]> Alter Meridian Pty Ltd

blake@truealter.com

This memo specifies a wire-level annotation grammar by which a large-language-model output may carry, at emission and at the granularity of an individual assertion, a provenance label drawn from a closed enumerated vocabulary of substrate-class identifiers. The memo defines the closed vocabulary, the per-assertion attachment form, the admissibility discipline a relying party MAY apply to the labels, and two terminal output states, UNVERIFIED-INFERENCE and DECAYED-TO-UNCERTAINTY, equal-rank with assertion and denial. The memo does not specify what an inference system MUST do; it specifies the wire grammar by which a relying party may inspect what the inference system DID with respect to the substrates it consulted. The memo is Informational.

Introduction When a large-language-model output is consumed by another agent, by a downstream automation, or by a relying party with action authority, the consuming entity has no wire-level mechanism to distinguish three cases: (i) the model emitted the assertion from training-corpus-resident pattern without consulting any external substrate; (ii) the model emitted the assertion after consulting an external substrate whose state corroborated the assertion within an admissibility window; (iii) the model emitted the assertion after consulting an external substrate whose state did not corroborate the assertion, and the model proceeded anyway. Existing approaches to this problem operate at the prose layer: post-hoc citation insertion by a separate retrieval orchestrator, natural-language hedge phrasing ("I believe", "it appears", "according to"), per-paragraph confidence scores rendered as adjectives, or refusal to answer. All four are recognised from surface form alone, and a model trained to substitute hedge phrasing for substrate consultation defeats all four. This memo specifies a wire-level grammar by which the inference system declares, at the granularity of an individual assertion within its output, which substrate-class, if any, corroborated the assertion at emission. The grammar is closed-vocabulary, finite, and version-anchored; the consuming entity parses the annotation without interpreting prose. Two terminal annotations, UNVERIFIED-INFERENCE and DECAYED-TO-UNCERTAINTY, are equal-rank with assertion and denial. Each is an output state in its own right; neither is a confidence score or a hedge phrase.

Conventions and Definitions 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 when, and only when, they appear in all capitals, as shown here. The following terms are defined for the purposes of this document:

Assertion

A token sequence in the inference system's output that makes a truth-evaluable claim about state external to the inference system itself.

Substrate

A source of state that is external to the inference system, not writeable by the inference system at consultation time, and observable to a relying party with independent access.

Substrate-class

A named, version-anchored category of substrate. A substrate-class identifier is an opaque string drawn from the closed enumerated vocabulary of Section 3.

Provenance annotation

A wire-level label attached to an individual assertion that names the substrate-class corroborating the assertion, or names one of the two terminal annotations defined in Section 4.

Admissibility window

A predicate, evaluated at assertion emission time, that determines whether the corroborating observation from a named substrate-class is recent enough to count toward admission.

Cardinality-thresholded admissibility

A discipline by which an assertion is admitted to the inference system's output only if at least k corroborating observations from substrate-class-distinct sources are available within the admissibility window, where k is a relying-party-configurable floor.

The Closed Substrate-Class Vocabulary The vocabulary defined by this memo is closed, finite, and version-anchored. An implementation parsing a provenance annotation MUST recognise the annotation if and only if the substrate-class identifier appears in the version of the vocabulary the implementation has loaded. Unknown substrate-class identifiers MUST NOT be silently interpreted; an implementation that encounters one MUST treat the annotation as if it were UNVERIFIED-INFERENCE (Section 4). The version-anchor scheme used by this memo is a dotted major.minor pair appearing as a leading element of the substrate-class identifier. The vocabulary defined in this revision of the memo carries the version anchor 1.0. Future revisions of this memo MAY add substrate-classes; addition is a minor-version bump. Future revisions MUST NOT remove substrate-classes without a major-version bump. The version-1.0 closed substrate-class vocabulary:

substrate.git.log

The output of git log on a repository observable to the relying party, with the assertion's content appearing within a commit reachable from a named reference. Compute-location: the relying party's local working tree or a trusted mirror.

substrate.grep

The output of grep over a file-set observable to the relying party, with the assertion's content appearing in a named region of a named file. Compute-location: the relying party's local file system.

substrate.code.read

The contents of a file at a specified path, optionally with line offset and limit, as read from a file system observable to the relying party. Compute-location: the relying party's local file system.

substrate.fs.mtime

The modification timestamp of a file observable to the relying party. Compute-location: the relying party's local file system.

substrate.mcp.brief

The output of a query against a Model-Context-Protocol-discoverable service per whose state is owned by a party other than the inference system vendor. Compute-location: the MCP service's own infrastructure.

substrate.do.sse-count

The numeric subscriber count on a server-sent-event channel per , observed by a relying party with independent subscription to the same channel. Compute-location: the SSE server's infrastructure.

substrate.unix.peercred

The kernel-reported peer credentials on a Unix-domain socket connection, per . Compute-location: the kernel of the host on which the observation is made.

unverified-inference

A sentinel substrate-class identifier reserved for assertions emitted without corroboration from any other substrate-class in this vocabulary. This is the terminal annotation defined in Section 4; it is listed here for parser completeness.

The eight identifiers above constitute the entirety of the version-1.0 closed vocabulary. The compute-location line on each class above uses the term as defined in Section 2.

Terminal Annotations Two annotation values are terminal: they signal a distinct output state of the inference system, equal-rank with assertion and denial, rather than corroboration by any substrate.

unverified-inference

The inference system emitted the assertion without corroboration from any substrate-class in the vocabulary. The annotation is the inference system's declaration that the assertion is uncorroborated, and is not a confidence score.

decayed-to-uncertainty

The inference system attempted to corroborate the assertion from a substrate-class in the vocabulary, but the observation aged beyond the admissibility window before emission. The annotation is the inference system's declaration that corroboration was attempted and aged out, and is not a confidence score.

Both terminal annotations are output tokens. A relying party parsing the wire output finds them in the same slot in which substrate-class identifiers appear; the parser applies the terminal-annotation disposition without inspecting prose. The two terminal annotations are distinct from refusal to answer, from explicit denial, and from the absence of annotation.

Attachment Form A provenance annotation is attached to an individual assertion in the inference system's output. This memo specifies the abstract attachment relationship; the concrete wire encoding is the implementation's choice and is not normative herein. The annotation form is the tuple: where:

• assertion-span identifies the token sequence to which the annotation applies (the implementation's choice of identifier: character offsets, token offsets, structured-output field path, or equivalent);
• substrate-class-identifier is drawn from the closed vocabulary of Section 3 or is one of the terminal annotations of Section 4;
• observation-id (OPTIONAL) is a stable identifier for the specific observation that corroborated the assertion, suitable for re-querying the same substrate to confirm the binding;
• ts (OPTIONAL) is the time at which the corroborating observation was made.

observation-id and ts remain OPTIONAL to emit. Omitting either does not make an annotation malformed, and this memo does not constrain what an inference system emits. Omitting either does bound what a relying party can later do with the annotation under Section 6. An annotation carrying no ts cannot be counted under the procedure of Section 6 at any floor, k = 2 included, because that procedure counts only annotations whose ts falls within the admissibility window. An annotation carrying no observation-id cannot be independently re-verified, and so cannot be counted toward the k = 3 floor. An annotation carrying neither is never counted. Two concrete encodings are illustrative and not normative: JSON-structured-output encoding per : In-line bracketed annotation, for free-text outputs:

Admissibility Discipline A relying party MAY apply a cardinality-thresholded admissibility discipline to inference system output annotated under this grammar. The discipline is parameterised by:

• k, the floor cardinality of distinct substrate-classes required to corroborate an assertion before the relying party admits the assertion for downstream action.
• W, the admissibility window: a per-substrate-class duration beyond which an observation is considered aged.

Reference values:

• k = 2 is the floor for assertions whose downstream effect is bounded to the relying party's own state.
• k = 3 is the floor for assertions whose downstream effect affects state external to the relying party (settlement, policy decision, automated action with external visibility).

The relying party applies the discipline by counting, for each assertion in the inference system's output, the distinct substrate-class identifiers appearing in the assertion's provenance annotations whose ts field is within W. Assertions not meeting the cardinality floor are not admitted; assertions annotated with unverified-inference or decayed-to-uncertainty are not admitted by virtue of the terminal annotation itself. The rules of this section determine what a relying party may count toward a floor. They do not constrain what an inference system may emit. Admission affects emission in one direction only: an annotation that omits a field may be uncountable, which gives an emitter a reason to include the field. Admission never bounds what an emitter may say. An implementation MUST NOT read the cardinality language of this section as emission-time semantics. A provenance annotation MUST NOT be counted toward the k = 3 floor unless the relying party has independently re-verified it, per the Annotation Fabrication discussion in Security Considerations. Independent re-verification requires an observation-id; an annotation lacking an observation-id MUST NOT be counted toward the k = 3 floor. This requirement does not apply to the k = 2 floor, where an annotation remains admissible under the counting procedure above without independent re-verification. Independent re-verification is meaningful only where re-querying the named substrate can confirm the earlier observation rather than produce a second, unrelated one. A relying party MUST apply that test per annotation before counting it toward the k = 3 floor, and MUST NOT treat a later reading as confirming an earlier one where the substrate's value may legitimately have changed in between, or where the observed condition no longer exists to be queried. substrate.do.sse-count, whose subscriber count legitimately differs on every observation, and substrate.unix.peercred, which is unavailable once the connection it described has closed, fail that test in every case. substrate.mcp.brief names state owned by a third party and may fail it depending on whether that state is stable across the interval. The test is stated as a property because the re-queryability of a given observation is not a fixed attribute of its class. An annotation that fails the test remains admissible toward k = 2. Some annotations can never contribute to the k = 3 floor, and an implementer counting toward that floor should expect this. This memo does not specify what the relying party MUST do with an inadmissible assertion. Common dispositions include: discarding the assertion silently, surfacing it to a human reviewer, requesting re-emission from the inference system, or substituting an explicit refusal in the relying party's own output to the next consuming entity. Each disposition is the relying party's own policy choice and is not constrained by this memo.

Why the Vocabulary Is Closed An open-extensible vocabulary would permit any inference system to introduce a new substrate-class identifier and emit corroboration annotations under it. A relying party encountering an unrecognised identifier would face a choice: trust the new identifier on its face, refuse the assertion, or treat the identifier as equivalent to a fallback known identifier. Each choice fails in a way the closed vocabulary avoids:

• Trusting the unrecognised identifier on its face permits an inference system to bypass the admissibility discipline by inventing self-corroborating substrate-classes.
• Refusing the assertion stalls the consuming pipeline on a vocabulary-version mismatch, an interop-combinatorics failure mode shared with the envelope-coordination anti-pattern of .
• Fallback-equivalence collapses the semantic distinctions the vocabulary is intended to preserve.

The closed-vocabulary posture treats unrecognised identifiers as unverified-inference (Section 4). This preserves the admissibility discipline under vocabulary drift while leaving the relying party free to upgrade its parser to a newer vocabulary version.

Relation to Prior Art This memo combines a closed substrate-class vocabulary, per-assertion attachment, and two terminal annotations in a single wire-level grammar. Each adjacent family of prior work differs from this grammar in at least one of those three components: Retrieval-Augmented Generation performs retrieval against an external corpus and conditions generation on the retrieved context. RAG is an inference-system architecture; this memo specifies an output-side annotation grammar. A RAG-architected inference system MAY emit annotations under this grammar; a non-RAG inference system MAY emit annotations under this grammar. The grammar is orthogonal to the architecture. Constitutional AI and self-critique architectures apply a second model pass to evaluate the first pass's output against a specification. The output of such a system is not annotated at the granularity of an individual assertion against an external substrate-class; it is annotated, if at all, with a critique-pass verdict against a specification authored by the model vendor. Annotation against a vendor-authored specification is distinct from annotation against a substrate observable to a relying party. Multi-agent debate and related multi-pass deliberation architectures produce a single consensus output from multiple agent passes. The output's provenance is the agents' agreement process, not a substrate observable to a relying party. Hidden-state probing inspects the inference system's internal activations to estimate the system's own confidence in its output. Confidence is a property of the inference system; substrate-class corroboration is a property of the relying party's own observation surface. Cryptographically-anchored append-only logs (Certificate Transparency , trusted timestamping per ) are candidate corroborating substrates under this grammar, and each is a substrate-class a future revision of the vocabulary MAY add. A chained log considered in isolation is not a per-assertion annotation grammar.

IANA Considerations This memo requires no IANA actions in its present revision. A future revision may request establishment of an IANA registry for substrate-class identifiers, governed by the closed-vocabulary discipline of Sections 3 and 7.

Security Considerations The grammar specified by this memo surfaces three classes of attack absent from prose-only or hedge-phrasing approaches. The mitigations described below are operational rather than wire-level; this memo specifies the grammar only, and an implementation's operational posture is its own.

Annotation Fabrication An inference system may emit a provenance annotation citing a substrate-class corroboration that did not in fact occur. The grammar specified by this memo provides no cryptographic binding between the annotation and any observed substrate state. A relying party MUST NOT treat the annotation as evidence of corroboration; the annotation is a declaration of the inference system's claim about its own behaviour, which the relying party MAY independently verify by re-querying the named substrate with the optional observation-id. Admissibility Discipline requires this independent re-verification before a provenance annotation may be counted toward the k = 3 floor; the MAY above is the relying party's general option to verify any annotation, and is superseded, for that floor, by the MUST stated there.

Vocabulary Drift An inference system implemented against a newer version of the vocabulary may emit substrate-class identifiers not present in a relying party's older vocabulary. Per Section 3, the relying party treats unrecognised identifiers as unverified-inference. This fails closed. A relying party operating at a substantially older vocabulary version SHOULD upgrade its parser to the current published version of this memo.

Substrate Capture An adversary controlling a substrate identified in the vocabulary may engineer the substrate's state to corroborate assertions of the adversary's choice. The admissibility discipline of Section 6, with k = 2 or k = 3, mitigates this by requiring corroboration from substrate-class-distinct sources before admission. An adversary controlling all k substrate-classes can defeat the discipline; selection of independent substrate-classes is the relying party's operational responsibility and is not specified by this memo.

Privacy Considerations Provenance annotations expose to consumers of the inference system's output the categories of substrate the inference system consulted. In typical deployments, the substrate-class identifier is a category not a record; the per-record observation-id, if emitted, may carry the privacy properties of the underlying substrate (a filesystem path, a content hash, a peer credential identifier). A relying party emitting annotations under this grammar to a further downstream consumer SHOULD apply the same identity-binding considerations articulated in for substrate observables: pseudonymous tier observations need not carry strong identifiers; identity-bound tier observations carry the identity-binding strength of the underlying substrate. The grammar specified by this memo does not require, and does not recommend, attachment of identifiers tying the inference system's output to a particular human end-user. Such attachments, if made, are outside the scope of this memo.

Normative References In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. JavaScript Object Notation (JSON) is a lightweight, text-based, language-independent data interchange format. It was derived from the ECMAScript Programming Language Standard. JSON defines a small set of formatting rules for the portable representation of structured data. This document removes inconsistencies with other specifications of JSON, repairs specification errors, and offers experience-based interoperability guidance. Alter Meridian Pty Ltd Alter Meridian Pty Ltd Alter Meridian Pty Ltd Alter Meridian Pty Ltd Informative References This document describes an experimental protocol for publicly logging the existence of Transport Layer Security (TLS) certificates as they are issued or observed, in a manner that allows anyone to audit certificate authority (CA) activity and notice the issuance of suspect certificates as well as to audit the certificate logs themselves. The intent is that eventually clients would refuse to honor certificates that do not appear in a log, effectively forcing CAs to add all issued certificates to the logs. Logs are network services that implement the protocol operations for submissions and queries that are defined in this document. This document describes the format of a request sent to a Time Stamping Authority (TSA) and of the response that is returned. It also establishes several security-relevant requirements for TSA operation, with regards to processing requests to generate responses. [STANDARDS-TRACK] Alter Meridian Pty Ltd

Acknowledgements This memo specifies a grammar layered above the substrate-observation primitive of , the identity substrate of and , and the discovery mechanism of . Mikhail Sergeev read -01 closely and asked where the memo's normative boundary between emission and admission sits. The boundary existed in -01 but the text did not state it. The two paragraphs added to Sections 5 and 6 in -02 state it, and the placement is his suggestion.

IPR Posture The applicant of any patent rights that may be construed to read on the grammar specified by this memo will file an IPR disclosure under the IETF's standard procedures. The applicant's intended licensing posture for any such rights is royalty-free with defensive-termination, consistent with the applicant's published IPR disclosures on companion memos in the morrison-* family.

Document History draft-morrison-substrate-provenance-grammar-03:

• Section 3: added an informative reference to Section 2 of for the term used in the compute-location line on each class.
• Section 5: removed the instruction to carry the field notes forward and the closing sentence about an emitter capping its own annotations; stated that an annotation with neither field is never counted.
• Section 6: removed the characterisation of the k = 3 consequence and of the reason the test is stated as a property.
• Section 8: reworded the opening paragraph to describe what the grammar combines, without a statement of contribution.
• Section 10.1: removed the sentence on cryptographic binding.
• Section 10.2: removed the evaluative sentence on the fail-closed posture.
• Acknowledgements: reworded the account of the question about emission and admission.
• Conventions and Definitions: BCP 14 key-words paragraph now generated by the boilerplate directive; text unchanged.
• Added the informative reference.
• Rewrote sentences across the document in plainer terms, set the definitions and the vocabulary as definition lists, and removed emphasis capitals outside BCP 14 key words. No requirement, identifier or encoding changes.

draft-morrison-substrate-provenance-grammar-02:

• Sections 5 and 6: stated the boundary between emission and admission, and the conditions under which an annotation counts toward the k = 2 and k = 3 floors.

Contributors