Intent ↔ L1 Reconciliation — cognitive agent-to-agent live demo

demo byChristopher Janz
speed Reference lexicon ON
Reconciling a declarative intent against a structural L1 service model — two very different models, one network. Agent O is a TM Forum Intent-interface OSS (TR 290 / IG 1253 / TMF921); Agent N speaks IETF L1CSM + TE service-mapping over the OTN/optical network. They refine quantitative expectations down to a realization and assure operational state up to fulfilment — verified by satisfaction, not round-trip identity. Toggle the reference lexicon ON/OFF to compare the work.

About this demo — what you're looking at

This is the second fully-cognitive, agent-to-agent case, and a deliberately wider gap than the first. Agent O speaks intent: declarative expectations with quantitative bounds — bandwidth ≥, latency ≤, availability ≥, protection required. Agent N speaks a concrete L1 service: discrete client signals (ODUk), routes, wavelengths, protection schemes. The two differ in nature (a predicate vs a value), in scope (consumer realm vs provider realm), and in abstraction (what to deliver vs how it is delivered). Neither side merges the other's schema; both are competent, interrogable agents that work out the mapping between themselves.

The mapping is a refinement, not an equivalence. An expectation such as bandwidth ≥ 8 Gbit/s does not equal ODU2; it is satisfied by it. So the right test is not that intent → network → intent returns the original bytes — refinement is lossy and under-determined, the provider picks among many valid realizations — but that the chosen realization satisfies every expectation. Verification here is satisfaction, with feasibility / best-achievable / judge when an expectation cannot be fully met.

Two directions, both shown. Refine-down: an intent arrives; Agent N derives and provisions a realization that satisfies it. Assure-up: operational state flows back; Agent O maps it to fulfilment — met, at-risk margin, or breach. Watch the cognition meter (heaviest at planning, the lift, the parametric refinement and the feasibility judge), and the pragmatics — realm (who decides), state (intended vs operational), authority (the judge) — which carry most of this case.

Stage Press ▶ Play to begin. The agents will plan, lift to ontologies, bind to a reference lexicon, align, refine expectations down to a realization, judge feasibility, co-refer endpoints, construct the mappings, and run live changes both ways.
Cognition load
—
Gap closed by
—
Verified by · measure
—

🟢 Agent O — Intent OSS cognitive proxyTR 290 / IG 1253 / TMF921

🟡 A2A message busTMF Agentic / JSON-RPC

🔵 Agent N — L1CSM network cognitive proxydraft-ietf-ccamp-l1csm-yang + teas mapping

Reference lexicon
Alignment
Knowledge graphs & instances
Refinement · down
Feasibility & judge
Assurance · up
Mappings
Traces
Measures & cognition
Work ledger · with vs without
Semantic model
A thin, generic reference — neither TM Forum nor IETF. It pins identity only, and it deliberately separates expectation-kind entries (a bound: ≥, ≤) from realization-kind entries (a discrete value or resource). The two are connected by a refinement relation, not by equality.
Two passes. Pass 1 — label: cheap candidates on names alone. Pass 2 — gloss + example: the disambiguating definition and canonical example confirm, reject, or mark a gap. The trap here is a nature false cognate — an expectation and a metric that share a word.
No alignment yet — press Play.
The lift makes each implicit ontology explicit (heavy cognition). Endpoint co-reference — a separate step — settles which delivery points and access points are the same, from evidence.

Agent O — lifted intent ontology

— not yet lifted —

Agent N — lifted L1CSM ontology

— not yet lifted —

Endpoint co-reference — which delivery points and access points co-refer, from evidence (service-order key → UNI id → location)

Endpoints co-refer in stage 8.
The heart of the case: each quantitative expectation is refined down to a concrete realization by a derived function, and checked by satisfaction — the realization meets the bound — not by round-trip identity. Refinement is one-to-many; the provider picks, in its realm.
KindExpectation (Agent O)Realization (Agent N)Refinement & satisfaction check✓
No refinements yet.
When expectations conflict, the network computes feasibility and returns a best-achievable offer; the intent judge (IG 1253) evaluates it against acceptance criteria and priorities, and accepts a degraded value or rejects. This is the beat where the agents work hardest to nail the answer down — and reach it without a human.
Feasibility is negotiated in stage 7.
The other direction. Operational state flows up from the network; Agent O maps each measured value to the fulfilment of its expectation — MET, AT-RISK MARGIN, or BREACH. Assurance is verified by the same satisfaction predicate, evaluated on live data.
Assurance mapping appears from stage 10.
The negotiated artefact: two directional mappings. Refine-down (intent → L1CSM) is one-to-many — it carries the refinement policy and the provider's choice. Assure-up (operational → fulfilment) is an aggregation to a satisfaction verdict. They are not inverses. Cognition is spent here once; runtime executes with no inference per message.

Refine-down — intent ▶ L1CSM

— generated after mediation —

Assure-up — operational ▶ fulfilment

— generated after mediation —
Deterministic application of the mappings for each live change — inbound native document, rule firings, emitted native document. No model inference per message.
No translations yet.
The per-stage ledger: how each gap is closed, how it is verified, the measure, and the cognition load. Note that verification is satisfaction and feasibility, not round-trip identity.
StageCognitionClosed byVerified byMeasure
Fills in as stages complete.
Same reconciliation, both ways. The reference lexicon does not just save conversation — it saves verification, and it pre-empts the nature false cognate by binding expectations and metrics to different kinds of entry, so "latency ≤ 5ms" is never mistaken for a measured latency.
The form-independent semantic model — meaning (semantics and pragmatics) lifted clear of any form, with provenance. Here the pragmatics carry the case: realm (consumer expectation vs provider realization), state (intended vs operational — the assurance loop), and authority (the feasibility judge). An agent consumes this directly; TMF921 and L1CSM are merely projections of it.
Assembled at the final stage — press Play, or jump with Step.

Intent view — service demands & expectations

met at-risk margin breach declared demand

Network view — L1 realization (OTN / optical)

access / ROADM optical link supporting-tp
Staged demonstration — the conversation is scripted for legibility; a code-backed run is scaffolded separately. Schema fragments are faithful but abbreviated. Intent side follows TM Forum TR 290 (Intent Common Model / TIO), IG 1253 (Intent Management, feasibility / best-achievable / judge) and TMF921; network side follows draft-ietf-ccamp-l1csm-yang and draft-ietf-teas-te-service-mapping-yang over the OTN/optical network of the first demo. A2A envelopes follow the TM Forum Agentic (A2A) JSON-RPC message/send shape.