Summary

  • draft-feng-netmod-naim-01 makes Canonical JSON authoritative, while its Markdown View and token-efficient LLM Context View are derived for different audiences and carry different information-loss rules.
  • Schema validation, deterministic projection, a successful model retry and a clean YANG parse are cumulative engineering evidence. None alone proves complete requirements, correct deployment semantics, standards adoption, server implementation, authorization, applied configuration or network outcome.

Suppose an operator says, “Allow the backup path only after the primary path has failed, and never for regulated traffic.” A modeling assistant asks about keys and writeability, emits a NAIM document, survives JSON Schema validation, regenerates its Markdown cleanly and produces YANG accepted by a validator. The dashboard is green. Yet the original sentence may still conceal three unresolved questions: who decides that the primary path has failed, how regulated traffic is identified, and whether “allow” means configure, authorize or execute. No serializer can recover facts that never entered the record.

That is the useful tension inside revision 01 of the NAIM draft. Natural AI Interface Modeling is proposed as a semantic intermediate representation between natural-language requirements and YANG. The draft says direct conversion is prone to guesses about configuration versus state, list keys, constraints, operational preconditions and cross-module relationships. It responds with a deliberately layered artifact rather than a single prompt.

Status discipline comes first. The Datatracker record lists Chong Feng's document as an active individual Internet-Draft, updated 18 July 2026, with no stream, no recorded intended RFC status and the state I-D Exists. The draft header says “Intended status: Standards Track.” That header records the author's destination; it is not NETMOD working-group adoption, IETF endorsement or an RFC. The history records revisions, not implementation.

Three views, three scopes of authority

The draft gives Canonical JSON the controlling role. Interoperable tools must treat it as authoritative. A Markdown View is generated deterministically for review and version control. If Markdown is accepted as an editing surface, represented fields must survive a round trip and fields present only in JSON must not disappear silently. Conflicts require an explicit resolution. This is more than formatting hygiene: it makes authority visible at the point where a convenient view could otherwise become a second, inconsistent master.

The LLM Context View has a different bargain. It is derived for token-efficient model consumption and must include names, semantic types, constraints, writeability, visibility, supported operations, preconditions and side effects. It must omit transport-construction details, internal metadata, raw JSON Schema, deviation declarations and extension definitions. It is intentionally not round-trippable and must not reconstruct Canonical JSON.

Those omissions are reasonable if the attempted task does not need them. They are dangerous when “shorter prompt” becomes “complete operational context.” A model can answer faithfully from the supplied view and still be wrong for a server whose deviations, feature set or external augments alter the effective schema. The relevant evidence is therefore not that context was optimized, but that each omitted field was classified against the task and target environment.

NAIM reinforces the separation through three layers. The Specification layer defines the document, schema and view-generation rules. The Skill layer contains AI-driven conversational modeling, summarization, YANG generation and reverse engineering. The Tool layer is deterministic conventional software: it validates structure, produces views, generates YANG from confirmed material and extracts structure from existing modules. Stable specification and replaceable skills are a sound design choice. They let operators test the common layer without granting one model permanent authority.

Retry narrows an error; it does not widen the facts

The Validation-Reflection-Retry loop is similarly bounded. A tool emits a structured error with a path and expected condition. The system can inject it into the model context and retry. After a maximum count it must stop, expose the remaining defect as a targeted question and should preserve the correction history.

A successful retry proves that the new output passed the checks that rejected the old one. It does not show that the model repaired the user's meaning. The loop may converge on schema-valid JSON by deleting a difficult optional field, choosing a plausible key or turning an unresolved operational condition into prose. The audit record must therefore preserve not only the final green result but also the input requirement, questions asked, assumptions made, deleted or defaulted fields, tool and schema versions, and every error-driven change.

The same boundary applies after YANG generation. RFC 7950 supplies rich language semantics: must, when, leafref, augment, features, configuration state and instance constraints. A validator can provide strong evidence when it receives the exact module closure, feature choices, deviations and test instances. It cannot establish that a natural-language when became the intended XPath, that a cross-module augment exists on the target server, or that a valid edit is authorized. RFC 8407 adds mature design guidance, not an adoption certificate for generated output.

Nor does server advertisement close the chain. RFC 8525 lets a server describe its YANG Library, module sets, features and deviations. That is indispensable evidence of a server's declaration. It must be joined to exact source bytes, client resolution traces and compiled-schema fingerprints. RFC 8342 then distinguishes intended from operational state. A successful edit can exist in intended configuration while the network is still converging, rejected by hardware, overridden or producing an unintended effect.

NAIM itself says it is not a JSON encoding of YANG and not a YANG replacement. It also says runtime operation validation belongs to the separate NAIM-OP draft. Revision 01 leaves the AI model, prompts, confidence scoring and runtime execution unspecified. Its structural checks are valuable precisely because their claims remain narrow.

An evidence ladder that preserves the gaps

For a production claim, retain at least six linked records:

  1. the original request, clarification dialogue and unresolved assumptions;
  2. exact Canonical JSON bytes, source provenance, schema version and hash;
  3. Markdown and LLM Context derivatives, converter identity and an omission/conflict ledger;
  4. exact YANG, full dependency closure, features, deviations, validator identity and positive and negative tests;
  5. server library snapshot, client resolution, authorization decision and transaction identifiers;
  6. intended and operational datastore observations, telemetry, rollback and measured network outcome.

RFC 8259 is a reminder that valid JSON has a deliberately limited meaning. The same intellectual discipline should govern every later green check.

This follows Lu Heng's argument for a minimum deterministic common layer: make shared validity locally testable, but leave later adoption to participants that run and verify it. His Running-Code Primacy and reality-layers notes sharpen the boundary between a published representation and an executable result. They are BTW's analytical lens here; Lu Heng is not being presented as a reviewer of NAIM.

Sources