Skip to main content

Gravit Verifiable Epistemic Decision Standard
draft-gravit-verifiable-epistemic-decision-00

Document Type Active Internet-Draft (individual)
Author Alex Konviser
Last updated 2026-08-09
RFC stream (None)
Intended RFC status (None)
Formats
Additional resources GitHub Organization
GitHub Repository
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-gravit-verifiable-epistemic-decision-00
Network Working Group                                        A. Konviser
Internet-Draft                                       Gravit Open Network
Intended status: Standards Track                           9 August 2026
Expires: 10 February 2027

             Gravit Verifiable Epistemic Decision Standard
             draft-gravit-verifiable-epistemic-decision-00

Abstract

   This document defines the normative requirements for the Gravit
   decision-making system.  Gravit's core principle is to base all
   definitive decisions exclusively on verifiable epistemic grounds.
   This specification establishes mandatory rules for evidence
   admissibility, epistemic classification, traceability, failure
   handling, auditability, and security in adversarial environments,
   including a taxonomy of admissible evidence and failure outcomes, a
   required Decision Record format, and an explicit threat model with
   corresponding verification mechanisms.  The goal is to ensure that
   all decisions are reproducible, auditable, and resistant to
   manipulation, thereby fostering trust and accountability in
   distributed and high-stakes settings.

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 10 February 2027.

Copyright Notice

   Copyright (c) 2026 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

Konviser                Expires 10 February 2027                [Page 1]
Internet-Draft                 Gravit VEDS                   August 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
   2.  Terminology . . . . . . . . . . . . . . . . . . . . . . . . .   4
   3.  Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . .   5
   4.  Requirements  . . . . . . . . . . . . . . . . . . . . . . . .   5
     4.1.  Epistemic Validity  . . . . . . . . . . . . . . . . . . .   5
       4.1.1.  Epistemic Classification Taxonomy . . . . . . . . . .   5
       4.1.2.  Classification Determination and Reclassification . .   6
       4.1.3.  Requirements  . . . . . . . . . . . . . . . . . . . .   6
     4.2.  Evidence Admissibility  . . . . . . . . . . . . . . . . .   7
       4.2.1.  Required Elements of Admissible Evidence  . . . . . .   7
       4.2.2.  Determination Procedure . . . . . . . . . . . . . . .   7
       4.2.3.  Requirements  . . . . . . . . . . . . . . . . . . . .   8
     4.3.  Traceability  . . . . . . . . . . . . . . . . . . . . . .   8
       4.3.1.  Contribution Attribution Format . . . . . . . . . . .   8
       4.3.2.  Audit Trail Reconstruction Procedure  . . . . . . . .   9
       4.3.3.  Requirements  . . . . . . . . . . . . . . . . . . . .   9
     4.4.  Failure Handling  . . . . . . . . . . . . . . . . . . . .  10
       4.4.1.  Failure State Taxonomy  . . . . . . . . . . . . . . .  10
       4.4.2.  Requirements  . . . . . . . . . . . . . . . . . . . .  11
     4.5.  Auditability and Assurance  . . . . . . . . . . . . . . .  12
       4.5.1.  Decision Record Content . . . . . . . . . . . . . . .  12
       4.5.2.  Timestamping and Record Integrity . . . . . . . . . .  12
       4.5.3.  Retention and Access  . . . . . . . . . . . . . . . .  13
       4.5.4.  Requirements  . . . . . . . . . . . . . . . . . . . .  13
     4.6.  Security and Adversarial Resistance . . . . . . . . . . .  13
       4.6.1.  Threat Model  . . . . . . . . . . . . . . . . . . . .  14
       4.6.2.  Required Verification Mechanisms  . . . . . . . . . .  14
       4.6.3.  Requirements  . . . . . . . . . . . . . . . . . . . .  15
   5.  Conformance . . . . . . . . . . . . . . . . . . . . . . . . .  16
     5.1.  Conformance Levels  . . . . . . . . . . . . . . . . . . .  16
     5.2.  Conformance Statement . . . . . . . . . . . . . . . . . .  16
   6.  Security Considerations . . . . . . . . . . . . . . . . . . .  17
     6.1.  Evidence Manipulation . . . . . . . . . . . . . . . . . .  17
     6.2.  Residual Risks  . . . . . . . . . . . . . . . . . . . . .  17
     6.3.  Opaque Decision Logic . . . . . . . . . . . . . . . . . .  18
   7.  IANA Considerations . . . . . . . . . . . . . . . . . . . . .  18
   8.  Normative References  . . . . . . . . . . . . . . . . . . . .  19

Konviser                Expires 10 February 2027                [Page 2]
Internet-Draft                 Gravit VEDS                   August 2026

   9.  Informative References  . . . . . . . . . . . . . . . . . . .  19
   Appendix A.  Acknowledgments  . . . . . . . . . . . . . . . . . .  19
   Appendix B.  Change Log . . . . . . . . . . . . . . . . . . . . .  19
   Author's Address  . . . . . . . . . . . . . . . . . . . . . . . .  19

1.  Introduction

   Gravit is a decision-making system architected to produce outputs
   that are rigorously grounded in evidence.  The foundational premise
   is that for a decision to be legitimate and trustworthy, it must be
   possible to independently verify the basis upon which it was made.

   The purpose of this standard is to codify this principle into a set
   of concrete, enforceable requirements.  It is intended to prevent
   decisions based on unverifiable assertions, unsupported hypotheses,
   or opaque inference paths.  This standard is particularly relevant
   for technical implementations in distributed, adversarial, or
   regulated environments where the legitimacy of a decision depends on
   its external verifiability and auditability.

   Section 4 organizes the normative requirements into six areas:
   epistemic classification (Section 4.1), evidence admissibility
   (Section 4.2), traceability (Section 4.3), failure handling
   (Section 4.4), auditability (Section 4.5), and security against
   adversarial evidence (Section 4.6).  Each area is specified through,
   as applicable, a taxonomy or model of the relevant concern, a
   procedure or set of mechanisms for applying it, and a list of
   normative requirements.  Several deployment-specific parameters that
   this document deliberately leaves undefined (for example, decision-
   impact thresholds, or a Decision Record retention period) are
   collected under a single Conformance Statement mechanism defined in
   Section 5.

   This document specifies what a conformant Gravit implementation must
   guarantee about the decisions it produces; it does not specify a
   particular protocol for performing the underlying evidence
   verification steps referenced in Section 4.2 and Section 4.6.  This
   document is informed by, and intended to be usable together with,
   prior work referred to as GEVP, which defines one mechanism for
   evidence verification (see [GEVP]).  An implementation satisfying the
   properties required by this document is not required to use GEVP
   specifically, provided its verification mechanism satisfies the
   requirements of Section 4.2 and Section 4.6.

   Requirements comparable to those in this document, particularly the
   marking and detectability of AI-generated or AI-manipulated content,
   are also addressed by external regulatory frameworks, such as Article
   50 of Regulation (EU) 2024/1689 [EU-AI-ACT-ART50].  This document

Konviser                Expires 10 February 2027                [Page 3]
Internet-Draft                 Gravit VEDS                   August 2026

   does not implement or claim conformance with any specific regulatory
   framework; it is cited here only to note that a Gravit
   implementation's evidence, provenance, and traceability guarantees
   may be relevant to satisfying such external obligations, subject to
   independent legal assessment.

2.  Terminology

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD 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.

   For the purposes of this document, the following terms are defined:

   Verifiable Epistemic Ground:  A fact, proof, data point, attestation,
      or inference chain that can be independently and objectively
      checked for validity, integrity, and correctness by a third party,
      without requiring privileged access to internal system states.

   Evidence:  Any artifact, including data, logs, cryptographic proofs,
      attestations, or trace records, that is used to support or justify
      a decision.

   Traceability:  The ability to reconstruct the complete, auditable
      chain of reasoning from a final decision back to the specific set
      of initial epistemic grounds and inference steps that produced it.

   Fallback Mode:  A documented, non-default operating mode in which
      Gravit may operate with reduced decision authority or rely on
      explicitly defined and authorized heuristic processes when normal
      verification procedures fail.  Authorization of a Fallback Mode
      configuration, including the conditions under which it may be
      entered and the parties responsible for approving it, is an
      operational and deployment concern outside the scope of this
      document, but implementations claiming conformance MUST document
      their authorization procedure in their Conformance Statement (see
      Section 5.2).

   Failure State:  One of a fixed set of machine-readable outcomes
      (REJECTED, DEGRADED, or PENDING; see Section 4.4) that Gravit
      reports when it does not produce an unqualified definitive
      decision.  A Failure State is itself a decision outcome and is
      subject to the same traceability requirements as a definitive
      decision (see Section 4.3).

   Decision Record:  The complete, structured artifact that binds a

Konviser                Expires 10 February 2027                [Page 4]
Internet-Draft                 Gravit VEDS                   August 2026

      definitive decision or Failure State to its evidentiary basis,
      provenance metadata, and timestamps, as required by Section 4.5.
      The Decision Record is the unit of storage referenced by all
      retention and audit requirements in this document.

3.  Scope

   This document applies to all decision-making functions performed by
   the Gravit system.  This includes, but is not limited to, direct
   decisions, decisions aggregated from multiple sources, and decisions
   delegated to or influenced by external subsystems.

   This document defines requirements for the verifiability and
   traceability of decisions, not the specific internal inference
   algorithms or machine learning models themselves.  Such algorithms
   are in scope only insofar as their operation affects the system's
   ability to meet the verification, traceability, and conformance
   requirements outlined herein.

4.  Requirements

4.1.  Epistemic Validity

4.1.1.  Epistemic Classification Taxonomy

   Gravit MUST classify every epistemic ground into exactly one of the
   following classes before it may contribute to a decision:

   Verified Fact:  An epistemic ground that has passed the admissibility
      procedure in Section 4.2.2 and whose content has additionally been
      independently confirmed accurate, and not merely attributable and
      intact, through reproduction, cross-corroboration, or equivalent
      means.

   Probabilistic Assessment:  An epistemic ground that assigns a
      likelihood or confidence value to a proposition, derived from a
      documented and reproducible method (e.g., a statistical inference
      procedure or a forecasting model with disclosed inputs), without
      asserting the proposition as definitively true.

   Speculative Claim:  An epistemic ground whose truth cannot currently
      be verified or falsified through reproducible means, including
      unconfirmed reports, unattributed assertions, or model outputs
      lacking a documented confidence method.

   The Verifiable Basis requirement of Section 4.1.3 MUST be satisfied
   using Verified Facts, Probabilistic Assessments, or a combination
   thereof only; a Speculative Claim MUST NOT, alone or in combination

Konviser                Expires 10 February 2027                [Page 5]
Internet-Draft                 Gravit VEDS                   August 2026

   with other Speculative Claims, be treated as satisfying that
   requirement.  Where a Probabilistic Assessment serves as part of the
   basis for a definitive decision, its confidence value SHOULD meet or
   exceed a documented, decision-class-specific threshold consistent
   with the Impact-Based Evidence Strength requirement in Section 4.2.3.

4.1.2.  Classification Determination and Reclassification

   1.  *Pre-Use Classification*: Gravit MUST determine the epistemic
       classification of a ground as a discrete step, completed before
       that ground is used in decision-making, in the same manner as the
       admissibility determination in Section 4.2.2.

   2.  *No Aggregation Laundering*: Gravit MUST NOT treat the
       aggregation of multiple Speculative Claims as producing an
       epistemic status equivalent to a Verified Fact or Probabilistic
       Assessment, unless that aggregation is performed by a documented,
       independently verifiable method that itself qualifies as a
       Probabilistic Assessment under Section 4.1.1.

   3.  *Reclassification on Contrary Evidence*: Gravit MUST reclassify
       an epistemic ground upon receipt of credible evidence
       contradicting its prior classification, and MUST re-evaluate any
       definitive decision for which the reclassified ground was part of
       the necessary basis under Section 4.1.1.

   4.  *Reclassification Record*: Gravit MUST record a reclassification
       event, including the prior and new classification and the
       triggering evidence, in the Decision Record of any decision
       affected by it, per Section 4.5.1.

4.1.3.  Requirements

   1.  *Verifiable Basis*: Gravit MUST base every definitive decision on
       verifiable epistemic grounds.

   2.  *Prohibition on Unverifiable Inputs*: Gravit MUST NOT base any
       decision, in whole or in part, on an unverifiable assertion,
       unsupported hypothesis, or unvalidated inference.

   3.  *Epistemic Classification*: Gravit MUST clearly distinguish
       between and internally label different types of epistemic
       grounds, per Section 4.1.1.

   4.  *Preference for Reproducibility*: Gravit SHOULD prefer evidence
       that can be independently reproduced or verified by a third
       party.

Konviser                Expires 10 February 2027                [Page 6]
Internet-Draft                 Gravit VEDS                   August 2026

   5.  *Use of Speculative Signals*: Gravit MAY use speculative signals
       only as non-decisive inputs, provided they do not determine the
       final outcome of a definitive decision without subsequent
       verification.

4.2.  Evidence Admissibility

4.2.1.  Required Elements of Admissible Evidence

   Gravit MUST NOT treat an item as admissible unless it carries, or can
   be resolved to, all of the following:

   a.  *Provenance*: an identifiable source or origin (e.g., signer
       identity, system of record, sensor identifier) sufficient to
       attribute the item to a specific, nameable origin.

   b.  *Integrity*: a means of detecting modification since origination
       (e.g., a cryptographic digest or signature covering the item's
       content).

   c.  *Freshness Bound*: a timestamp or validity interval sufficient to
       determine whether the item is stale relative to the decision
       being made; Gravit MUST define, per deployment, the maximum age
       at which an otherwise-valid item ceases to be admissible for a
       given decision class.

   d.  *Interpretability*: a documented schema, encoding, or semantics
       sufficient for the item's content to be unambiguously interpreted
       without reliance on undocumented local convention.

   An item missing any element in this list MUST be treated as
   inadmissible under Section 4.2.3, regardless of the apparent
   plausibility of its content.

4.2.2.  Determination Procedure

   1.  *Pre-Use Determination*: Gravit MUST determine admissibility as a
       discrete step that completes, with a recorded outcome, before the
       item is used in any decision-making process; admissibility MUST
       NOT be determined implicitly as a side effect of the decision
       computation itself.

   2.  *Binary Outcome*: The admissibility determination for a given
       item MUST produce exactly one of two outcomes (admissible or
       inadmissible) for a given decision context; Gravit MUST NOT
       propagate a "partially admissible" item into decision-making
       without first resolving it to one of these two outcomes.

Konviser                Expires 10 February 2027                [Page 7]
Internet-Draft                 Gravit VEDS                   August 2026

   3.  *Inadmissibility Recording*: Gravit MUST record the specific
       element from Section 4.2.1 that caused an item to be found
       inadmissible, for inclusion in the Decision Record if the item's
       absence contributed to a REJECTED or DEGRADED outcome (see
       Section 4.4.1).

   4.  *Non-Reuse of Rejected Items*: Gravit MUST NOT reuse an item
       found inadmissible in a later determination for the same decision
       context unless the specific deficiency identified under
       requirement 3 of this section has been remedied.

4.2.3.  Requirements

   1.  *Source and Proof Linkage*: Gravit MUST reject any input that
       cannot be linked to a verifiable source, a cryptographic proof,
       or a traceable inference chain.

   2.  *Pre-Use Validation*: Gravit MUST validate the admissibility of
       evidence before using it in any decision-making process, per
       Section 4.2.2.

   3.  *Admissibility Criteria*: Gravit MUST NOT treat an item as
       admissible evidence if its provenance, integrity, or
       interpretation cannot be established, per Section 4.2.1.

   4.  *Evidence Preservation*: Gravit SHOULD preserve evidence,
       including items found inadmissible, in a form suitable for later
       audit and independent review.

   5.  *Impact-Based Evidence Strength*: Gravit MAY require stronger
       forms of evidence (e.g., cryptographic proofs instead of digital
       signatures) for decisions with higher potential impact, provided
       the elevated requirement is defined in terms of the elements in
       Section 4.2.1 and not as an undocumented, ad hoc criterion.

4.3.  Traceability

4.3.1.  Contribution Attribution Format

   Where Gravit records the contribution of an evidence item to a
   decision, per the Contribution Attribution requirement in
   Section 4.3.3, that record MUST include:

   a.  a reference to the specific evidence item, resolvable to the item
       as determined admissible under Section 4.2.2;

   b.  the item's epistemic classification at the time of use, per
       Section 4.1.1;

Konviser                Expires 10 February 2027                [Page 8]
Internet-Draft                 Gravit VEDS                   August 2026

   c.  a description of the role the item played in the outcome (e.g.,
       necessary condition, weighted factor, corroborating support,
       disqualifying condition); and

   d.  where the decision process is quantitative, the item's individual
       weight or score contribution to the outcome, to the extent the
       process supports isolating that contribution.

   An attribution record that identifies only that an item was "used,"
   without the role or weight information in this list, MUST NOT be
   treated as satisfying the Contribution Attribution requirement.

4.3.2.  Audit Trail Reconstruction Procedure

   To satisfy the Audit Trail Completeness requirement in Section 4.3.3,
   Gravit SHOULD support, at minimum, the following reconstruction steps
   from a Decision Record (see Section 4.5.1) alone, without reliance on
   Gravit's internal runtime state:

   1.  identification of every evidence item referenced in the Decision
       Record;

   2.  independent verification of each item's admissibility, per
       Section 4.2.1;

   3.  independent verification of each item's epistemic classification,
       per Section 4.1.1; and

   4.  confirmation that the recorded decision logic maps the referenced
       evidence, with its recorded roles or weights per Section 4.3.1,
       to the recorded outcome.

   An implementation that cannot support step 4 of this procedure for a
   given decision SHOULD record that limitation explicitly as part of
   the assurance boundaries required by Section 4.5.4.

4.3.3.  Requirements

   1.  *Decision-Evidence Relationship*: Gravit MUST maintain a
       traceable relationship between each definitive decision and the
       evidentiary basis used to produce it.

   2.  *Contribution Attribution*: Gravit MUST record which specific
       evidence items influenced a decision and, where applicable, the
       nature and extent of their contribution to the final result, per
       Section 4.3.1.

Konviser                Expires 10 February 2027                [Page 9]
Internet-Draft                 Gravit VEDS                   August 2026

   3.  *Segregation of Premises*: Gravit MUST NOT merge verified and
       unverified premises in a way that prevents their clear and
       unambiguous separation during an audit.

   4.  *Audit Trail Completeness*: Gravit SHOULD provide an audit trail
       sufficient to fully reconstruct the decision process, per
       Section 4.3.2.

   5.  *Trace Fidelity*: A recorded trace or explanation MUST NOT
       represent a reasoning path that did not, in fact, produce the
       recorded outcome; Gravit MUST NOT substitute a plausible post-hoc
       rationalization for the actual decision logic in any record used
       to satisfy this section.

   6.  *Limited Disclosure*: Gravit MAY expose only a limited trace when
       full disclosure would violate security, privacy, or policy
       constraints.

4.4.  Failure Handling

4.4.1.  Failure State Taxonomy

   Gravit MUST classify every non-definitive outcome into exactly one of
   the following Failure States, and MUST NOT define additional outcome
   types that are not reducible to one of these three without updating
   this specification:

   REJECTED:  No verifiable epistemic basis exists and none is expected
      to become available without a material change in inputs.  Gravit
      MUST NOT automatically retry a REJECTED request with the same
      evidence set.

   DEGRADED:  A decision was produced, but only by operating in an
      authorized Fallback Mode (see Section 2) with reduced decision
      authority.  A DEGRADED outcome MUST be labeled as such and MUST
      NOT be represented to downstream consumers as equivalent to an
      unqualified definitive decision.

   PENDING:  Verification could not be completed within the applicable
      time or resource bounds, but a verifiable basis may become
      available (e.g., outstanding corroboration, an unresolved proof).
      A PENDING outcome MUST carry either a retry policy or an explicit
      expiry, after which it MUST transition to REJECTED.

Konviser                Expires 10 February 2027               [Page 10]
Internet-Draft                 Gravit VEDS                   August 2026

4.4.2.  Requirements

   1.   *Refusal with No Basis*: Gravit MUST refuse to produce a
        definitive decision when no verifiable epistemic basis is
        available, and MUST classify the refusal as REJECTED per
        Section 4.4.1.

   2.   *Fail-Closed Default*: When Gravit cannot determine which
        Failure State applies, or cannot complete classification within
        the applicable time bound, it MUST default to REJECTED rather
        than producing an unqualified definitive decision or silently
        defaulting to a permissive outcome.

   3.   *Degraded Mode Entry*: Gravit MUST enter DEGRADED or REJECTED
        when evidence verification fails for a required decision, and
        MUST NOT enter DEGRADED except through an authorized Fallback
        Mode configuration as defined in Section 2.

   4.   *Prohibition on Unverified Heuristics*: Gravit MUST NOT
        substitute heuristics or unverified approximations for formal
        verification unless explicitly operating in a documented and
        authorized fallback mode, as defined in Section 2.

   5.   *No Silent Retry with Weakened Evidence*: Gravit MUST NOT re-
        attempt a decision using a relaxed admissibility threshold (see
        Section 4.2) as an automatic response to a REJECTED or DEGRADED
        outcome; any change to the admissibility threshold MUST be an
        explicit, separately authorized action.

   6.   *PENDING Bounds*: Gravit MUST bound the duration or resource
        cost of a PENDING outcome and MUST transition it to REJECTED
        upon expiry, as required by Section 4.4.1.

   7.   *Failure Record*: Gravit MUST produce a Decision Record (see
        Section 2) for every REJECTED, DEGRADED, or PENDING outcome,
        subject to the same traceability requirements in Section 4.3 as
        a definitive decision.

   8.   *Explanation of Refusal*: Gravit SHOULD return a clear and
        actionable explanation for its refusal to decide when such
        disclosure is safe and operationally useful, and SHOULD include
        the applicable Failure State in that explanation.

   9.   *Request for Additional Evidence*: Gravit MAY request additional
        evidence from trusted sources before attempting to decide again,
        and MAY represent the interim state as PENDING while such
        evidence is outstanding.

Konviser                Expires 10 February 2027               [Page 11]
Internet-Draft                 Gravit VEDS                   August 2026

   10.  *Repeated-Failure Signaling*: Gravit MAY escalate or flag a
        source or evidence path that produces repeated REJECTED outcomes
        for subsequent handling under Section 4.6, but such signaling
        MUST NOT itself be used as grounds for a definitive decision.

4.5.  Auditability and Assurance

4.5.1.  Decision Record Content

   Each Decision Record (see Section 2) MUST contain, at minimum:

   a.  a unique, stable identifier for the decision or Failure State;

   b.  the outcome (definitive decision value, or Failure State per
       Section 4.4.1);

   c.  references to every evidence item that influenced the outcome,
       and its individual contribution where determinable, per
       Section 4.3;

   d.  the epistemic classification (verified fact, probabilistic
       assessment, speculative claim; see Section 4.1) of each
       referenced evidence item;

   e.  a verifiable timestamp of both evidence collection and decision
       production, per Section 4.5.2;

   f.  an explicit statement of assurance boundaries, per requirement 3
       of this section, when the decision is DEGRADED or relies on any
       non-reproducible evidence.

4.5.2.  Timestamping and Record Integrity

   1.  *Verifiable Timestamps*: Gravit MUST associate each Decision
       Record with a timestamp that is itself verifiable (e.g., derived
       from a trusted time source or a cryptographic timestamping
       scheme), and MUST NOT rely solely on an unauthenticated local
       clock value for records used in audit or dispute resolution.

   2.  *Tamper Evidence*: Gravit SHOULD store Decision Records in a form
       that is tamper-evident (e.g., via cryptographic hashing or hash
       chaining across records), such that undetected post-hoc
       modification of a stored Decision Record is computationally
       infeasible.

Konviser                Expires 10 February 2027               [Page 12]
Internet-Draft                 Gravit VEDS                   August 2026

   3.  *Immutability After Finalization*: Gravit MUST NOT modify a
       finalized Decision Record in place; corrections MUST be
       represented as new, separately timestamped records that reference
       the record they supersede.

4.5.3.  Retention and Access

   1.  *Retention Period*: An implementation claiming conformance MUST
       define and document a minimum retention period for Decision
       Records, commensurate with the operational, regulatory, or
       contractual audit requirements of its deployment context; this
       document does not mandate a specific duration.

   2.  *Machine-Readable Structure*: Gravit SHOULD store Decision
       Records in a structured, machine-readable format to support
       automated audit tooling and independent re-verification.

   3.  *Access Control on Audit Data*: Gravit SHOULD restrict write
       access to stored Decision Records to the decision-production
       process itself, and SHOULD make read access available to
       designated auditors independent of that process.

4.5.4.  Requirements

   1.  *Decision Auditability*: Gravit MUST ensure that definitive
       decisions are auditable against the full, recorded evidence set.

   2.  *Provenance Metadata*: Gravit MUST preserve sufficient metadata,
       per Section 4.5.1, to support post hoc verification of decision
       provenance.

   3.  *Explicit Assurance Boundaries*: Gravit SHOULD make assurance
       boundaries explicit, clearly documenting what was verified, what
       was assumed, and what remains unknown.

   4.  *External Verification Support*: Gravit MAY publish partial
       proofs or attestations to support external verification.

   5.  *Minimizing Opaque Paths*: Gravit SHOULD minimize the use of
       opaque decision paths where a verifiable path is feasible.

   6.  *Independent Re-Verification*: Gravit SHOULD structure Decision
       Records such that an independent auditor with access to the
       referenced evidence can reconstruct the outcome without access to
       Gravit's internal runtime state.

4.6.  Security and Adversarial Resistance

Konviser                Expires 10 February 2027               [Page 13]
Internet-Draft                 Gravit VEDS                   August 2026

4.6.1.  Threat Model

   Gravit MUST be designed and operated under the assumption that an
   adversary may possess some or all of the following capabilities with
   respect to evidence entering the system:

   Forgery:  The ability to generate evidence-like artifacts purporting
      to originate from a legitimate source that did not, in fact,
      produce them.

   Tampering:  The ability to modify the content of evidence in transit
      or at rest, between the time it was originated and the time it is
      used in a decision.

   Replay:  The ability to capture previously valid evidence and
      resubmit it outside the context, decision, or freshness window
      (see Section 4.2.1) for which it was originally produced.

   Source Spoofing:  The ability to cause evidence to be attributed to a
      source other than its actual origin, including impersonation of a
      trusted source's identity or credentials.

   Collusion:  The ability to cause multiple, ostensibly independent
      sources to act in coordination, so as to satisfy a corroboration
      requirement without providing genuinely independent confirmation.

   Verification-Path Flooding:  The ability to submit high volumes of
      low-quality, malformed, or spurious evidence with the intent of
      exhausting verification resources or inducing the system to relax
      its admissibility or verification standards under load.

   This list defines the minimum threat coverage required for
   conformance; it does not preclude an implementation from defending
   against additional threats specific to its deployment context.

4.6.2.  Required Verification Mechanisms

   For each threat identified in Section 4.6.1, Gravit MUST apply the
   corresponding countermeasure:

   1.  *Cryptographic Integrity Binding*: To defend against Tampering,
       Gravit MUST verify a cryptographic digest or signature covering
       the evidence content, per the Integrity element defined in
       Section 4.2.1, before the evidence is used.

   2.  *Authenticity Binding*: To defend against Forgery and Source
       Spoofing, Gravit MUST verify a binding between the evidence and a
       specific, nameable origin, per the Provenance element defined in

Konviser                Expires 10 February 2027               [Page 14]
Internet-Draft                 Gravit VEDS                   August 2026

       Section 4.2.1, using credentials or identity proofs that the
       claimed source could plausibly, and only the claimed source could
       plausibly, have produced.

   3.  *Replay Defense*: To defend against Replay, Gravit MUST enforce
       the Freshness Bound element defined in Section 4.2.1 and SHOULD
       additionally bind evidence to a nonce, sequence number, or
       decision-specific context value where the evidence format
       supports it, such that a previously valid item cannot be
       revalidated for a different decision without detection.

   4.  *Independent Corroboration Threshold*: To defend against
       Collusion, Gravit SHOULD require corroboration from more than one
       source for decisions above a defined impact threshold (see the
       Impact-Based Evidence Strength requirement in Section 4.2.3).
       Gravit MUST NOT treat two sources as independent for the purpose
       of satisfying a corroboration requirement when they are known to
       share a controlling party, a common infrastructure dependency, or
       another relationship that would allow a single actor to influence
       both.

   5.  *Resource Bounding Under Load*: To defend against Verification-
       Path Flooding, Gravit MUST bound the verification resources
       committed per source or per request, and MUST respond to resource
       exhaustion by producing a REJECTED or PENDING Failure State (see
       Section 4.4.1) for the affected requests rather than by relaxing
       admissibility criteria or verification depth to relieve load.

4.6.3.  Requirements

   1.  *Adversarial Assumption*: Gravit MUST assume that all incoming
       evidence may be incomplete, stale, malformed, or adversarially
       manipulated, consistent with the threat model in Section 4.6.1.

   2.  *Integrity and Authenticity Verification*: Gravit MUST verify the
       integrity and authenticity of evidence before relying on it, per
       Section 4.6.2.

   3.  *Rejection of Tampered Evidence*: Gravit MUST NOT accept forged,
       replayed, or tampered evidence as a valid basis for decision-
       making.

   4.  *Uniform Failure Path for Verification Failure*: Gravit MUST
       treat evidence that fails integrity, authenticity, or replay
       verification as inadmissible under Section 4.2 and process it
       through the ordinary admissibility and failure-handling paths
       defined in Section 4.4, rather than through a separate,
       undocumented error path.

Konviser                Expires 10 February 2027               [Page 15]
Internet-Draft                 Gravit VEDS                   August 2026

   5.  *Source-Based Verification*: Gravit SHOULD apply stronger
       verification checks to evidence originating from untrusted or
       partially trusted sources.

   6.  *Isolation of Trust Levels*: Gravit SHOULD isolate the processing
       path for evidence from untrusted or partially trusted sources
       from privileged internal state, such that a flaw in the handling
       of a single untrusted evidence item cannot by itself grant
       broader access to, or control over, the decision-making process.

   7.  *Source Confidence vs. Verification*: Gravit MAY assign
       confidence levels to sources as an auxiliary signal for
       prioritizing or scheduling verification work (for example,
       determining the order in which evidence items are checked, or the
       amount of independent corroboration sought).  Such confidence
       levels MUST NOT be used as a substitute for the direct
       verification of the evidence itself, and MUST NOT independently
       increase or decrease the admissibility of evidence as determined
       under Section 4.2.

5.  Conformance

5.1.  Conformance Levels

   An implementation of Gravit conforms to this specification only if it
   satisfies all requirements marked with the keywords MUST, MUST NOT,
   REQUIRED, SHALL, or SHALL NOT.

   Requirements marked SHOULD, SHOULD NOT, or RECOMMENDED are considered
   best practices; any deviation from them MUST be explicitly justified
   by documented operational constraints and risks, and that
   justification MUST be recorded in the implementation's Conformance
   Statement (see Section 5.2).

   Requirements marked MAY or OPTIONAL are optional capabilities and do
   not affect an implementation's conformance.

5.2.  Conformance Statement

   This document deliberately leaves a number of parameters to be
   defined at the deployment level rather than fixing them normatively.
   An implementation claiming conformance MUST publish and maintain a
   Conformance Statement that documents, at minimum:

   a.  its Fallback Mode authorization procedure, including the
       conditions under which Fallback Mode may be entered and the
       parties responsible for approving it (see Section 2 and
       Section 4.4.1);

Konviser                Expires 10 February 2027               [Page 16]
Internet-Draft                 Gravit VEDS                   August 2026

   b.  its decision-impact classes and, for each, the corresponding
       evidence-strength requirements and, where applicable, the minimum
       Probabilistic Assessment confidence threshold (see Section 4.2.3
       and Section 4.1.1);

   c.  its Decision Record retention period (see Section 4.5.3);

   d.  its criteria for treating two sources as independent for the
       purpose of the Independent Corroboration Threshold mechanism (see
       Section 4.6.2); and

   e.  any SHOULD-level or RECOMMENDED-level deviations in effect and
       their justification, per Section 5.1.

   An implementation MUST NOT claim conformance to this specification
   while any parameter listed in this section remains undocumented.  The
   Conformance Statement is a deployment-level artifact and is distinct
   from the Decision Record defined in Section 4.5.1; this document does
   not mandate a specific format or distribution mechanism for it.

6.  Security Considerations

   This section discusses security considerations in terms of the threat
   model defined in Section 4.6.1, the residual risks that remain after
   applying the mitigations in Section 4.6.2, and considerations that
   fall outside the scope of the normative requirements in Section 4.

6.1.  Evidence Manipulation

   The primary security risk for systems like Gravit is the acceptance
   of maliciously crafted or corrupted evidence, which could lead to
   incorrect decisions.  Implementations MUST treat all incoming
   evidence as untrusted until it has successfully passed the
   verification requirements defined in Section 4.2 and Section 4.6,
   covering, at minimum, the Forgery, Tampering, Replay, and Source
   Spoofing threat classes identified in Section 4.6.1.

6.2.  Residual Risks

   The mitigations required by this document reduce, but do not
   eliminate, the risks associated with the threat model in
   Section 4.6.1.  Implementers and deployers should be aware of the
   following residual risks:

   *  *Fail-Closed as an Availability Target*: The Fail-Closed Default
      and Resource Bounding Under Load requirements (see Section 4.4.2
      and Section 4.6.2) mean that an adversary who cannot forge
      admissible evidence may instead attempt a denial-of-service attack

Konviser                Expires 10 February 2027               [Page 17]
Internet-Draft                 Gravit VEDS                   August 2026

      by flooding Gravit with low-quality or malformed evidence, with
      the goal of consuming verification resources and forcing
      legitimate decisions into REJECTED or PENDING.  This document
      deliberately accepts unavailability as the safe failure mode in
      preference to producing an unverified decision; deployers
      requiring resilience against this specific attack SHOULD apply
      deployment-level defenses (e.g., request prioritization, per-
      source rate limiting, or capacity overprovisioning) in addition
      to, and outside the scope of, the requirements in this document.

   *  *Collusion Resistance Is Not Absolute*: The Independent
      Corroboration Threshold requirement (see Section 4.6.2) is stated
      at the SHOULD level and depends on an implementation's ability to
      correctly identify relationships between sources.  An adversary
      capable of establishing sources that appear independent by every
      criterion the implementation checks, but are not in fact
      independently controlled, can defeat corroboration.  Deployments
      supporting decisions with severe potential impact SHOULD elevate
      this requirement to a local MUST and invest in stronger,
      deployment-specific independence criteria.

   *  *Trust Anchor and Credential Management*: The Authenticity Binding
      mechanism (see Section 4.6.2) presumes the existence of a
      mechanism for establishing and distributing the credentials, keys,
      or identity proofs against which source authenticity is checked.
      The provisioning, rotation, revocation, and compromise recovery of
      such trust anchors are deployment-specific security concerns that
      this document does not define; an otherwise-conformant
      implementation remains vulnerable to Forgery and Source Spoofing
      if its trust anchor management is weak.

6.3.  Opaque Decision Logic

   A second significant risk is the presence of hidden or opaque
   decision logic that prevents external auditing and obscures the
   reasoning process.  To address this, Gravit implementations SHOULD
   retain comprehensive provenance and metadata, per Section 4.5.1, to
   support later forensic review, especially in systems where decisions
   affect safety, access control, governance, or financial outcomes.
   Relying on heuristics or fallback modes outside of documented and
   justified circumstances, contrary to the Prohibition on Unverified
   Heuristics requirement in Section 4.4.2, introduces brittleness and
   should be considered a security vulnerability in its own right.

7.  IANA Considerations

   This document has no actions for IANA.

Konviser                Expires 10 February 2027               [Page 18]
Internet-Draft                 Gravit VEDS                   August 2026

8.  Normative References

   [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>.

   [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>.

9.  Informative References

   [GEVP]     Konviser, A., "Gravit Evidence Verification Protocol
              (GEVP)", Work in Progress, Internet-Draft, draft-gravit-
              gevp-06, 2026,
              <https://datatracker.ietf.org/doc/draft-gravit-gevp/>.

   [EU-AI-ACT-ART50]
              European Parliament and Council of the European Union,
              "Regulation (EU) 2024/1689 of the European Parliament and
              of the Council of 13 June 2024 laying down harmonised
              rules on artificial intelligence (Artificial Intelligence
              Act), Article 50 (Transparency obligations for providers
              and deployers of certain AI systems)", 13 June 2024,
              <https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng>.

Appendix A.  Acknowledgments

   This specification was informed by the requirement language defined
   in BCP 14 [RFC2119] [RFC8174] and the general industry need for
   transparent, traceable, and auditable decision-making systems in
   adversarial environments.

Appendix B.  Change Log

   [This section to be removed by the RFC Editor before publication;
   retained here for -00 tracking purposes.]

   -00: Initial version.

   This document is a companion policy standard to draft-gravit-gevp
   (currently at -06, POSTED), which is maintained as a separate
   document line and is not renumbered or superseded by this document's
   version counter.

Author's Address

Konviser                Expires 10 February 2027               [Page 19]
Internet-Draft                 Gravit VEDS                   August 2026

   Alex Konviser
   Gravit Open Network
   CH- Zurich
   Switzerland
   Email: ietf@gravit.space
   URI:   https://gravit.space/

Konviser                Expires 10 February 2027               [Page 20]