| Internet-Draft | Identity Streams | November 2026 |
| Morrison | Expires 22 May 2027 | [Page] |
When one system tells another about a trait it has inferred about a person, the receiver today learns the trait and a confidence, and nothing about what the inference was drawn from. Because the receiver cannot tell a trait the person stated in answer to a question from one read off the person's activity on a linked platform, it cannot apply the lawful basis, consent and context limits that belong to each. This document defines a source tag carried with every inferred trait on the wire. The tag names, for each stream that contributed to the trait, a registered stream class, the provenance class of the observation, where the inference was computed, the lawful basis and purposes the stream was consented for, and the contexts the trait may be used in. It reuses the provenance-class vocabulary of the Morning Brief specification, establishes a registry of stream classes, and states the rules a receiver follows, including refusing a trait whose sources it cannot check. The same tag carries facts the person states about themselves, such as when they are available or the terms on which they offer something, and those facts keep the limits of the stream they were given on. It defines a wire field only. How traits are inferred, stored or scored is out of scope.¶
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 22 May 2027.¶
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. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
An identity-inference system derives statements about a person (a disposition, a competency, a style of working) from what the person does, without asking the person to declare them. Such statements increasingly pass between systems. An identity service answers an organisation's query with them and a credential issuer hands them to a verifier, while an agent may pass to another what it has learnt about the person it acts with.¶
The obligations attached to an inferred statement depend on where it came from. Data-protection law ties lawful basis and purpose to the collection of the data an output is derived from [GDPR]. Consent given for inference from one stream does not cover inference from another. Some inferences are prohibited in some contexts whatever consent exists. The EU AI Act prohibits inferring the emotions of a natural person in the workplace and in education institutions, except for medical or safety reasons [EU-AI-ACT]. A trait derived on the person's own device from passively observed activity differs from one the person supplied in a structured exchange, even when both carry the same name and value.¶
A receiver handed a bare trait knows none of this. It can honour the obligations only by trusting that the sender already did, and it cannot carry them forward if it passes the trait on. The Compute-Location Gate specification [COMPUTE-GATE] already requires an inference service to keep the provenance class on every derived statement inside its own architecture, and says a statement that has lost its provenance class is not consent-checkable. This document extends that requirement across the boundary between systems.¶
A source tag, sources, carried with each inferred trait
(Section 3).¶
The fields of each source entry, and the CDDL [RFC8610] that defines them, with CBOR [RFC8949] and JSON [RFC8259] serialisations.¶
A registry of stream classes with general initial entries (Section 4), and registries for lawful bases, purposes and use contexts.¶
The rules an issuer follows when it emits a tagged trait (Section 5) and the rules a receiver follows when it accepts one (Section 6).¶
A stream class for facts the subject states, and the rules that keep a stated fact distinct from an inference (Section 5.2).¶
This document does not specify how traits are inferred, the trait vocabulary, which facts a subject may state, how sources are weighted or combined, the container a trait travels in, or any settlement of value owed to the subject. The tag is designed to sit inside an existing signed container, such as a COSE_Sign1 structure [RFC9052], a signal in a Morning Brief [MORNING-BRIEF], or a verifiable credential, and it inherits that container's integrity protection.¶
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.¶
The natural person a trait is about.¶
A statement about a subject, such as a value on a named dimension
together with a confidence. A trait is either derived by inference
or, where every source is of class subject-statement, stated by
the subject and carried as stated (Section 5.2).¶
A distinct channel of observation about a subject, under one grant of consent or other lawful basis. Two linked accounts on the same platform are two streams. A stream has exactly one stream class.¶
The kind of channel a stream is, drawn from the registry in
Section 10.1. A stream class names a kind of channel and
identifies no party: peer-attestation says that a peer attested
without saying which.¶
How the observation was made, as defined by the Morning Brief specification [MORNING-BRIEF] and drawn from its "Morning Brief Provenance Classes" registry.¶
The system that emits a tagged trait.¶
The system that accepts a tagged trait, whether to use it or to pass it on.¶
The setting in which a receiver intends to use a trait, such as a hiring decision or an educational assessment, drawn from the registry in Section 10.4.¶
Every trait an issuer emits under this specification carries a
sources member: a non-empty array of source entries, one for each
stream that contributed to the trait. A trait derived from three
streams carries three entries. An issuer MUST NOT merge entries for
different streams into one entry, and MUST NOT omit a contributing
stream because its contribution was small.¶
The sources member MUST be inside the integrity protection of the
container that carries the trait. A tag that can be stripped or
altered without breaking the container's signature provides none of
the properties in this document.¶
tagged-trait = {
trait : tstr, ; trait name, issuer vocabulary
value : any, ; trait value, issuer vocabulary
? confidence : float16 / float32,
sources : [+ source],
? issued : tdate,
}
source = {
stream-class : tstr, ; registered, see Section 4
stream : tstr, ; opaque stream reference
provenance : provenance-value,
computed : "device" / "service",
basis : tstr, ; registered lawful basis
purposes : [+ tstr], ; registered purposes
contexts : [+ tstr], ; registered use contexts
? window : [tdate, tdate],; first and last observation
? share : float16 / float32, ; 0.0 to 1.0
? consent-ref : tstr, ; URI to check consent status
? return-ref : tstr, ; URI naming where value owed
; to the subject is settled
}
provenance-value = "active" / "passive-aggregate"
¶
The JSON serialisation uses the same member names. Times are RFC 3339 date-time strings in both serialisations.¶
stream-classThe stream's class. It MUST be a value in the "Identity Stream Classes" registry (Section 10.1).¶
streamAn opaque reference to the stream, chosen by the issuer. It MUST be stable for a given subject, stream and receiver because a receiver has to match a later revocation to the traits it affected. It MUST NOT be the same value for the same stream across different receivers (Section 9.1), and MUST NOT encode an account name, platform user identifier or any other identifier of the subject. A random value such as a version 4 UUID [RFC9562] kept per subject, stream and receiver meets these requirements.¶
provenanceThe provenance class of the observations the stream contributed.
The value is drawn from the Morning Brief provenance-class registry
[MORNING-BRIEF]. That registry deliberately omits any token for
passive observation of a single individual. Rule 7 of Section 5
says when an entry may carry active, and Section 5.1 says how a
derivation made on the subject's device is tagged.¶
computedWhere the inference from this stream was computed: device for a
device under the subject's control, service for infrastructure
operated by the issuer or on its behalf.¶
basisThe lawful basis under which the stream was collected and the inference made, from the "Identity Stream Lawful Bases" registry (Section 10.2).¶
purposesThe purposes the stream's consent or other basis covers, from the "Identity Stream Purposes" registry (Section 10.3). These are the purposes as recorded when the basis was established. An issuer MUST NOT list a purpose the subject did not agree to at that time, and MUST NOT widen the list later by reinterpreting an earlier grant.¶
contextsThe use contexts the trait may be used in, as far as this stream is concerned, from the "Identity Stream Use Contexts" registry (Section 10.4).¶
windowThe times of the first and last observation drawn from the stream for this trait.¶
shareThe issuer's statement of how much of the trait this stream contributed, between 0.0 and 1.0. The shares of one trait need not sum to 1.0, because issuers combine sources in different ways. The value is informative.¶
consent-refA URI at which a receiver holding the right credentials can learn whether the stream's consent is still in force. It MUST NOT identify the subject to a party that does not already know the subject.¶
return-refA URI naming where value the receiver owes the subject for this stream is settled, for deployments where reading a trait carries a payment to the subject. How settlement works is out of scope.¶
A trait drawn from two streams, shown in the JSON serialisation. All values are examples.¶
{
"trait": "collaborative_problem_solving",
"value": "high",
"confidence": 0.62,
"sources": [
{
"stream-class": "direct-response",
"stream": "6f1c2e0a-8b7d-4d2e-9a51-3c0f7e2b9d14",
"provenance": "active",
"computed": "service",
"basis": "consent",
"purposes": ["trait-inference", "disclosure-on-query"],
"contexts": ["general", "workforce"],
"window": ["2026-08-02T09:15:00Z", "2026-09-30T18:40:00Z"],
"share": 0.7
},
{
"stream-class": "linked-account",
"stream": "c2a94d71-0f3e-4b8a-bf62-91d5e7a0c3f8",
"provenance": "active",
"computed": "service",
"basis": "consent",
"purposes": ["trait-inference"],
"contexts": ["general"],
"consent-ref": "https://identity.example/consent/c2a94d71",
"share": 0.3
}
]
}
¶
Because the second stream's consent does not cover disclosure on query and its contexts do not include workforce, a receiver in a hiring setting cannot use this trait at all (Section 6, rules 5 and 6).¶
A stream class describes the channel, at a level of generality that holds across implementations. The initial classes are:¶
direct-responseThe subject's own answers in an exchange whose purpose, as the subject understood it, was to contribute to inference about them.¶
relayed-responseThe subject's own answers, submitted through a software agent
acting at the subject's direction rather than typed by the
subject. The choice is the subject's, but signals that depend on
the subject's own pace and phrasing are absent, and receivers
SHOULD weigh it below direct-response.¶
authored-workWork the subject has attached and attested as their own, such as writing, code or a portfolio item. Inference is drawn from the work itself.¶
peer-attestationA statement about the subject made by another identified person. The class does not identify the peer.¶
linked-accountData from an account the subject holds on a third-party platform and has linked to the issuer. The platform's authorisation of access is not, by itself, a lawful basis for inference (Section 5, rule 4).¶
third-party-metadataFacts about the subject's presence on a third-party service, such as counts, timestamps or structure, without content.¶
public-recordThe subject's work or activity in a publicly available corpus. Being public does not make it consented for inference.¶
service-interactionHow the subject uses the issuer's own service, observed by the issuer in the course of providing it.¶
device-activityActivity observed on a device under the subject's control, inferred from on that device (Section 5.1).¶
subject-statementA fact the subject stated about themselves to be passed on as they
stated it, such as when they are available, where they will work,
or the terms on which they offer something. It differs from direct-response, whose answers feed
an inference (Section 5.2).¶
New classes are added by the procedure in Section 10.1. A class MUST describe a kind of channel and MUST NOT be defined by one platform, product or implementation.¶
An issuer MUST attach a source tag to every trait it emits under this specification, and MUST NOT emit a trait whose sources it cannot enumerate. A trait whose provenance has been lost inside the issuer is not emitted.¶
An issuer MUST record the stream, stream class, provenance class and compute location of each observation when the observation is taken, and MUST NOT reconstruct them afterwards from the trait.¶
An issuer MUST NOT emit a source entry whose provenance is any
value outside the Morning Brief provenance-class registry, and
MUST NOT emit a trait about a named subject with a source entry
whose provenance is passive-aggregate. An aggregate describes
a cohort, and a statement about one person drawn only from it is
not an aggregate.¶
For linked-account and third-party-metadata streams, the
issuer MUST hold a basis for inference that is separate from, and
separately revocable from, the platform's authorisation of
access. The purposes list reflects that separate basis.¶
An issuer MUST NOT list a use context in contexts that the
stream's basis and purposes do not support, and MUST NOT list
workforce or education for any trait that is an inference of
emotion, in jurisdictions where such inference is prohibited in
those settings. A categorical prohibition applies whatever the
tag says.¶
When consent for a stream is withdrawn, the issuer MUST stop
emitting traits that list that stream, and SHOULD make the
withdrawal visible at consent-ref where one was given.¶
A source entry of any stream class carries provenance set to
active only where the subject initiated the act that released
that stream's evidence for inference, for example by linking the
account. Section 3.1 of [COMPUTE-GATE] defines the active class
this way. Where the subject initiated no such act, an issuer MUST NOT emit a source entry for that stream in a trait about a named
subject. Section 5.2 applies the same test to a stated fact, where
the subject's act is the stating.¶
The Compute-Location Gate specification [COMPUTE-GATE] requires an inference from passive observation of a single individual to be computed on the subject's device. The derived statement is retained on that device and is never transmitted to a server as a passive observation. Section 4 of [COMPUTE-GATE] gives the one route by which such a derivation becomes visible to a server. The subject contributes it by a fresh active-class act, and the contribution enters as active-class signal and is gated afresh.¶
A trait contributed this way is tagged with computed set to device,
provenance set to active (the contribution is the subject's act), and
the stream class of the underlying observation, normally
device-activity. The underlying observations never appear on the
wire. An issuer MUST NOT emit a device source entry for a
derivation the subject did not contribute by such an act, and MUST NOT emit a
service source entry for any inference drawn from passive
observation of a single individual.¶
A trait whose every source entry has the stream class
subject-statement is a stated fact. Its value is what the subject
said, and the issuer relays it without vouching for it.¶
An issuer MUST NOT emit, under subject-statement, a value the
subject did not state, and MUST NOT carry under it a value the
issuer inferred. Changing the representation without changing the
meaning, such as converting a time zone or a currency format, is
permitted.¶
A source entry of this class carries provenance set to active,
since stating is the subject's act, and computed names where the
statement was recorded. An issuer SHOULD omit confidence from a
stated fact.¶
The second time in window is when the subject last stated or
confirmed the fact. Stated facts such as availability go stale,
and an issuer SHOULD include window on them. When the subject
withdraws or replaces a stated fact, the issuer MUST stop emitting
the old value, as for withdrawn consent (Section 5, rule 6).¶
A trait drawn from a subject-statement source together with
sources of any other class is an inference, and its value is not
the subject's statement.¶
A stated fact is unverified, and a claim that a third party has
checked, such as a licence or an insurance policy, is not carried as
a subject-statement source.¶
A receiver processes each trait as follows, and stops at the first rule that fails.¶
If the trait carries no sources member, or an empty one, the
receiver MUST NOT use the trait and MUST NOT pass it on.¶
If the sources member is outside the integrity protection of
the container, the receiver MUST treat the trait as untagged.¶
If any source entry carries a provenance value outside the
Morning Brief provenance-class registry, the receiver MUST reject
the whole trait as malformed. If the trait is about a named
subject and any entry carries passive-aggregate, the receiver
MUST reject it.¶
If any source entry carries a stream-class, basis, purpose or
context the receiver does not recognise, the receiver MUST treat
that entry as permitting no purpose and no context.¶
The receiver MUST NOT use the trait for a purpose unless every source entry lists that purpose. A trait inherits the narrowest purposes of its sources.¶
The receiver MUST NOT use the trait in a use context unless every
source entry lists that context. A receiver whose intended use
context is workforce or education MUST also refuse any
trait it knows to be an inference of emotion, whatever the tag
lists.¶
Before an action with material effect on the subject, such as a
hiring or admissions decision, the receiver SHOULD check
consent-ref for every entry that carries one, and MUST NOT act
on a trait whose consent the check reports withdrawn.¶
A receiver that passes a trait on MUST pass the sources member
unchanged, and MUST NOT add contexts or purposes to it. A receiver
that derives a new trait from received traits MUST carry forward
every source entry of every input trait.¶
A receiver MUST NOT present a trait as stated by the subject unless
every source entry has the stream class subject-statement, and
MUST NOT present a stated fact as verified.¶
Rules 5 and 6 take the intersection of the sources rather than the union because a trait is an inseparable product of its inputs. A receiver cannot use the part of a trait that came from a stream the subject agreed to share and leave behind the part that came from one they did not.¶
The provenance-class vocabulary is the Morning Brief's
[MORNING-BRIEF]. This document adds no provenance class and
inherits that registry's rule that no token permitting single-subject
passive inference may be registered. A Morning Brief signal that
carries an inferred trait SHOULD carry the trait's sources member
in its body, and its own provenance_class SHOULD be the most
restrictive of the entries.¶
The Compute-Location Gate [COMPUTE-GATE] decides, inside one
inference service, where an inference may compute and whether it may
proceed, keyed to the pair of provenance class and signal stream.
This document carries the outcome of that decision with the trait
once it leaves the service. The stream and stream-class fields
correspond to that specification's signal_stream, with the stream
reference made opaque and receiver-specific for the reasons in
Section 9.1.¶
The tag is the issuer's statement. A receiver can check that the
tag is well formed and signed, but cannot check that it is true. An issuer that
tags a passively observed stream as direct-response, or lists
contexts its consent never covered, has produced a conformant tag
that misstates the facts, and nothing on the wire detects it. An
inference tagged subject-statement is the same case, and it presents
the issuer's conclusion as the subject's own words.
Receivers SHOULD accept tagged traits only from issuers whose
practices they have a basis to rely on, and SHOULD prefer issuers
whose inference pipelines are subject to audit.¶
A party between issuer and receiver could remove source entries, so
that a trait appears to come only from its most permissive stream,
or add contexts. Placing sources inside the container's signature
(Section 3) defeats this for any receiver that verifies the signature.
An attacker gains nothing by moving the tag outside the signature,
as rule 2 of Section 6 treats an unprotected tag as absent.¶
A receiver could derive a new trait from a received one and emit it without the original sources, and the restrictions would then fall away at the second hop. Rule 8 of Section 6 forbids this. It is a rule on conduct; a receiver that breaks it has broken this specification, and the second receiver cannot tell.¶
Consent can be withdrawn after a trait is emitted. A tagged trait
held by a receiver remains signed and well formed after its basis has
gone. consent-ref and rule 7 of Section 6 let a receiver check
before acting, and receivers SHOULD NOT retain tagged traits longer
than their purpose needs.¶
The considerations of [RFC6973] apply throughout. The tag itself discloses information about the subject, and the rules below limit it.¶
A list of the streams that contributed to a trait says which kinds of channel the subject has. It can show that they linked a platform account, that peers attested about them, or that they use the issuer's service in a particular way. Stream classes are deliberately coarse so that the tag names a kind of channel and identifies no platform or person. Issuers MUST NOT encode a platform name, peer identity or account identifier in any field.¶
An opaque stream reference that was the same across receivers
would let receivers who compare notes link the same subject across
their separate dealings. The requirement in Section 3 that the
reference differ per receiver prevents this.¶
An issuer SHOULD emit only the traits a query needs, and the tag for
each lists only the streams that contributed to that trait. The
window, share, consent-ref and return-ref fields are
optional, and an issuer SHOULD omit any a receiver does not need.
window in particular can reveal when the subject was active, and
SHOULD be coarsened to the day or omitted.¶
A stated fact can say where the subject will be and when. An issuer
SHOULD carry such a fact at the coarsest precision the receiver's
purpose needs, such as an area for a place or a half-day for a
time, and MUST NOT carry an exact address as a
subject-statement source. An exact place needed to deliver
something belongs in the agreement between the parties.¶
The purposes and contexts lists carry purpose limitation across
the boundary between systems. Their value depends on issuers
recording them honestly when the basis was established and never
widening them afterwards (the purposes field, Section 3.3).¶
Some inferences are prohibited in some settings regardless of consent, the clearest being emotion inference in the workplace and in education [EU-AI-ACT]. The tag lets a receiver see which contexts the sources permit; it cannot make a prohibited inference permitted, and both issuer and receiver rules require refusal where a categorical prohibition applies.¶
The rule in Section 5.1 keeps raw passive observations of an individual off the wire. Only the derived trait the subject chose to contribute crosses it, tagged so the receiver knows it was computed on the subject's device from that subject's own activity.¶
This document requests IANA to create four registries under a new "Identity Streams" registry group. Each entry carries a value, a short description and a reference. The registration policy for each is Specification Required [RFC8126]. The designated expert SHOULD refuse a value that duplicates an existing one, and MUST refuse one whose meaning depends on a single platform, product or implementer.¶
Initial entries, each with this document (Section 4) as the
reference: direct-response, relayed-response, authored-work,
peer-attestation, linked-account, third-party-metadata,
public-record, service-interaction, device-activity,
subject-statement.¶
The expert MUST refuse a class that names a party rather than a kind of channel.¶
Initial entries: consent (the subject agreed to the stated
purposes), explicit-consent (consent meeting the stricter standard
some regimes require for special categories of data), contract
(necessary to perform an agreement with the subject),
legal-obligation, legitimate-interests. Each has this document
as the reference.¶
Initial entries:¶
trait-inference: deriving traits from the stream.¶
subject-reflection: showing the subject what was inferred about
them.¶
trait-history: keeping the history of a trait over time.¶
disclosure-on-query: disclosing the trait to a third party that
queries for it.¶
verification: confirming a statement about the subject to a
party that asks.¶
Each has this document as the reference.¶
Initial entries:¶
general: no setting with specific legal restrictions.¶
workforce: recruitment, employment and work allocation.¶
education: admissions, assessment and teaching.¶
credit: lending and creditworthiness.¶
health: health and care.¶
civic: access to public services or civic standing.¶
Each has this document as the reference. The expert SHOULD admit a context where its setting carries legal restrictions distinct from those already registered.¶
draft-morrison-identity-streams-00 (November 2026): first version.
Defines the per-trait source tag, the stream-class, lawful-basis,
purpose and use-context registries, and issuer and receiver rules,
and carries facts the subject states under a subject-statement
stream class. A source entry about a named subject is active
only where the subject initiated the release of its evidence, and a
derivation made on the subject's device reaches a service only when
the subject contributes it.¶