Verdict Records: Results on Claims that Cite Receipts as Evidence
draft-wright-verdict-record-02
This document is an Internet-Draft (I-D).
Anyone may submit an I-D to the IETF.
This I-D is not endorsed by the IETF and has no formal standing in the
IETF standards process.
| Document | Type | Active Internet-Draft (individual) | |
|---|---|---|---|
| Author | Darren Wright | ||
| Last updated | 2026-11-02 | ||
| RFC stream | (None) | ||
| Intended RFC status | (None) | ||
| Formats | |||
| Stream | Stream state | (No stream defined) | |
| Consensus boilerplate | Unknown | ||
| RFC Editor Note | (None) | ||
| IESG | IESG state | I-D Exists | |
| Telechat date | (None) | ||
| Responsible AD | (None) | ||
| Send notices to | (None) |
draft-wright-verdict-record-02
Network Working Group D. Wright
Internet-Draft Sencor AI S.L.
Intended status: Experimental 2 November 2026
Expires: 6 May 2027
Verdict Records: Results on Claims that Cite Receipts as Evidence
draft-wright-verdict-record-02
Abstract
This document defines a verdict record: a statement, bound to its
issuer, of a result reached on a claim, together with the evidence
the result relies on and the rule by which the result can be re-
derived. Evidence is cited by digest and can be a receipt or record
in any format, such as a SCITT Receipt or an Agent Action Capsule.
The document defines how a checker checks a verdict record and what
such a check does and does not show. It defines no receipt format
and changes none.
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 6 May 2027.
Copyright Notice
Copyright (c) 2026 IETF Trust and the persons identified as the
document authors. All rights reserved.
Wright Expires 6 May 2027 [Page 1]
Internet-Draft Verdict Records November 2026
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.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1. Requirements Language . . . . . . . . . . . . . . . . . . 3
1.2. Terminology . . . . . . . . . . . . . . . . . . . . . . . 3
2. The Verdict Record . . . . . . . . . . . . . . . . . . . . . 4
2.1. Encoding and Digests . . . . . . . . . . . . . . . . . . 4
2.2. Members . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3. The Claim . . . . . . . . . . . . . . . . . . . . . . . . 6
3. Carrying a Verdict Record . . . . . . . . . . . . . . . . . . 6
4. Checking a Verdict Record . . . . . . . . . . . . . . . . . . 7
4.1. Results of a Step . . . . . . . . . . . . . . . . . . . . 7
4.2. Inputs . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.3. Steps . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.4. State and Report . . . . . . . . . . . . . . . . . . . . 8
4.5. Checker Policy . . . . . . . . . . . . . . . . . . . . . 9
5. What a Check Shows and Does Not Show . . . . . . . . . . . . 9
6. Reassessment and Withdrawn Evidence . . . . . . . . . . . . . 10
7. Relation to Receipt Formats . . . . . . . . . . . . . . . . . 11
8. Related Work . . . . . . . . . . . . . . . . . . . . . . . . 11
9. Security Considerations . . . . . . . . . . . . . . . . . . . 13
10. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 14
11. References . . . . . . . . . . . . . . . . . . . . . . . . . 14
11.1. Normative References . . . . . . . . . . . . . . . . . . 14
11.2. Informative References . . . . . . . . . . . . . . . . . 14
Appendix A. Examples . . . . . . . . . . . . . . . . . . . . . . 17
A.1. A Model Decision . . . . . . . . . . . . . . . . . . . . 17
A.2. A Rule Re-run on Cited Delegation Records . . . . . . . . 20
Appendix B. Open Questions . . . . . . . . . . . . . . . . . . . 22
Appendix C. Changes Since -01 . . . . . . . . . . . . . . . . . 23
Appendix D. Changes Since -00 . . . . . . . . . . . . . . . . . 23
Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 24
Wright Expires 6 May 2027 [Page 2]
Internet-Draft Verdict Records November 2026
1. Introduction
Several formats record that something happened or that something was
said. A SCITT Signed Statement, registered on a Transparency
Service, yields a Receipt [RFC9943]. An Agent Action Capsule records
one agent action and its disposition
[I-D.mih-scitt-agent-action-capsule]. A Signed Evaluation Receipt
records one evaluation result
[I-D.gruszka-signed-evaluation-receipts]. Each records an act or a
statement; a Receipt is issued by the Transparency Service on which
the statement is registered.
This document defines a companion to those formats: a verdict record,
which states a result reached on a claim and cites such receipts and
records as its evidence. It does not define how evidence is
captured, and it does not replace any receipt format.
A verdict record names the rule that produced its result. Where the
rule is mechanical, any checker that holds the evidence can re-run it
and compare. Where no mechanical rule can re-derive the result, the
record says so, and a check then shows only what the result was based
on, and when.
A check of a verdict record never shows that a claim is true. It
shows that the issuer asserted the result, on the stated evidence,
under the stated rule.
An architecture for auditing agent interactions
[I-D.kuehlewind-audit-architecture] describes an Auditor that
determines whether recorded behaviour matched the user's intent and
the authorization in force, and whether recorded evidence supports or
contradicts a claim. A verdict record is one way to record such a
determination, so that a party that trusts neither the agent nor its
operator can check what it was based on. The records that
architecture defines can be its evidence.
1.1. Requirements Language
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.
1.2. Terminology
Claim: A statement whose support is in question. A claim can be
carried by a record or receipt in any format, or stated as text.
Wright Expires 6 May 2027 [Page 3]
Internet-Draft Verdict Records November 2026
Evidence item: A sequence of bytes, such as a receipt, a signed
statement or a record, that a verdict record cites by its digest.
Issuer: The party whose key binds a verdict record. The record
names it in its writer member.
Checker: Whoever runs the checks of Section 4. A checker needs no
access to the issuer's systems.
Presenter: Whoever hands the checker a verdict record and its
evidence items. That can be the issuer, a party that relies on
the record, or anyone.
Rule: A named, specified procedure that re-derives a result from the
inputs a record names (Section 2.2). The name judgement means
that no mechanical rule can re-derive the result. A rule name
other than judgement SHOULD identify the version of the
specification it names, so that a changed rule has a new name.
2. The Verdict Record
2.1. Encoding and Digests
A verdict record is a JSON object [RFC8259] in canonical form. In
canonical form, object members are sorted by name, compared as
sequences of Unicode code points; there is no whitespace outside
strings; strings escape only the quotation mark, the reverse solidus
and control characters, and write every other character as its UTF-8
bytes; numbers are integers; and arrays keep their order. Re-
serializing the parsed record MUST give the same bytes.
A verdict record MUST be strict JSON: valid UTF-8 that does not start
with a byte-order mark, in which no object at any depth has the same
member name twice and no escape forms a lone surrogate. A checker
MUST refuse JSON nested deeper than 32 levels at any point; the
record itself is at level 1, and each object or array inside it adds
one level.
An integer is written as an optional minus sign and digits only, with
no leading zero, no fraction and no exponent, and lies between -(2^53
- 1) and 2^53 - 1. A time is UTC, written YYYY-MM-
DDTHH:MM:SS.ffffffZ with exactly six fraction digits. Hex is
lowercase, two characters per byte, with no prefix. A member of the
wrong JSON type is refused at the step that reads it. Bytes that
break the rules of this section, or a record that breaks Section 2.2
other than in evidence and check, are refused before any step runs:
the state is refused and the report names the rule broken.
Wright Expires 6 May 2027 [Page 4]
Internet-Draft Verdict Records November 2026
The digest of a verdict record, and of an evidence item, is the
SHA-256 hash [FIPS180-4] of its exact bytes, written in hex. This
document defines no salted or hiding form of digest.
2.2. Members
A verdict record has exactly the members below. No other member is
allowed. The members v, kind, key, version, writer and body MUST be
present. A record without evidence or check can be parsed, but fails
step 1 or step 2 of Section 4.3.
v: A string. This document defines the value "verdict-record/1".
kind: A string: the record kind. This document defines two kinds:
"check-run", a run of a mechanical rule on the inputs the record
names, and "model-decision", a decision made by an AI model. A
kind starting with "x-" is private.
key: A non-empty string: the record's key within its kind, chosen by
the issuer. It MUST be opaque.
version: An integer of 1 or more: the version of this record for
this key.
supersedes: Null for version 1; for a later version, the digest of
the previous version.
writer: A non-empty string: the issuer, the party whose key binds
the record.
actor: An object: who reached the result. Its members are type
("person" or "system") and id, and for a system also model and
model_version.
act_time: A time: when the result was reached, by the issuer's
statement.
evidence: An array of the evidence items this record relies on.
Each entry is an object with exactly two members: sha256, the
digest of the item (Section 2.1), and format, a string that names
the format by whose rules the item is verified (Section 7).
check: An object with exactly the members rule (a string, the name
of the rule), inputs (optional) and result (the result the rule
gave).
Wright Expires 6 May 2027 [Page 5]
Internet-Draft Verdict Records November 2026
The member inputs is an object. Each of its members names one
input of the rule, and its value says where the checker reads that
input:
* a string: the name of a member of body. The input is that
member's value.
* an object with the member evidence, an integer index into the
evidence array, optionally the member pointer, and no other
member. With no pointer, the input is the digest the checker
computes from the item it holds for that entry. With pointer,
a JSON Pointer [RFC6901], the input is the value at that
pointer in the item's bytes, parsed as strict JSON by the rules
of Section 2.1; the item need not be in canonical form.
An issuer SHOULD bind an input to evidence rather than copy a
value from an evidence item into body: a copy is the issuer's
statement, while a bound input is read from the cited bytes.
body: An object: the content of the record. Its members depend on
kind. It holds the claim (Section 2.3) and any values the rule
reads from body.
2.3. The Claim
The member body.claim names the claim the verdict is about. It is an
object with exactly one of two members: text, a string that states
the claim, or ref, an evidence entry (as in evidence) for the record
or receipt that carries the claim. When ref is used, the same entry
MUST also appear in evidence.
3. Carrying a Verdict Record
This document does not define how a verdict record is signed, time-
stamped or made available. From the container that carries the
record, a checker needs three things: the exact bytes of the record;
a key that bound those bytes, shown to belong to the party the record
names in writer; and an upper time bound, a time no later than which
the record existed, which the checker establishes itself.
A SCITT Signed Statement whose payload is the verdict record,
registered on a Transparency Service whose Receipt the checker
verifies, is one such container [RFC9943].
If the container does not show the record, the state is held, with
the reason (Section 4.4).
Wright Expires 6 May 2027 [Page 6]
Internet-Draft Verdict Records November 2026
4. Checking a Verdict Record
4.1. Results of a Step
PASS: the step ran and its condition holds.
FAIL: the step ran and its condition does not hold. The result
names the step.
NOT RUN: an input the step needs is missing, or is not shown to come
from outside the issuer's reach (Section 4.2). The result names
the step and the input it needs.
When a step cannot run, the check goes on with its other steps. If a
later step FAILs, the check FAILs at that step. Otherwise it is NOT
RUN at the first step that could not run.
Two checkers that implement this document and hold the same inputs
MUST give the same result and name the same step. No step reads the
checker's clock.
4.2. Inputs
Every input that a step relies on MUST come from outside the issuer's
reach. In this document an input is outside the issuer's reach in
two ways only: the checker computes it from bytes it holds (a digest,
a signature, or the re-run of a rule), or it is the checker's own
input (its policy (Section 4.5), its trust anchors, this document,
and the rule specifications it holds). A checker MUST NOT take its
own inputs from the record or from the issuer.
A statement made by the issuer alone, such as the content of the
record, can be reported. It MUST NOT make a step PASS that would
otherwise not pass.
A step MUST NOT rely on the absence of a statement that the issuer or
the presenter could withhold.
One exception: under the default policy (Section 4.5), step 3 can
rely on a value in body that check.inputs names. That value is the
issuer's statement, and the report names each such input, so a reader
can see which inputs came from the issuer.
4.3. Steps
Step 1. The member check MUST be present. It MUST name a rule and
give a result. The rule judgement is allowed only for kinds that the
checker's policy lists; otherwise step 1 FAILs.
Wright Expires 6 May 2027 [Page 7]
Internet-Draft Verdict Records November 2026
Step 2. The member evidence MUST be an array of entries as in
Section 2.2. For each entry, the checker MUST hold an item whose
digest equals the entry's sha256; if it does not, step 2 is NOT RUN,
naming the entry.
Each item is verified by the rules of the format its entry names. An
item that does not verify, or whose format the checker's policy does
not list, is not relied on: step 2 is then NOT RUN, naming it.
An item that is itself a verdict record is checked by this section,
as a record of its own, and verifies only when its own state is
released. Where an item's verification gives it an upper time bound,
that bound MUST NOT be later than the upper time bound of the record
that cites it; if it is later, step 2 is NOT RUN. The checker's
policy sets how deep such citations are followed: the record being
checked is at depth 0, so following one citation needs a depth of 1
or more. Beyond that depth, step 2 is NOT RUN.
Step 3. If the rule is not judgement, the checker MUST hold the
rule's specification. It resolves each input that check.inputs names
(Section 2.2) and re-runs the rule on them. If an input is bound to
an evidence item the checker does not hold, or that step 2 does not
rely on, step 3 is NOT RUN. If an input cannot be resolved (no such
member of body, no such evidence entry, a binding object with another
member, a pointer that does not resolve, or an item that is not JSON
by the rules of Section 2.1), or is of the wrong JSON type, that is a
FAIL of this step, not a result of the rule. If one input cannot be
read for the first reason and another for the second, the step FAILs.
If the checker's policy reads inputs from evidence only
(Section 4.5), an input read from body makes step 3 NOT RUN. The
result MUST equal check.result; if it does not, step 3 FAILs.
The check passes when steps 1 to 3 PASS.
4.4. State and Report
A checker reports the record's state: released, when steps 1 to 3
PASS; held, when no step FAILs and at least one is NOT RUN; refused,
when a step FAILs; or no verdict, when the checker's own policy is
invalid. With the state it reports each step's result and, for each
step that FAILed or did not run, the step and, where it did not run,
the input it needed. Under the default policy the report also names
each rule input read from body (Section 4.2).
Wright Expires 6 May 2027 [Page 8]
Internet-Draft Verdict Records November 2026
A checker MUST NOT report any word or member that speaks for the
whole record and that a reader could take as "this is true", such as
verified, valid, true, trusted, correct, proven, confirmed or OK. A
summary line, where one is printed, names the state and every step's
result.
A state speaks as of the record's upper time bound. The report
states that time, and the SHA-256 digest of the checker's policy
(over its canonical form, Section 2.1) and of each set of trust
anchors it used, so that another checker can see what the state rests
on.
4.5. Checker Policy
A checker's policy is its own input. It lists the kinds for which
judgement is allowed, the evidence formats the checker can verify,
and the depth of citation it follows, as an integer of 0 or more. It
can also state that rule inputs are read from evidence only; by
default they are read from body or from evidence, as each input says.
A policy that lacks one of the first three, or states any of them in
a form other than these, is invalid, and a checker whose policy is
invalid gives no verdict.
Reading inputs from evidence only means that step 3 reads bytes the
record cites by digest, not values written into body. Those bytes
are outside the issuer's reach only as far as their verification in
step 2 shows that another party made them. A record may still carry
values in body; under this policy an input bound to body makes step 3
NOT RUN (Section 4.3), and other values in body play no part in the
check.
5. What a Check Shows and Does Not Show
A released verdict record shows that its issuer asserted the result,
on the stated evidence, under the stated rule, no later than its
upper time bound. It never shows that the claim is true.
For the rule judgement, a check shows only what the result was based
on, and when. It does not show that the judgement is correct.
The values in body that a rule reads are the issuer's statement. A
check does not show that they are true. An input bound to an
evidence item shows only that the rule read the cited bytes; it does
not show that what those bytes state is true.
Wright Expires 6 May 2027 [Page 9]
Internet-Draft Verdict Records November 2026
A check does not show that the issuer kept its key secret, or that no
other version of the record exists beyond those the checker holds.
Nor does it show that no evidence item has been withdrawn
(Section 6).
A check does not show that the evidence a record cites is all the
evidence there was, or that the issuer wrote a record for every
result it reached. A verdict record provides no completeness
property. Where completeness matters, it comes from outside the
record: from the container, or from the formats of the evidence
items, for example a sequence that makes an omitted record visible.
6. Reassessment and Withdrawn Evidence
A result can change: evidence arrives late, a cited item is withdrawn
or superseded, or the rule or policy changes. A verdict record does
not change. A reassessment is a new record. An issuer that
reassesses its own result writes a new version of the same key, whose
supersedes is the digest of the earlier version; both versions remain
checkable. A party that reassesses another issuer's record writes a
record of its own and cites the earlier record as evidence. If it
cites the earlier record with the format "verdict-record/1", the new
record is released only while the earlier one is; to reassess a
record that is no longer released, it cites the earlier record's
bytes under another format.
Where a record cites another verdict record, the cited record is
checked each time the citing record is checked, and the citing record
relies on it only if the cited record is released; otherwise step 2
is NOT RUN (Section 4.3). A change that alters the state of a cited
record therefore reaches every record that cites it, to the depth the
checker follows, the next time that record is checked. Because the
check is deterministic (Section 4.1), such a change comes only from a
change in the items the checker holds or in its own inputs, such as
its trust anchors or policy. No notice is sent to the citing
records. The aim is comparable to truth maintenance [DOYLE79], but
nothing keeps justification links: the dependency is recomputed when
a record is checked. A later version of a cited record does not by
itself change the state of the earlier version; how a checker that
holds a later version should treat a citation of the earlier one is
an open question (Appendix B).
Because no step relies on the absence of a statement (Section 4.2), a
checker cannot conclude that an evidence item has not been withdrawn.
A withdrawal, retraction or revocation changes a check only where the
verification rules of the item's format take it as input, or where
the checker's trust anchors or policy change; no step of Section 4.3
reads it otherwise.
Wright Expires 6 May 2027 [Page 10]
Internet-Draft Verdict Records November 2026
7. Relation to Receipt Formats
Receipts and records made by other formats, such as SCITT Signed
Statements and Receipts [RFC9943], Agent Action Capsules
[I-D.mih-scitt-agent-action-capsule] and Signed Evaluation Receipts
[I-D.gruszka-signed-evaluation-receipts], enter a verdict record as
evidence items. A verdict record adds no property to them and
requires none of them.
The format member of an evidence entry is a media type [RFC6838]
where one is registered for the item, and otherwise a URI that names
the specification of the item's format. For an item that is itself a
verdict record, the format is the string "verdict-record/1". A
checker verifies an item only by the specification so named.
The digest in an evidence entry is the digest of the exact bytes that
the item's format verifies. For a format whose items are signed
envelopes, such as COSE_Sign1 or DSSE, those bytes include the
envelope. Some formats identify an item by a different digest, for
example a digest of the envelope's payload; that digest does not
replace the one a verdict record uses. An issuer that wants to
relate the two can cite the payload as an item of its own.
8. Related Work
Several documents define records of what an AI agent did, was
permitted to do, or was decided about. A verdict record differs from
each in the same way: it records a result reached on a claim, cites
such records as evidence, and names a rule that a checker re-runs
where the rule is mechanical. This section describes the published
revisions cited.
* The audit architecture [I-D.kuehlewind-audit-architecture] defines
Interaction, Action, Delegation and Authorization Transition
Records, and an Auditor that reaches a determination over them. A
verdict record can carry that determination, with those records as
its evidence.
* The AI Agent Execution Profile of SCITT
[I-D.emirdag-scitt-ai-agent-execution] records material agent
actions in a hash-chained, independently held evidence chain, and
places content evaluation out of its scope. Its records can be
evidence items.
Wright Expires 6 May 2027 [Page 11]
Internet-Draft Verdict Records November 2026
* The pre-execution authorization profile
[I-D.munoz-scitt-permit-profile] binds an allow, deny or challenge
decision to the canonical bytes of the request later dispatched.
It records a decision about an action before it runs; a verdict
record records a result on a claim, and can cite such a decision
as evidence.
* The AI-agent action receipt profile
[I-D.noa-scitt-ai-agent-receipt] records an action with its
recorded verdict at a governed boundary, and states that it does
not claim the agent was correct.
* The proofbundle decision receipt [PROOFBUNDLE-DR] records a signed
verdict on a proposed action, with a policy digest, evidence
digests, what was not checked, and the conditions that would
change the decision. It identifies the policy by digest; it does
not specify a re-run of the policy by a checker.
* Permission receipts [I-D.izmaylov-agent-permission-receipts] and
MUSUBI settlement [I-D.oga-musubi-settlement] each specify a
verifier that recomputes from records whether recorded actions
stayed within a signed permission or grant. Each fixes one rule.
A verdict record names its rule per record and compares the
issuer's result with the re-run.
* The Gravit Epistemic Verification Protocol [I-D.gravit-gevp] moves
claims through states backed by records from verifiers, with a
committed state reached by convergence. A check of a verdict
record never shows that a claim is true; each record is one
issuer's statement.
* What an agent audit record can and cannot establish
[I-D.zhang-audit-record-limits] states limits that hold for any
record format, including that a verifier never reports an
unchecked property as passed. Section 5 and Section 4.4 follow
the same discipline.
* Decision receipts for approvals, declined actions and executed
actions [I-D.gruszka-audit-decision-receipts] give each of those a
signed record with exact bytes and map them onto the record types
of the audit architecture. They can be evidence items. That
document cites this one; its digests cover the envelope's payload,
so a verdict record that relates the two cites the payload as an
item of its own (Section 7).
* Execution-trace verification for agents
[I-D.durga-agenttrust-trace-verification] chains delegation hop
records by the digest of their canonical form and requires that
Wright Expires 6 May 2027 [Page 12]
Internet-Draft Verdict Records November 2026
scopes do not widen along the chain. Its hop records can be
evidence items, and two of its Section 5 checks can be applied by
the rule of a verdict record (Appendix A.2).
* Post-quantum agentic delegation
[I-D.durga-pquip-pqc-agentic-delegation] defines hybrid signatures
over delegation chains and aggregated execution receipts. Its
receipts can be evidence items.
* The Verifiable Agent Trust Envelope [VATE] records verifier-side
admission decisions and links them to receipts. Where a required
receipt arrives after its deadline, it records a new assessment
and keeps the earlier one. Section 6 takes the same approach.
* PALA-1 [I-D.sparysh-pala-audit] separates internal consistency,
completeness and existence in time, and reports completeness as
not checked where it cannot be checked. A verdict record makes no
completeness claim (Section 5).
9. Security Considerations
A verdict record is its issuer's statement. Its result can be wrong,
and a mechanical rule can be re-run correctly on values that are
false. A released state is a statement about what was asserted and
on what evidence, not about the claim.
A presenter can withhold evidence items or later versions of a
record. Because no step relies on the absence of a statement that
could be withheld (Section 4.2), withholding can leave a step NOT RUN
but cannot make a step PASS.
The rule judgement admits any result. A checker's policy limits it
to the kinds for which the checker accepts a result that no rule re-
derives.
The check is deterministic and reads no clock (Section 4.1), so two
checkers holding the same inputs report the same state.
Every evidence item is given to the checker in full. An issuer ought
to consider, before citing an item, whether it is fit to pass to
every party that may check the record.
Key compromise, revocation and the trustworthiness of time sources
are properties of the container (Section 3) and are outside the scope
of this document.
Wright Expires 6 May 2027 [Page 13]
Internet-Draft Verdict Records November 2026
10. IANA Considerations
This document has no IANA actions. A later version may ask IANA for
registries of record kinds, rule names and evidence format names,
which this version lists in Section 2.2, Section 4.3 and Section 7.
11. References
11.1. Normative References
[FIPS180-4]
National Institute of Standards and Technology, "Secure
Hash Standard (SHS)", FIPS PUB 180-4,
DOI 10.6028/NIST.FIPS.180-4, August 2015,
<https://doi.org/10.6028/NIST.FIPS.180-4>.
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/info/rfc2119>.
[RFC6901] Bryan, P., Ed., Zyp, K., and M. Nottingham, Ed.,
"JavaScript Object Notation (JSON) Pointer", RFC 6901,
DOI 10.17487/RFC6901, April 2013,
<https://www.rfc-editor.org/info/rfc6901>.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, <https://www.rfc-editor.org/info/rfc8174>.
[RFC8259] Bray, T., Ed., "The JavaScript Object Notation (JSON) Data
Interchange Format", STD 90, RFC 8259,
DOI 10.17487/RFC8259, December 2017,
<https://www.rfc-editor.org/info/rfc8259>.
11.2. Informative References
[DOYLE79] Doyle, J., "A Truth Maintenance System", Artificial
Intelligence, vol. 12, no. 3, pp. 231-272, 1979.
[I-D.durga-agenttrust-trace-verification]
Durga, R. K., "Cryptographic Execution-Trace Verification,
JCS Canonicalization, and Fail-Closed Delegation
Attestation for Autonomous AI Agents", Work in Progress,
Internet-Draft, draft-durga-agenttrust-trace-verification-
00, 29 October 2026,
<https://datatracker.ietf.org/doc/html/draft-durga-
agenttrust-trace-verification-00>.
Wright Expires 6 May 2027 [Page 14]
Internet-Draft Verdict Records November 2026
[I-D.durga-pquip-pqc-agentic-delegation]
Durga, R. K., "Post-Quantum Hybrid Cryptographic
Delegation Chains and Merkle-Aggregated Execution Receipts
for Autonomous AI Agents", Work in Progress, Internet-
Draft, draft-durga-pquip-pqc-agentic-delegation-00, 31
October 2026, <https://datatracker.ietf.org/doc/html/
draft-durga-pquip-pqc-agentic-delegation-00>.
[I-D.emirdag-scitt-ai-agent-execution]
Emirdag, P. and D. Feichtinger, "AI Agent Execution
Profile of SCITT", Work in Progress, Internet-Draft,
draft-emirdag-scitt-ai-agent-execution-01, 12 October
2026, <https://datatracker.ietf.org/doc/html/draft-
emirdag-scitt-ai-agent-execution-01>.
[I-D.gravit-gevp]
Konviser, A., "Gravit Epistemic Verification Protocol
(GEVP)", Work in Progress, Internet-Draft, draft-gravit-
gevp-11, 23 October 2026,
<https://datatracker.ietf.org/doc/html/draft-gravit-gevp-
11>.
[I-D.gruszka-audit-decision-receipts]
Gruszka, K., "Decision Receipts for Approvals, Declined
Actions and Executed Actions in Agent Audit Records", Work
in Progress, Internet-Draft, draft-gruszka-audit-decision-
receipts-03, 2 November 2026,
<https://datatracker.ietf.org/doc/html/draft-gruszka-
audit-decision-receipts-03>.
[I-D.gruszka-signed-evaluation-receipts]
Gruszka, K., "Signed Evaluation Receipts", Work in
Progress, Internet-Draft, draft-gruszka-signed-evaluation-
receipts-04, 26 October 2026,
<https://datatracker.ietf.org/doc/html/draft-gruszka-
signed-evaluation-receipts-04>.
[I-D.izmaylov-agent-permission-receipts]
Izmaylov, P., "Signed Permissions and Action Receipts for
Automated Agents", Work in Progress, Internet-Draft,
draft-izmaylov-agent-permission-receipts-01, 24 October
2026, <https://datatracker.ietf.org/doc/html/draft-
izmaylov-agent-permission-receipts-01>.
[I-D.kuehlewind-audit-architecture]
Kühlewind, M. and H. Birkholz, "An Architecture for
Auditing Agent Delegation and Interactions", Work in
Progress, Internet-Draft, draft-kuehlewind-audit-
Wright Expires 6 May 2027 [Page 15]
Internet-Draft Verdict Records November 2026
architecture-01, 7 September 2026,
<https://datatracker.ietf.org/doc/html/draft-kuehlewind-
audit-architecture-01>.
[I-D.mih-scitt-agent-action-capsule]
Mih, S., "An Agent Action Capsule Profile for SCITT", Work
in Progress, Internet-Draft, draft-mih-scitt-agent-action-
capsule-05, 26 September 2026,
<https://datatracker.ietf.org/doc/html/draft-mih-scitt-
agent-action-capsule-05>.
[I-D.munoz-scitt-permit-profile]
Munoz, C., "A SCITT Profile for Pre-Execution AI Action
Authorization Records", Work in Progress, Internet-Draft,
draft-munoz-scitt-permit-profile-01, 19 July 2026,
<https://datatracker.ietf.org/doc/html/draft-munoz-scitt-
permit-profile-01>.
[I-D.noa-scitt-ai-agent-receipt]
Toraman, T., "A SCITT Profile for AI-Agent Action
Receipts", Work in Progress, Internet-Draft, draft-noa-
scitt-ai-agent-receipt-01, 14 August 2026,
<https://datatracker.ietf.org/doc/html/draft-noa-scitt-ai-
agent-receipt-01>.
[I-D.oga-musubi-settlement]
Oga, T., "MUSUBI Settlement: Recomputing Whether an Agent
Acted Within a Signed Grant", Work in Progress, Internet-
Draft, draft-oga-musubi-settlement-02, 22 October 2026,
<https://datatracker.ietf.org/doc/html/draft-oga-musubi-
settlement-02>.
[I-D.sparysh-pala-audit]
Sparysh, A. and O. Verteletskyi, "PALA-1: A Tamper-Evident
Audit Record Format for Constrained and Disconnected
Deployments", Work in Progress, Internet-Draft, draft-
sparysh-pala-audit-02, 22 October 2026,
<https://datatracker.ietf.org/doc/html/draft-sparysh-pala-
audit-02>.
[I-D.zhang-audit-record-limits]
Zhang, K., "What an Agent Audit Record Can and Cannot
Establish", Work in Progress, Internet-Draft, draft-zhang-
audit-record-limits-00, 27 October 2026,
<https://datatracker.ietf.org/doc/html/draft-zhang-audit-
record-limits-00>.
Wright Expires 6 May 2027 [Page 16]
Internet-Draft Verdict Records November 2026
[PROOFBUNDLE-DR]
b7n0de, "Decision Receipt predicate decision-receipt/
v0.1", July 2026,
<https://github.com/b7n0de/proofbundle/blob/v6.4.0/docs/
predicates/decision-receipt.md>.
[RFC6838] Freed, N., Klensin, J., and T. Hansen, "Media Type
Specifications and Registration Procedures", BCP 13,
RFC 6838, DOI 10.17487/RFC6838, January 2013,
<https://www.rfc-editor.org/info/rfc6838>.
[RFC8785] Rundgren, A., Jordan, B., and S. Erdtman, "JSON
Canonicalization Scheme (JCS)", RFC 8785,
DOI 10.17487/RFC8785, June 2020,
<https://www.rfc-editor.org/info/rfc8785>.
[RFC8792] Watsen, K., Auerswald, E., Farrel, A., and Q. Wu,
"Handling Long Lines in Content of Internet-Drafts and
RFCs", RFC 8792, DOI 10.17487/RFC8792, June 2020,
<https://www.rfc-editor.org/info/rfc8792>.
[RFC9943] Birkholz, H., Delignat-Lavaud, A., Fournet, C., Deshpande,
Y., and S. Lasker, "An Architecture for Trustworthy and
Transparent Digital Supply Chains", RFC 9943,
DOI 10.17487/RFC9943, June 2026,
<https://www.rfc-editor.org/info/rfc9943>.
[VATE] Sato, T., "Verifiable Agent Trust Envelope (VATE):
recipient inbox reference example", November 2026,
<https://github.com/Poke-nushi/Verifiable-Agent-Trust-
Envelope/blob/b64ab7efeaf00ee285c1fdf15759dbcf792374f8/
reference/recipient-inbox/README.md>.
Appendix A. Examples
A.1. A Model Decision
The record below is a decision by a model on a claim carried by
another record. Its rule is judgement, so a checker re-runs no rule;
it holds both evidence items, verifies each by its own format, and
reports the state. Long lines are folded for display as described in
[RFC8792]; the record itself has no whitespace and no line breaks.
The claim record cited by both body.claim.ref and the first evidence
entry is these bytes:
Wright Expires 6 May 2027 [Page 17]
Internet-Draft Verdict Records November 2026
=============== NOTE: '\' line wrapping per RFC 8792 ================
{"made_by":"builder.example","statement":"Release 2.1 of example-ser\
vice passed its acceptance suite."}
The verdict record, with whitespace added for display:
Wright Expires 6 May 2027 [Page 18]
Internet-Draft Verdict Records November 2026
=============== NOTE: '\' line wrapping per RFC 8792 ================
{
"act_time": "2026-10-23T10:00:00.000000Z",
"actor": {
"id": "checker-7",
"model": "example-model",
"model_version": "1",
"type": "system"
},
"body": {
"claim": {
"ref": {
"format": "application/json",
"sha256": "f4747f1fa1ad63eaad90e41b587794f8f5c82f9215b9a8e23\
a6c204edabc7d48"
}
}
},
"check": {
"result": "supported",
"rule": "judgement"
},
"evidence": [
{
"format": "application/json",
"sha256": "f4747f1fa1ad63eaad90e41b587794f8f5c82f9215b9a8e23a6\
c204edabc7d48"
},
{
"format": "https://datatracker.ietf.org/doc/draft-gruszka-sign\
ed-evaluation-receipts/",
"sha256": "0c964f45c450a0253fc09eb2316e6ad271944be201c46ebf2c5\
d40e529f35306"
}
],
"key": "verdict-0001",
"kind": "model-decision",
"supersedes": null,
"v": "verdict-record/1",
"version": 1,
"writer": "verifier.example"
}
Its digest, over its canonical bytes, is:
2a85c94e39c56f50185b3b69b8f56026946364631799774a83b4f93399841c45
Wright Expires 6 May 2027 [Page 19]
Internet-Draft Verdict Records November 2026
A.2. A Rule Re-run on Cited Delegation Records
This record checks two delegation hop records from
[I-D.durga-agenttrust-trace-verification], contributed by that
document's author.
Its rule, named urn:example:verdict-record:rule:hop-continuity-
attenuation:v1 for this example, applies items 2 and 3 of Section 5
of that document and not the other checks of that section. It takes
exactly four inputs: parent_hash and parent_digest, which are
strings, and parent_scopes and child_scopes, which are arrays of
strings. Its result is "attenuated_and_linked" when parent_hash
equals parent_digest and every member of child_scopes is a member of
parent_scopes, and "not_attenuated_or_not_linked" otherwise. Here
parent_digest is the digest of the cited bytes, which equals that
document's SHA256(JCS(hop)) only when the item is already in JCS
form, as these are. Each input is bound to an evidence item, so body
holds only the claim, and a checker that reads inputs from evidence
only (Section 4.5) can re-run the rule.
The parent hop, evidence item 0:
=============== NOTE: '\' line wrapping per RFC 8792 ================
{"expires_at":1793463600,"granted_scopes":["mcp:tool:erp.ledger.read\
","mcp:max_rows:100"],"issued_at":1793460000}
The child hop, evidence item 1:
=============== NOTE: '\' line wrapping per RFC 8792 ================
{"expires_at":1793463600,"granted_scopes":["mcp:tool:erp.ledger.read\
"],"issued_at":1793460000,"parent_record_hash":"1336337e88adc67853b8\
0b4998363210dc5c0797cc7b7c722128cecc56a38c14"}
The verdict record, with whitespace added for display:
=============== NOTE: '\' line wrapping per RFC 8792 ================
{
"act_time": "2026-10-31T18:00:00.000000Z",
"actor": {
"id": "agenttrust-trace-checker",
"model": "none",
"model_version": "0",
"type": "system"
},
"body": {
Wright Expires 6 May 2027 [Page 20]
Internet-Draft Verdict Records November 2026
"claim": {
"text": "hop[1] preserves SHA-256 parent_record_hash continuit\
y and monotonic scope attenuation from hop[0]"
}
},
"check": {
"inputs": {
"child_scopes": {
"evidence": 1,
"pointer": "/granted_scopes"
},
"parent_digest": {
"evidence": 0
},
"parent_hash": {
"evidence": 1,
"pointer": "/parent_record_hash"
},
"parent_scopes": {
"evidence": 0,
"pointer": "/granted_scopes"
}
},
"result": "attenuated_and_linked",
"rule": "urn:example:verdict-record:rule:hop-continuity-attenuat\
ion:v1"
},
"evidence": [
{
"format": "https://datatracker.ietf.org/doc/draft-durga-agentt\
rust-trace-verification/",
"sha256": "1336337e88adc67853b80b4998363210dc5c0797cc7b7c72212\
8cecc56a38c14"
},
{
"format": "https://datatracker.ietf.org/doc/draft-durga-agentt\
rust-trace-verification/",
"sha256": "8886795d4403b5cd8820eb813ee0724f1d6d5d383124364b09c\
dc5020639e00c"
}
],
"key": "agenttrust/trace/hop-1",
"kind": "check-run",
"supersedes": null,
"v": "verdict-record/1",
"version": 1,
"writer": "agenttrust-reference-verifier"
}
Wright Expires 6 May 2027 [Page 21]
Internet-Draft Verdict Records November 2026
Its digest, over its canonical bytes, is:
62017940a8ddddd5ddd7ff46c8ff2a6ce79dfcccb775d767b8dff8db0ebe19ed
The hops carry no signature, which the format of
[I-D.durga-agenttrust-trace-verification] requires. A checker that
applies that requirement reports held, at step 2. A checker that
verified these hop records without a signature, which this document
does not yet define (Appendix B), and that held the rule's
specification above, would report released. If the child hop widened
the scopes, the re-run would give a different result and the record
would be refused; if the checker did not hold the child hop, the
record would be held. Test vectors for these cases are being added
to the test vectors for this document (Appendix B).
Appendix B. Open Questions
This section is to be removed before publication.
* Where the work belongs: SCITT, whose Signed Statements can carry a
verdict record, or the proposed AUDIT work, whose Auditor reaches
the determination a verdict record states.
* Whether a record should state what its issuer did not check, and
why, as the decision receipt [PROOFBUNDLE-DR] and
[I-D.gruszka-audit-decision-receipts] do.
* Whether a record should state what evidence would change its
result.
* Whether the canonical form of Section 2.1 should instead be a
profile of the JSON Canonicalization Scheme [RFC8785], with the
same restriction to integers. The two give the same bytes except
in the ordering of member names that contain characters outside
the Basic Multilingual Plane.
* Whether a record should state the cutoff it applied to the arrival
of evidence, so that evidence arriving later is visibly outside
the assessment rather than absent from it (Section 6, [VATE]).
* What it means for an item to verify when it carries no signature
its format requires, as for the delegation hop records of
Appendix A.2, and for a plain media type such as application/json.
* Whether a checker that holds a later version of a cited verdict
record should treat a citation of an earlier version as
superseded, and if so how, given that no step may rely on the
absence of a later version (Section 6).
Wright Expires 6 May 2027 [Page 22]
Internet-Draft Verdict Records November 2026
* Where rule names and their results are registered, so that a rule
such as the one in Appendix A.2 can have a stable name outside an
example.
* Test vectors and a reference checker are published at
https://github.com/diggertydog/verdict-record-test-vectors, with
the open points they exposed. A checker written by a party other
than the author is still wanted. Contributions of evidence items
in other formats, with their digests, are welcome for Appendix A.
Appendix C. Changes Since -01
This section is to be removed before publication.
* A rule input can be bound to an evidence item, by index and JSON
Pointer, so that a re-run reads the cited bytes rather than copies
in body; a checker's policy can read inputs from evidence only
(Section 2.2, Section 4.3, Section 4.5).
* Added Section 6: reassessment as a new record, and how a withdrawn
item reaches a check.
* Rule names should identify a version.
* Stated that a cited verdict record verifies only when released,
how citation depth and nesting depth are counted, that malformed
bytes are refused before any step runs, when a policy is invalid,
and over which bytes the policy digest is taken.
* Added a second example, with delegation hop records as evidence
(Appendix A.2). In -01 such a rule could read only values copied
into body, and a record whose copies disagreed with its cited
evidence was still released; under -02 that remains so unless the
checker's policy reads inputs from evidence only.
* Added related work and open questions.
Appendix D. Changes Since -00
This section is to be removed before publication.
* Pointed to the audit architecture's Auditor in the introduction.
* Stated that a verdict record provides no completeness property.
* Stated which bytes an evidence digest covers for enveloped items.
* Added Related Work and Open Questions.
Wright Expires 6 May 2027 [Page 23]
Internet-Draft Verdict Records November 2026
Author's Address
Darren Wright
Sencor AI S.L.
Spain
Email: [email protected]
Wright Expires 6 May 2027 [Page 24]