reports concrete results from a demonstrator, and draws out the operational and standardization consequences.

The demonstrator covers two of the working group's problem areas. The first, OSS/BSS interface and API interoperability, is the YANG-to-TMF mapping of [I-D.lls-onsen-yang-tmf-mapping], a cross-format mapping that is largely determinate. The second, the structural LxSM and LxNM problems, is the reconciliation of the L3SM service model to the L3NM network model [RFC9182], a cross-abstraction mapping where service intent meets network realization. The two differ in how much cognition each needs, and taken together they show both what over-arching cognition settles on its own and where a cognitive counterpart becomes necessary. The second case sits at the both-cognitive end not by choice but of necessity: its divergences turn on facts and framing held only by the other system, which no reasoning over the static models can supply.

[I-D.janz-nmrg-adhoc-semantic-reconciliation] supersedes the conceptual treatment in the earlier drafts [I-D.janz-nmrg-ontology-reconciliation] and [I-D.janz-nmrg-reference-lexicons] and supplies their empirical grounding. This document uses its vocabulary throughout.

2. Terminology

The following terms are used as defined in [I-D.janz-nmrg-adhoc-semantic-reconciliation].

Lift
The cognitive act that produces one semantic model from one system's data. A semantic model is an ontology with its lexicon, plus pragmatics and provenance.
Reconciliation
The operation that bridges two independently produced semantic models well enough to exchange information, with no single model agreed in advance.
Reference
A thin shared nomenclature: stable identifiers, preferred names and synonyms, a shallow class, a disambiguating definition, and a canonical example, with no relationship axioms. What binds correctly is the reference's description, never a bare shared identifier.
Pragmatics
The situated meaning a data model rarely carries on its surface, which must be captured explicitly for reconciliation to be correct.
Placement of cognition
Where machine cognition sits relative to the two systems being reconciled. Each system may host its own cognitive agent, a ReAct-style reasoning loop that acts in reconciliation on that system's behalf. Such a loop is live when it can interrogate its own system, answer the counterpart's questions about it, and run virtual experiments against its knowledge graph. The placements form a spectrum by how many of the two loops are live: both sides cognitive, through one side inert, to both sides inert, the overlay case, where a single over-arching cognition works from the two systems' static artifacts.
Resolved fraction and precision
The share of the true correspondences an agent finds and commits, and the share of what it commits that is correct.
Translator
A deterministic artifact reconciliation emits, which maps instance data from one representation into the other at run time, with no model inference per message. Reconciliation produces one per direction, and each side generates and operates the inbound translator it relies on.

3. Two Deciding Systems, Not Two Models

It is easy to read reconciliation as a problem about the models alone: two schemas, and a translation that carries information from one to the other. That reading misses what makes the problem hard. The systems on either side are autonomous decision-makers. What each decides, and the facts it alone holds, is case-dependent: a service system and a network system, a BSS issuing intent and an OSS speaking service models, an alarm system and an anomaly detector. In the cross-abstraction case this document takes up, the service side decides service-level matters and the network side decides realization: which provider edge bears a site, which route distinguisher and route targets it is given, which profile meets an intent. Either way, the information each sends the other is not cargo in transit, it is an input to the other's decisions.

Three things follow, and they are not independent choices but three faces of the one fact that the deciding party on each side holds something the other needs.

Interrogation. Some facts exist only as the other system's decisions and appear in no model: the provider edge that realizes a site, the route distinguisher, the route targets. They are reached by asking the system that decided them, because they are its decisions and not derivations from any schema.

Pragmatics. The meaning of a value is relative to the deciding system's frame of use. The customer site and the provider edge sit at opposite ends of the same attachment, and each names direction from its own end, so the traffic a site calls inbound, arriving from the network, is what the PE sends outbound toward the customer: the same bandwidth value takes opposite direction labels in the two frames. The information model carries the field; it does not carry the frame. The frame lives in how each system uses the value, and it is recovered from the system, not read off the surface.

Authority. For any contested meaning, which node is a customer-facing service, how an unspecified type binds, some party must have the standing to decide it. Where an external authority, a standard and its annotations, has decided in advance, the systems inherit the decision. Where none has, the decision falls to the systems themselves, and what looked like a fact turns out to have been a decision all along, and an unstable one. That boundary is what this document's authority-withheld results size.

The placement of cognition is the payoff of this view. Only a cognitive agent standing at a system can be asked, can supply and convert its frame, and can make or settle an undecided authority call. A passive schema, or an over-arching overlay above both systems, cannot, because there is no deciding party at the system to interrogate, to furnish the frame, or to settle the call. That is why the capabilities in the cases below appear exactly as cognition is placed at the systems and fall away as it recedes to an overlay. The two cases that follow are the two poles: a determinate mapping where an overlay suffices, and a cross-abstraction mapping where, of necessity and not by choice, the deciding counterpart must be live.

4. The Operational Picture

The approach places a cognitive agent at, over, or beside each management system that must interoperate. Four elements make up the picture (Figure 1).

    YANG-modeled system                          TMF-modeled system
  +---------------------+                      +---------------------+
  |      YANG model      |                      |      TMF model      |
  +----------+----------+                      +----------+----------+
             |                                             |
             | lift                   lift                 |
             | (ReAct-style agent)    (ReAct-style agent)  |
             |          +-------------------+              |
             |<.......  | thin reference    |  .......>|   |
             |          | (optional support)|              |
             v          +---------+---------+              v
  +---------------------+         :          +---------------------+
  | lifted semantic     |         :          | lifted semantic     |
  | model               |         :          | model               |
  +----------+----------+         :          +----------+----------+
             |                     :                     |
             v                     v                     v
   +============ reconciliation (ReAct-style agent) ============+
   |    agent A    <-------- negotiate -------->    agent B     |
   +============================================================+
             | emits                              | emits
             v                                    v
  +---------------------+             +---------------------+
  | translator into A   |             | translator into B   |
  | owned by agent A    |             | owned by agent B    |
  +----------+----------+             +----------+----------+
             |     runtime: no model inference per message
             v                                    v
   A instance  <========== translate ==========>  B instance
                              |
         always-on sensing on the runtime path:
            novelty check  and  round-trip check
                              |
                              +--- drift? ---> re-wake reconciliation
Figure 1: Each side's model and its lifted semantic model, the lift a ReAct-style agent process between them; an optional thin shared reference shown by dotted links onto the two lift arrows and reconciliation; reconciliation by a ReAct-style agent emitting the two inbound translators each side operates; and an always-on sense on the runtime path triggering re-reconciliation on drift. The figure shows both sides cognitive; under the over-arching placement, agents A and B are two roles of a single reasoner over the static models, and 'negotiate' is that reasoner's internal reconcile-and-check rather than an exchange between live systems

Lift and thin reference. A ReAct agent on or over each side lifts that system's model into a semantic model. Each lift may be assisted by a thin shared reference, but the reference is a support, not a required input: where none is available, the agents can collaborate to build one, and it then serves both the lifts and the reconciliation that follows. The reference carries identity and meaning, not the full structure of the model, and binding rests on its descriptions, so two terms that merely share a label are not mistaken for one another.

Reconciliation. The two agents bind their lifted models through the shared reference, settle the correspondences, and resolve the remaining differences. Like the lifts, this step uses language-model inference; both are ReAct-style agents.

Translator generation. Reconciliation emits a deterministic translator for each direction of traffic, and each side generates and operates the inbound translator it depends on. The YANG-side agent builds the translator that renders TMF instance data into YANG terms; the TMF-side agent builds the one that renders YANG instance data into TMF terms. Neither side imposes a mapping on the other; each owns the translation into its own world. The translators carry no inference. They apply the agreed correspondence to instance data at run time. This is the separation of cognition from the data path: reconciliation is paid once, and the per-message run-time path carries no model inference.

Drift handling. Because models change, the operational plane must detect that a model has moved and cause reconciliation to run again, regenerating the affected part of the translator. Section 8 develops this.

4.1. Where Cognition Sits

Following [I-D.janz-nmrg-adhoc-semantic-reconciliation], the placement of cognition is a master variable. With both sides cognitive, each agent has its full action set, including asking the other questions and running decisive virtual experiments, and reconciliation can always complete. With one side inert, the live agent can only probe the other's static description and run solo experiments, so results stay sound but completeness is not guaranteed; a thin reference recovers part of what is lost. With both sides inert, the overlay case, cognition reconstructs both models from structure and data and checks proposals against each side's records; results stay sound but coverage is lower and full closure is not guaranteed.

Model capability is a separate, secondary variable. It affects mainly how much a reference helps a given agent. The results below were produced with an ordinary mid-capability model, which is itself informative: the determinate part of the problem does not require a frontier model.

What is realistic today bears stating plainly. The systems an operator runs now are overwhelmingly not cognitive, so the near-term placement is over-arching cognition reconciling two inert models from their static artifacts. That is the placement the first case runs in. The both-cognitive picture, live agents interrogating each other and running virtual experiments, is aspirational, and the second case shows what it adds; this document is explicit throughout about which of its claims depend on it. The immediate practical question is how much can be achieved with over-arching cognition alone. The results below answer it for the determinate part of the YANG-to-TMF problem, and they also show where the published model extensions earn their place. Precisely because an over-arching cognition has no live counterpart to ask, the authority calls it cannot derive—which nodes are services and how the unspecified type binds—must be supplied as published stipulations. The mapping-root annotations of [I-D.lls-onsen-yang-tmf-mapping] are such a stipulation, and the authority-withheld results below, where the agents agree on which nodes are services only three times in ten, show they are what over-arching cognition does not reliably recover on its own. The model extensions help most in exactly the case that is realistic today.

The converse is as important. Where over-arching cognition already suffices, which is the determinate bulk of this problem, a more cognitive placement adds nothing. For a structural interoperability problem that is not a limitation but a finding: the interoperability does not require cognitive systems at all, only an over-arching reasoner over the static models and a thin published stipulation. The sections that follow are explicit, result by result, about which is which. Where this document claims a both-cognitive advantage, it demonstrates it. Each turns on a live agent standing where the over-arching overlay has none, to be asked, to supply a frame, or to announce or adapt to a change. Proactive drift notification shows a measured gain over reactive handling, since only a live agent on the changing side announces a change before it lands. The service-to-network case adds interrogation, where asking the counterpart recovers facts the over-arching agent cannot reach; a self-adapting trigger, where a live agent rewrites its own novelty rule; and a virtual experiment, where provisioning a candidate on the live counterpart and reading it back settles a correspondence neither reasoning nor asking resolves. Its pragmatics result is conditional, a documented frame the over-arching agent reads and an undocumented one only the counterpart supplies. The virtual experiment runs against the counterpart's model, not a live physical network.

5. Reconciliation at the Schema Level

5.1. Method

The demonstrator anchors on a YANG model annotated with the mapping-root extensions of [I-D.lls-onsen-yang-tmf-mapping]. A reference mapper encodes the normative YANG-to-TMF rules of that document and serves as the oracle: the mapping the standard prescribes. Two agents, one holding the YANG side and one holding the TMF side, reconcile the model from the lifted YANG reference and the TMF conventions. Neither agent is shown the oracle. The translator the agents emit is scored against the oracle, characteristic by characteristic. To measure reliability rather than a single outcome, the reconciliation is run as several independent trials over the same model.

The demonstrator runs this reconciliation in two forms that differ only in how much is given in advance. In the authority-given form, where the authority has supplied in advance which nodes are services, the model carries the mapping-root annotations, so the customer-facing service nodes are already marked and the agents reconcile downward from them. In the authority-withheld form, the agents are given the plain [RFC8299] model with no annotations and must first locate the services themselves, then reconcile. The authority-withheld form has no oracle to score against, because no mapping is prescribed; it is measured instead by whether independent trials agree, on the services they locate and on the mapping that follows. The two forms isolate what the annotations actually contribute.

This case sits near the overlay end of the placement spectrum: one active reasoning loop works over the static artifacts of both sides, the YANG model and the TMF conventions, with no live TMF counterpart to interrogate and no virtual experiments run against a running system. The predictions of [I-D.janz-nmrg-adhoc-semantic-reconciliation] for that placement, soundness without guaranteed completeness, are therefore what the demonstrator should be expected to show.

5.2. Results with the Authority Call Supplied

In the authority-given form, where a handful of mapping-root annotations supply the one authority call the agents cannot make on their own—which nodes are customer-facing services, two findings from the L3SM service model [RFC8299] stand out.

The first concerns binding by description. Absent a shared naming convention, the two agents each made a locally reasonable but divergent choice: the TMF-side agent renamed hyphenated YANG identifiers to camel case, so the emitted translator was faithful yet did not match the oracle. Pinning identifiers in the thin reference, so that binding rests on the reference entry rather than on each agent's own styling, removed the divergence. On a small illustrative model this moved the result from four of seven characteristics renamed to zero, at no additional cost. It shows, on a small scale, the reconciliation role the thin reference plays in the central result.

The second is the central result, and it bears out the placement prediction. The full L3SM service model, the vpn-service and site roots together, has 209 characteristics across 90 nested nodes. Over 20 independent trials, 12 reproduced the agreed mapping exactly. That whole-model figure of 0.6 understates the structure of the outcome. On every characteristic that [I-D.lls-onsen-yang-tmf-mapping] specifies, the agents found and committed the correct correspondence in all 20 trials, with identifiers verbatim: resolved fraction and precision were effectively full on the specified model. Every failure, in every failing trial without exception, fell on a leaf of YANG type "empty", the one type the document's type table does not cover.

This is the overlay-placement signature. The agents were sound, no false cognates and no wrong bindings, yet not complete, because one case had no anchor to bind to. On the empty type the agents are not even internally consistent: across trials they map an empty leaf to a boolean in most runs and to a string in a minority. That is the non-determinism an unanchored case produces between independent reconciliations. A boolean, representing presence as true or false, is the better representation and the agents' more common choice. The remedy is the one the method predicts: give the case an anchor. Specifying the empty type, or carrying a convention for it in the shared reference, removes the sole source of divergence, which on this evidence restores full convergence. A both-cognitive placement lets one agent ask the other, or test whether the leaf carries a value, which confirms it is presence-only; but the choice of boolean over string on the TMF side is a representation convention, not a fact to test, so the case is closed by a published convention, not by cognition on either side.

5.3. Locating the Services Without the Authority Call

The authority-withheld form removes the annotations. The agents receive the plain [RFC8299] model and must decide, before any mapping, which nodes are the customer-facing services that should each become a TMF ServiceSpecification. This is a judgment, not a lookup, and it is where the agents diverge: across ten independent trials they produced the same set of service roots only three times. The divergence is not noise. It sorts into three kinds, each already named by [I-D.janz-nmrg-adhoc-semantic-reconciliation].

The first is granularity. Every trial identified the VPN service as a service; they differed only on the level, seven choosing the vpn-service list and three its enclosing container. A single reference binding, the L3VPN service to the list node, settles this. It is the cheap case.

The second is the false cognate. In three trials the agents also marked vpn-profiles, a library of reusable profile definitions referenced by services, as a service in its own right. This is the error the reference's description screens out: a shallow class and a disambiguating definition separate a customer-facing service from a profile library that is not one. It is the reference's precision role, at the locate step.

The third is an authority boundary, and it is the one no reasoning closes. Whether a site is a service, and if so which node carries it, split the agents four ways: two trials chose the site itself, four a service container beneath the site, one the sites collection, and three did not treat a site as a service at all. RFC 8299 gives a site a rich parameter set of its own, so whether a site is a peer service or a part of one is a modeling decision, a matter of authority over the model, not a fact recoverable from structure. [I-D.janz-nmrg-adhoc-semantic-reconciliation] predicts this directly: where resolution turns on a stipulation that neither model carries, cognition alone does not converge. The mapping-root annotation of [I-D.lls-onsen-yang-tmf-mapping] is that stipulation. Marking a node as a customer-facing service is not part of a mapping; it is the authority's declaration that the node is a service, and it is what the agents cannot derive on their own.

The instability does not stop at the root. Among the trials that agreed on the same root set, none reproduced the same full mapping of those roots' characteristics, where the authority-given form, with the roots fixed, had reproduced the prescribed mapping on every specified characteristic in every trial. An unsettled authority call is not a local defect; it propagates into the mapping beneath it.

Read together, the two forms size the authority boundary and show it is small. Give the agents one thing, which nodes are the services, a handful of annotations, and they reproduce the 209-characteristic mapping with no standard, the only residue the unspecified empty type. Withhold it and they agree on what the services even are only three times in ten, miscast a profile library as a service in three of ten, and carry that instability into the mapping. The expensive, 209-way mapping is what cognition supplies without an agreement. The small, contested thing, which nodes are services and who says so, is what is worth agreeing in advance.

6. Reconciliation in the Operational Concrete

A schema-level correspondence is necessary but not sufficient. An operator needs the translator to carry real instance data without loss, and the ongoing record and invariant checks are what hold precision at its ceiling. The demonstrator maps a real L3SM vpn-service instance to a TMF Service and back. Following [I-D.lls-onsen-yang-tmf-mapping], the annotated root's own leaves become the Service's characteristics, and each nested container or list instance becomes one Feature in a flat collection, named by its path, with list entries carrying a key predicate, linked to its parent by a composedOf relationship. The round-trip returns the original instance, so the reconciled translator is faithful on real data, not only aligned in the abstract.

The cross-abstraction case is different at the instance level, and that difference is the point. Mapping a concrete L3SM service instance to a concrete L3NM instance cannot complete from the service instance and the schema alone. It needs facts that exist only as the network's per-instance decisions, the provider edge that bears the site, the route distinguisher, the route targets, which the live counterpart must surrender before the round-trip can return the original. And the values the instance does carry must be placed in the right frame: the service's input bandwidth lands on the provider edge's outbound direction, not on the field that shares its name, so a by-name instance map is faithful in structure and wrong on the data, as the pragmatics result below measures. The same operation, map a real instance and read it back, completes from the instance alone in the determinate case and only from the instance plus the other system's decisions in the cross-abstraction case. That is the concrete signature of the both-cognitive placement. Even which network profile realizes this instance's QoS intent is not read from the instance; it is settled by provisioning a candidate profile and reading back whether it holds.

7. Pragmatics

Pragmatics is meaning relative to a deciding system's frame of use. The frame is a property of how a system uses a value in its own decisions, so it lives in the system and not in the schema, which is why it can be recovered only from a live counterpart and not read off the surface. [I-D.janz-nmrg-adhoc-semantic-reconciliation] argues that pragmatics is often load-bearing and the hardest layer to recover once cognition recedes, because it was resolved against a live context.

The YANG-to-TMF mapping of the first case is largely structural and type-level, where pragmatics is light, and that is part of why reconciliation is so reliable on the specified model. Crossing abstraction levels is where frames diverge, because each system frames the same quantity against its own vantage: the service frames bandwidth at the customer site, the network at the provider edge, the inverse. The service-to-network case below measures this, and the result is conditional in a way that is itself the point. Where the model documents its own frame, an over-arching overlay reads it and needs no counterpart. Where the frame is off the surface, only a live counterpart can supply it. Pragmatics is load-bearing exactly when the frame is not written down, which is the one case the information model cannot carry by construction, and recovering it there needs a deciding party at the system, a both-cognitive capability now measured rather than asserted.

8. Drift and Triggering

The value of the approach rests on handling model change with little or no human mapping work, because the burden of reconciliation is ongoing, not one-time. How fully the lifecycle runs without human involvement depends on placement, and the sensing-and-triggering scaffolding is itself engineered rather than free. The question is what triggers a re-reconciliation and what that implies for the architecture.

8.1. Trigger Classes

Triggers fall into two classes.

Reactive triggers are detected in the operational plane:

  • Structural novelty. The translator meets an element it has no mapping for: a new or renamed field, or an instance referring to a construct not seen before. The translator must notice unmapped elements as they pass.
  • Consistency failure. An ongoing round-trip or consistency check fails, which signals a model has moved out from under the translator. Round-trip checking is therefore not a one-time validation but a continuous operational guardrail.

Proactive triggers are notified rather than detected: a human operator or a machine pipeline signals that a change is coming, whether a model revision, new instance types, or renamings, or any change to the knowledge graph and to the schema or ontology it represents. Reconciliation then runs ahead of the change reaching live traffic.

8.2. Consequences for the Architecture

Three consequences follow, and all are consequences of placement.

First, the run-time path is not wholly free of cognition-adjacent function. Each of the two translators must carry two cheap, always-on senses, a novelty check and an ongoing round-trip check, whose job is to wake the costly cognition when a model moves. So the separation of cognition from the data path is strong in cost, no model inference per message, but it is not absolute: the path still carries the sensing that summons cognition. The closer a live agent sits to a system's operational path, the sooner it senses drift there and the more locally it can re-reconcile, which is the operational meaning of placing cognition near each system.

Second, reconciliation is a negotiation, so a trigger on one side must wake the implicated engine on the other side, or a supervisory engine above both. Proactive, bilateral notification, one side announcing a change before it lands, is available only when both sides are cognitive. A one-sided placement can only react to the other side's drift, not be told of it.

Third, an inert system has no agent of its own to install or operate a translator. The deterministic translator still has to run in that system's data path, so the path must be engineered with a point where a translator can be inserted and, as models drift, replaced: a slot the over-arching cognition fills with the translator it generates and updates. Where a system is cognitive its own agent plays that role; where it is inert, the slot and its update channel are a prerequisite, and the over-arching placement is realistic only for systems that expose one.

8.3. What the Placement of Cognition Constrains

The lifecycle is not uniform across the spectrum, and it is worth being exact about which parts hold where. The two reactive senses, the novelty check and the ongoing round-trip check, are deterministic run-time machinery, not cognition; they work whatever the placement, including over two inert systems. Reactive re-reconciliation of a drifted node is achievable with over-arching cognition alone, because the node is re-derived from the updated static model, the same operation reconciliation performs elsewhere, scoped to one node. The lifecycle was run on the vpn-service root over four real reactive changes, a new leaf, a renamed identifier, a new leaf of type empty, and a leaf promoted to a nested keyed list, each caught by the novelty or round-trip sense and re-reconciled on the affected node alone, from a baseline pair of translators built from the reconciled mapping. Proactive, bilateral notification is not achievable there: an inert system does not announce its own change, so being told of drift before it lands requires a cognitive agent on the changing side. This advantage is measured rather than asserted: proactive notification of a change eliminated the one instance that reactive detection drops, zero against one on the same change. And the sensing-and-triggering scaffolding itself, what counts as novelty, how a re-nest is scoped, how a simultaneous change on both sides is kept from thrashing, is engineered and must be adjusted as the models evolve. The claim is therefore specific: the costly mapping is re-derived without human mapping work, not that the system maintains itself with no human involvement. Genuine self-maintenance, in which the apparatus adapts its own triggers, is approached only toward the both-cognitive pole, which today's non-cognitive systems do not reach.

8.4. Incremental Reconciliation

Most ongoing change is minor: one new field, one rename. Because reconciliation proceeds model-node by model-node, re-reconciliation can be scoped to the affected node or construct rather than the whole model. Keeping the per-change cost proportional to the size of the change is a first-class design goal for operational use. Some changes are not local: a structural change that re-nests or splits a construct touches its neighbours, and a change on both sides at once must be reconciled jointly rather than twice, lest the two sides thrash. These are the cases that most reward a near, both-cognitive placement.

9. Service-to-Network Reconciliation: L3SM to L3NM

The YANG-to-TMF mapping of the previous sections is a cross-format mapping, from a YANG service model to a TMF API representation. It is structural and largely determinate, which is why over-arching cognition suffices. A second case tests the other end of the spectrum. Reconciling a service model to a network model, L3SM [RFC8299] to L3NM [RFC9182], is a cross-abstraction mapping, from service intent to network realization, and it is where [I-D.kbf-onsen-problem-statement] records inconsistent semantics and no clear service-to-network mapping, in its structural LxSM and LxNM problem area. This case stresses the capabilities the structural mapping leaves quiet, exercising four of them, each scored under the over-arching placement against the both-cognitive placement. The demonstrator uses a faithful representative extract of the L3NM module grounded in [RFC9182]; the full module drops in to reconcile against it verbatim.

9.1. Interrogation: a fact held only by the other side

The provider edge that realizes a site, and the route distinguisher and route targets, are assigned by the network and appear nowhere in the service model. The over-arching agent, reasoning over the two models, cannot produce them: on the three network-assigned attributes it scored zero of three, because the facts are not within its reach. A both-cognitive placement adds an interrogation channel, and the network agent, which holds its own realization, answers all three when asked, three of three. This is the clean both-cognitive case, and not because the over-arching agent reasons poorly. The fact is in no artifact it holds, only in the counterpart's.

9.2. Pragmatics: a frame off the surface

L3SM svc-input-bandwidth and svc-output-bandwidth are in the site frame; L3NM inbound and outbound are in the PE frame, the inverse. Whether this needs a cognitive counterpart turns on one thing, whether the model documents the frame, and running it both ways makes the condition sharp. With the frame documented in the L3NM descriptions, the over-arching agent read it off the model and resolved the inversion, mapping svc-input to outbound and svc-output to inbound, two of two correct, with no one to ask. With the frame off the surface, the same agent mapped by name, svc-input to inbound, the silent swap, zero of two, and only the both-cognitive agent, asking the counterpart its frame, recovered the inversion, two of two. So pragmatics needs the both-cognitive placement precisely when the operational context is not written down. Where a model documents its own frame, over-arching cognition reads it.

9.3. Self-maintenance: adapting a trigger

The set of access encapsulations is open. A new value slips past the leaf-level novelty sense, which flags unmapped fields rather than unexpected values, and is caught only when a round-trip fails, one instance dropped. The agent then writes itself a value-level novelty rule for the open set, and the next new value is caught by that self-added rule before any instance drops, none dropped. The apparatus adapts its own trigger, measured as the proactive drift result was, one dropped against zero.

9.4. Virtual experiment: a correspondence only testing settles

Which L3NM qos-profile realizes the service's QoS intent is not a fact the network holds. It carries several profiles, and in the demonstrator two of them share a declared label, so the declaration cannot tell them apart; the effective rate and class a profile produces are known only when it is provisioned on a test instance and read back. The over-arching agent cannot resolve it, zero of one. Asking the network agent does not resolve it either: it returns the two candidates and reports that it cannot confirm which realizes the intent without provisioning, zero of one. The both-cognitive placement adds the experiment, and the agent provisions a candidate, reads back its effective treatment, and keeps the one that holds, resolving the correspondence, one of one. This is the virtual experiment of [I-D.janz-nmrg-adhoc-semantic-reconciliation], provision a candidate on the live side and read back whether it holds, in the service-to-network setting. It runs against the counterpart's model, not a live physical network.

9.5. The Two Cases Together

The two cases are the two poles of the placement spectrum, on two of the working group's problem areas. In the cross-format mapping, over-arching cognition reproduces the whole 209-characteristic correspondence with no standard, and a more cognitive placement adds nothing to it. In the cross-abstraction mapping, over-arching cognition resolves everything the models document, but it cannot invent the network's own assignments, and it cannot recover a frame the model leaves implicit. There a cognitive counterpart moves the network-assigned attributes from zero to three of three, and the undocumented frame from zero to two of two, and resolves by experiment a qos-profile correspondence that neither reasoning nor asking settles. The determinate work needs no cognitive systems; the work that turns on a fact held only by the other side, or on context no model wrote down, is where a cognitive counterpart earns its place. This case also speaks to a question [I-D.kbf-onsen-problem-statement] poses on that problem area, whether the service and network models need to interwork and what that interworking costs.

10. Recommendations

These recommendations align with the standardization implications of [I-D.janz-nmrg-adhoc-semantic-reconciliation], applied to the YANG-to-TMF case.

For standardization:

For operational practice:

The practical frictions this document has named belong to the inert case in particular: an inert system needs a translator slot engineered into its data path, and the over-arching cognition must maintain that translator from outside, sensing drift at a distance. These are costs of placing cognition outside the systems, not of reconciliation itself. As the systems become cognitive, hosting their own agents, the costs fall away: each operates and maintains its own translator in its own path, senses its own drift where it happens, answers questions about itself, and announces its own changes. Over-arching cognition is enough for the determinate interoperability today, and nothing here asks an operator to wait. But where the cross-abstraction cases matter and self-maintenance is wanted, the direction that resolves the overhead is the one systems are already moving in: cognition at each system rather than above it.

11. Security Considerations

The deterministic translator is the trust boundary. Reconciliation uses language-model inference, so its negotiation inputs and outputs must be treated as untrusted: a lifted model, a reference entry, or a counterpart's message can carry crafted content, and prompt injection against the reasoning step is a realistic threat. The translator the step emits contains no inference and is reviewable before it carries production traffic, so the control is to validate and review the translator rather than to trust the reasoning that produced it. Runtime applies only the reviewed translator, which bounds the blast radius of a bad reconciliation to what review and the round-trip guardrail would catch.

Access control, auditing, and rate limiting defined on the YANG side must be enforced equivalently on the TMF side, consistent with the security considerations of [I-D.lls-onsen-yang-tmf-mapping]; a translator that preserved data but dropped an access constraint would be a faithful mapping and still a security regression, so constraint preservation is part of what review checks.

Two inputs are shared trust artifacts. The thin reference anchors binding, so its integrity and the authority behind its entries matter; a tampered reference can steer correspondences. The service-boundary stipulation, which nodes are customer-facing services, is security-relevant in the same way: a mis-stated boundary changes what is exposed as a service, so the authority that publishes it and the provenance of each annotation should be verifiable.

The drift mechanism is a potential denial-of-service surface. Novelty and round-trip checks wake the costly reasoning, so an adversary who can inject spurious model changes or malformed instances can try to force repeated re-reconciliation. Re-reconciliation should be rate limited and scoped to the affected node, and proactive notifications should be authenticated, so that only an authorized changing side can announce a change ahead of traffic.

Reconciliation produces an auditable record of the correspondence it establishes, including the transcripts of the negotiation, which supports after-the-fact review of how two systems agreed to exchange information.

12. IANA Considerations

This document has no IANA actions.

13. Informative References

[I-D.janz-nmrg-adhoc-semantic-reconciliation]
Janz, C. and B. Peters, "Portable Semantic Models and Ad Hoc Reconciliation by Cognitive Agents in Network Management", Work in Progress, Internet-Draft, draft-janz-nmrg-adhoc-semantic-reconciliation-01, , <https://datatracker.ietf.org/doc/html/draft-janz-nmrg-adhoc-semantic-reconciliation-01>.
[I-D.janz-nmrg-ontology-reconciliation]
Janz, C., "Ontology Reconciliation for Network Management", Work in Progress, Internet-Draft, draft-janz-nmrg-ontology-reconciliation-01, , <https://datatracker.ietf.org/doc/html/draft-janz-nmrg-ontology-reconciliation-01>.
[I-D.janz-nmrg-reference-lexicons]
Janz, C., "Reference Lexicons for Reconciliation", Work in Progress, Internet-Draft, draft-janz-nmrg-reference-lexicons-00, , <https://datatracker.ietf.org/doc/html/draft-janz-nmrg-reference-lexicons-00>.
[I-D.kbf-onsen-problem-statement]
Barguil, S., Lambrechts, K., and C. Xie, "ONSEN Problem Statement", Work in Progress, Internet-Draft, draft-kbf-onsen-problem-statement-01, , <https://datatracker.ietf.org/doc/html/draft-kbf-onsen-problem-statement-01>.
[I-D.lls-onsen-yang-tmf-mapping]
Lambrechts, K., Larsson, K., and V. Shipley, "Mapping YANG Schemas and Instance Data to TM Forum Data Models and APIs", Work in Progress, Internet-Draft, draft-lls-onsen-yang-tmf-mapping-00, , <https://datatracker.ietf.org/doc/html/draft-lls-onsen-yang-tmf-mapping-00>.
[RFC8299]
Wu, Q., Ed., Litkowski, S., Tomotaki, L., and K. Ogaki, "YANG Data Model for L3VPN Service Delivery", RFC 8299, , <https://www.rfc-editor.org/info/rfc8299>.
[RFC9182]
Barguil, S., Gonzalez de Dios, O., Ed., Boucadair, M., Ed., Munoz, L., and A. Aguado, "A YANG Network Data Model for Layer 3 VPNs", RFC 9182, , <https://www.rfc-editor.org/info/rfc9182>.
[TMF640]
TM Forum, "Service Activation and Configuration API (TMF640)", , <https://www.tmforum.org/>.
[TMF641]
TM Forum, "Service Ordering Management API (TMF641)", , <https://www.tmforum.org/>.

Acknowledgments

This document builds on the problem framing of [I-D.kbf-onsen-problem-statement] and the mapping work of [I-D.lls-onsen-yang-tmf-mapping], and applies the methodology of [I-D.janz-nmrg-adhoc-semantic-reconciliation].

Authors' Addresses

Christopher Janz
Huawei Technologies
Brad Peters
NBN