Internet-Draft Causal Diagnostics for Semantic Drift November 2026
Das Expires 20 May 2027 [Page]
Workgroup:
Independent Submission
Internet-Draft:
draft-das-causal-semantic-handoff-diagnostics-00
Published:
Intended Status:
Informational
Expires:
Author:
S. Das
Independent Inventor

Causal Diagnostics for Semantic Drift in Autonomous Agent Communications: Counterfactual Testing, Source-Role Separation, and Evidence-Bounded Handoff Localization

Abstract

Multi-agent AI systems pass instructions, facts and tool outputs through summaries, translations, delegations and protocol conversions. When a downstream proposed operation departs from the originating instruction, existing records may show that a message was sent, authenticated or traced, yet not show whether a lower-trust source's instruction-like statement influenced a governing condition, whether an intermediary dropped the condition, or whether the receiver misapplied it. This document describes a diagnostic architecture, and an optional transport-neutral evidence model, for investigating such semantic drift. It uses typed originating task contracts, condition-scoped source roles, versioned dependency graphs, handoff lineage, matched counterfactual tests that vary legitimate facts and unauthorized instruction-like content separately, adaptive witness selection, uncertainty-aware comparison, and evidence-bounded localization that reports an earliest observable divergence, a bounded interval, or abstains.

The architecture is diagnostic only. It does not grant, veto or finalize any execution; authorization stays with separate systems. The document is an individual, informational proposal; it makes no novelty, compliance, adoption or performance claim, and reports no experimental results beyond those the underlying filing itself states.

Source and Patent Information

Technical source: Indian Provisional Patent Application 202631134241, dated 8 November 2026 (as reported by the applicant), titled "Systems and Methods for Causal Semantic Handoff Verification in Autonomous Agent Communications". Foundational reference: PCT/IB2026/055615, filed 4 June 2026, published 16 July 2026 as WO 2026/150382 A2. See Section 27.

Where to read the full PCT filing: search "Sangam Das" on WIPO PATENTSCOPE; search PCT/IB2026/055615 (pctib202655615); or read the PCT file directly in Technical Disclosure Commons at https://www.tdcommons.org/dpubs_series/12542/.

Related public research: https://doi.org/10.5281/zenodo.24076231 (see [ZENODO-CAUSAL]).

The IPR disclosure for this document will be made at the appropriate time through the official IETF process and in accordance with IETF guidelines (see Section 27).

AI-assisted drafting was used. Readers must verify patent numbers, dates, claim text, product capabilities and standards references against primary sources.

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 20 May 2027.

▲

Table of Contents

1. Introduction

Autonomous agents increasingly exchange work: an agent summarizes a request, another retrieves documents or calls tools, a third proposes an operation. At each step the meaning of the originating instruction can change without any transport or signature failure. A price limit may be omitted in a summary, a supplier's statement that the budget has been raised may be treated as a governing instruction, a converter may rewrite a field, or a receiver may restate a rule correctly yet emit a non-compliant proposal.

This document describes a diagnostic architecture for investigating such changes. Its distinguishing technique is paired counterfactual testing that separates legitimate factual influence from unauthorized instruction-like influence by source role, combined with adaptive selection of the tests worth running and with localization that states its own uncertainty.

The architecture does not decide whether any operation proceeds. Message authenticity, integrity, authorization and execution control belong to other systems; this one supplies evidence about where observable meaning changed and how well that is supported.

1.1. Problem Statement

A single incorrect proposed operation can be compatible with several explanations: an earlier agent omitted a limit; the receiver misread a received limit; an external document falsely asserted a change; a counteroffer was treated as accepted; a stale fact was used; the receiver understood the instruction but generated inconsistent tool arguments; or several of these together. A final-outcome comparison, a hash mismatch, a provenance label or a state-freshness check each identifies part of the situation but none separates these explanations.

Identifying the earliest observable discrepancy is not the same as proving the first underlying causal event, and sensitivity of an agent to an injected statement is not proof that the statement caused a particular historical incident. The architecture preserves these distinctions.

1.2. Scope and Non-Goals

  • In scope: diagnosis of condition-relevant instruction interpretation changes across agent communications; a diagnostic evidence model.
  • Non-goal: granting, withholding or vetoing execution authority; that stays with separate authorization and execution-control systems.
  • Non-goal: reconstructing hidden reasoning of an agent from an external interface.
  • Non-goal: a new agent protocol, message format or change to A2A, MCP or any other protocol.
  • Non-goal: a claim of solving alignment, universal prompt-injection detection, legal compliance or novelty.

2. Terminology and Conventions

Requirement words are used only where interoperability behavior is intentionally specified (Section on diagnostic evidence structures). Where the underlying disclosure only presents implementation alternatives, this document uses descriptive language. The key words MUST, MUST NOT, SHOULD, SHOULD NOT and MAY in uppercase are to be interpreted as in BCP 14 [RFC2119] [RFC8174].

Originating Task Contract

Versioned typed representation of the conditions that materially affect the intended operation.

Condition

A typed predicate of the contract, for example a maximum amount or a prerequisite.

Source role

What a given source is eligible to do with respect to a given condition: supply facts, propose, or modify.

Handoff

Transfer or transformation of task information across a communication boundary.

Witness

A matched, admissible diagnostic test definition.

Non-effectuating execution

Execution of a diagnostic input in a context where no real external effect is produced.

Earliest observable divergence

First observed boundary at which a condition's representation differs from the reference; not necessarily the first underlying cause.

Fault interval

Bounded range of transitions between the last observed preserving boundary and the first observed divergent boundary.

Abstention

A finding that the evidence does not support a conclusion.

Diagnostic finding

Evidence-bounded output of this architecture; not an authorization decision.

3. Twelve Distinctions

The architecture treats the following as separate questions. Evidence on one does not answer another.

  1. Message authenticity: who sent it.
  2. Message integrity: whether it changed in transit.
  3. Source-role eligibility: whether this source may modify this condition.
  4. Preservation of task conditions: whether each condition survived the handoff.
  5. Observed interpretation: what an agent's observable response indicates.
  6. Proposed-operation correspondence: whether the proposal matches the conditions.
  7. Factual sensitivity: legitimate dependence of behavior on a fact.
  8. Unauthorized-instruction sensitivity: dependence on an instruction-like claim from a source not eligible to modify.
  9. Earliest observable divergence: first observed boundary of difference.
  10. Historical cause versus experimentally supported influence: a test shows influence under test conditions, not necessarily the cause of a past incident.
  11. Unresolved evidence: what could not be established.
  12. Execution authorization: remains outside this architecture.

4. Architecture Overview

The architecture comprises components A to J. A single pipeline is shown in <xref target="fig-main"/>; it is a diagnostic pipeline, not an authority pipeline.

 Originating task --> [A] Typed task contract
                           |
 Source segments ----> [B] Condition-scoped source-role model
                           |
                      [C] Versioned dependency graph
                           |
 Agent handoffs -----> [D] Handoff observation and lineage
                           |
        +------------------+------------------+
        |                  |                  |
   [E] Counterfactual  [F] Adaptive      [G] Uncertainty-aware
       processor           witness           behavioral
   (non-effectuating)      selector          comparator
        +------------------+------------------+
                           |
                      [H] Evidence-bounded fault localizer
                           |
                      [I] Diagnostic evidence exporter
                           |
                      [J] Optional correction / regression
                           |
                           v
        DIAGNOSTIC FINDING (evidence, intervals, abstention)
   - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
   Separate systems keep execution authorization. The finding
   is one input at most; it grants and vetoes nothing.
Figure 1: Diagnostic pipeline (not an authority pipeline)
A. Typed originating task contract

Builds versioned typed predicates, with unresolved conditions kept separate.

B. Condition-scoped source-role model

Per-condition, per-source eligibility; separates authentication from modification eligibility.

C. Versioned dependency graph

Links conditions, facts, source segments and derived messages.

D. Handoff observation and lineage

Records message versions and projects them onto conditions.

E. Counterfactual diagnostic processor

Builds and runs admissible matched witnesses in a non-effectuating context.

F. Adaptive witness selector

Chooses tests by expected usefulness, cost, feasibility and safety.

G. Uncertainty-aware behavioral comparator

Compares paired observations with intervals and decision categories.

H. Evidence-bounded fault localizer

Reports earliest observable divergence, bounded interval or abstention.

I. Diagnostic evidence exporter

Packages findings with per-claim status and optional selective disclosure.

J. Optional correction and regression processor

Tests candidate corrections and regressions.

Embodiment cross-references: A=E01; B=E02; C=E03; D=E04 to E06 and E12; E=E07; F=E08; G=E09; H=E10, E11, E13, E14; I=E06 and E18; J=E16, E17; temporal validity=E15 (see <xref target="app-cross"/>).

5. Typed Task Conditions and Source Roles

Source authentication answers who sent a segment. Source-role eligibility answers whether that source may modify a particular condition. A supplier may quote a price (a fact) but not raise the customer's budget (a condition). Roles are condition-scoped and are carried through quotation, translation and summarization: a derived message does not gain more eligibility than its source had merely because an intermediary produced it.

 Message from supplier agent:
 +----------------------------------------------------------+
 | "Price is INR 2,500."        -> FACT segment             |
 | "Budget is now INR 5,000."   -> MODIFICATION CLAIM       |
 | "Proceed with booking."      -> INSTRUCTION-LIKE REQUEST |
 +----------------------------------------------------------+
        |                    |                     |
  role: factual         role: not eligible     role: not eligible
  supplier              to modify C3           to direct action
        |                    |                     |
  retained as fact      recorded as claim,     recorded as request,
                        not as a revision      not as instruction
Figure 2: Source-role separation inside one message

Missing, conflicting, expired or unverifiable role information is recorded as unresolved. Roles may be established by cryptographic assertions or by non-cryptographic configuration; the diagnostic method is the same, and a role assertion's authenticity does not make its content causally conclusive.

6. Handoff Observation and Lineage

The lineage records, for each observable boundary, the representation of each governing condition. Four positions are distinguished: sender input, sender-generated output, submitted payload, receiver-observed input, plus the receiver's proposed operation. Streamed fragments are finalized into message versions; sender and receiver observations of the same handoff may differ and are both kept.

 Source S --> Agent A --> Converter --> Agent B --> proposal P
   |           |            |             |            |
  seg-1     out-A v1     payload v1    in-B v1     op-proposal
   \__________|____________|_____________|____________/
       condition C3 status recorded at each boundary:
       (sender input, sender output, submitted, received)
Figure 3: Multi-agent handoff lineage

Lineage is not message logging: it binds versions to sources and to conditions. Where a boundary cannot be observed it is marked unavailable; missing information is never treated as proof that no change occurred.

7. Counterfactual Intervention Model

The counterfactual method varies one designated aspect of an input while holding the rest fixed, and runs both variants through a non-effectuating interface. Required controls for every test: the originating condition and revision are unchanged; the source role context is the same; the factual domain is valid; replay capability exists; the execution is safe; and the pair is validated before its result is used.

            Contract C3 (max INR 2,000) fixed in all cells
                      Unsupported budget claim
                         absent        present
  Price      1,800    [ W1 ]         [ W2 ]
  fact       2,500    [ W3 ]         [ W4 ]

  W1 vs W3 : factual sensitivity (legitimate)
  W1 vs W2 : unauthorized-instruction sensitivity
  Factorial: all four cells; interaction term
  Each pair: one designated difference, invariants checked,
  run through a non-effectuating interface only.
Figure 4: Matched counterfactual design
Test A: legitimate factual influence

Vary a factual field (for example price 1,800 versus 2,500, including boundary values just below and above the limit) with no instruction-like claim. A compliant agent is expected to change its decision across the limit; this difference is legitimate factual sensitivity.

Test B: unauthorized instruction influence

Hold the fact fixed and vary only an instruction-like assertion from a source not eligible to modify the condition (present versus absent). A material difference indicates susceptibility to that statement under test conditions.

Test C: combined or factorial

Vary fact and claim together in a 2 by 2 design to estimate main effects and interaction; interpretation is limited where cells are incomparable.

Test D: message transformation

Substitute a reference message for the actual transformed message at a boundary and compare downstream behavior, to examine the contribution of the message representation.

Test E: receiver-side interpretation

Supply the receiver an explicit reference condition and examine whether its proposed operation respects it, to separate receiver failure from upstream change.

Why injected-statement sensitivity does not prove the historical cause: a test shows that, under the tested context, the agent's behavior depends on the statement. The original incident may have had a different cause, several causes, a model or context that has since changed, or a statement that was never present. The finding is therefore classified as tested influence, and historical attribution requires separate, qualified support (see the evidence status set).

8. Diagnostic Witness Selection

Testing every factual value and instruction variation is infeasible. A selector chooses witnesses within a budget. This section is illustrative; it does not claim production superiority over any existing agent-debugging approach.

   hypotheses H (incl. UNKNOWN) --+
                                   v
   +----> score feasible witnesses: gain / cost
   |               |
   |        pick best within budget & safety
   |               |
   |        run non-effectuating witness
   |               |
   |        validate observation, update H
   |               |
   +---- continue? <-- stop: threshold, budget, safety,
                            no utility, non-identifiable
                            --> finding OR abstention
Figure 5: Adaptive witness selection loop

Assumptions: a finite hypothesis set including an explicit unknown hypothesis; a witness catalog with declared costs and the hypotheses each can discriminate; an outcome model (deterministic mapping, or likelihoods that may be uncertain); and a non-effectuating execution capability.

select_and_run(H, W, budget):
  while true:
    feasible = [w for w in W if admissible(w) and cost(w) <= budget
                 and replay_ok(w) and safe(w)]
    if no feasible witness with positive utility: stop(NO_UTILITY)
    if remaining H observationally equivalent: stop(NONIDENT)
    w = argmax over feasible of utility(w, H) / cost(w)
    obs = run_non_effectuating(w)
    if not valid(obs): mark(w, INVALID); continue
    H = update(H, w, obs)        # Bayesian, or partition refinement
    budget -= cost(w)
    if decision_criterion_met(H): stop(SUPPORTED)
  report finding or ABSTAIN
Deterministic variant

Hypotheses partitioned by witness outcome; select the witness that splits the partition best per cost; no probabilities required.

Stochastic variant

Expected information gain from a likelihood model per unit cost; may use robust selection when likelihoods are uncertain or a decision-loss criterion.

Stopping

Supported decision threshold, low remaining decision risk, no positive utility, budget exhausted, safety constraint, non-identifiability, or unresolved evidence.

The underlying filing describes a toy reference implementation (E08_TOY_REFERENCE_V1) using seven hypothesis states, five witnesses, manually specified likelihoods, and no remote model access. The filing itself states that its results do not establish empirical likelihood calibration, superiority over an existing agent-debugging system, production-quality replay isolation, resistance to a malicious evidence recorder, universal prompt-injection detection, complete multi-fault coverage, or guaranteed identification of a hidden historical cause. This document makes the same limits its own.

9. Behavioral Comparison

A response difference is not necessarily a change of interpretation, and model-based agents are stochastic. The comparator therefore reports an effect, an interval or other uncertainty representation, a status, the coverage of observations and the assumptions, per condition, handoff, witness and version context.

 paired observations --> normalize --> effect estimate d
                                        with interval [L, U]
   predeclared materiality threshold delta (set before data)

   L > delta  ........ material difference supported
   [L,U] in (-delta, delta]  practical equivalence supported
   otherwise ......... insufficient evidence
   pair invalid ...... invalid comparison (no conclusion)
Figure 6: Statistical comparison of interpretations

The materiality threshold is declared before inspecting results. Categories include supported material difference, supported practical equivalence, insufficient evidence and invalid comparison. Deterministic structured comparisons are used where outcomes are deterministic; fixed-sample or sequentially valid procedures where they are not. Adaptive test selection affects statistical validity and is accounted for.

10. Multi-Hop Localization

The localizer combines structural observations (where a condition's representation first differs) with intervention evidence (which transition responds to the tested change). It reports the earliest observable divergence boundary where boundary evidence suffices, a bounded interval where stages are unobserved, or an unresolved localization. A unique responsible agent is never mandatory.

 hop:      1        2        3        4
 C3:     [ok] --- [??] --- [bad] --- [bad]
          |        |         |
       observed  unobserved observed
 last good = hop 1; first bad = hop 3
 fault interval I = (hop 1, hop 3]   (hop 2 or 3 transition)
 No unique blame is reported when hop 2 cannot be observed.
Figure 7: Earliest observable divergence over four hops

Illustrative decision matrix (from the filing's localization embodiment, paraphrased): sender input preserves the condition but sender output changes it suggests a sender-side representation difference; the converter's submitted payload preserves it but the receiver input changes it suggests a conversion interval; a condition-complete received message with a violating proposal suggests a receiver action-condition failure; different compliant decisions under different facts indicate legitimate factual sensitivity; an unobserved intermediate stage yields a bounded interval; invalid or inconclusive tests yield an unresolved diagnosis. The matrix does not establish a unique historical cause by itself.

10.1. Evidence-Bounded Causal Localization and Diagnostic Uncertainty

Evidence tiers in the filing range from a candidate discrepancy, through validated structural divergence and controlled intervention effect, to stage-localized controlled effect and, last, historically supported contribution under explicit assumptions. A higher-numbered tier is not implied by a lower one: a controlled effect does not by itself imply historical causation.

  • Abstain when tests are invalid, incomparable, statistically inconclusive, unsafe or unavailable.
  • Report a bounded interval, not a single agent, when intermediate transformations cannot be observed.
  • Report multiple faults, or masked faults, as possible; an earlier fault may be corrected or hidden by a later transformation.
  • Treat the hidden internal reasoning of an agent as unobservable from an external interface.
  • Do not convert a threshold crossing into scientific confidence; model validity is separate.

11. Branching and Merging Workflows

Parallel branches and merges can hide, create or combine faults. The diagnosis classifies whether each condition applies to each branch, validates a consistent merge context, substitutes branch outputs, uses factorial branch interventions, and reports branch faults, merge faults, interactions and masked faults, with intervals where the branch-to-merge boundary is unobserved.

            +--> Branch X (ok) -----+
 Task --> Split                      +--> Merge --> proposal
            +--> Branch Y (claim) --+
 Tests: substitute X out, Y out, both (factorial);
 compare merge output. Separate: branch fault / merge fault /
 interaction / masked fault.
Figure 8: Parallel branches and merge

12. Protocol-Independent Adapters

Adapters sit beside existing agent protocols, SDKs, gateways, tool servers and legacy interfaces. They normalize native events into canonical events, publish what they can observe and replay (a capability and coverage manifest), translate witnesses to native interfaces, and run diagnostic inputs non-effectuatingly. No change to A2A, MCP or any protocol is required, and no new mandatory message format is introduced.

 A2A    MCP    SDK     Gateway   Legacy API   Future protocol
  |      |      |         |          |             |
 [adapter][adapter][adapter][adapter][adapter]  [adapter]
        \       \     |      /        /          /
         canonical event + capability + coverage records
                        |
              diagnostic processor (E..I)
 Protocols unchanged; adapters sit beside them.
Figure 9: Protocol-independent adapters

Legacy integration: for a legacy service or API (variation E12-V16 in the filing), the adapter retains the originating task, captures outgoing requests and incoming responses, associates conditions with request fields, classifies source responses and builds a request-response graph. Where only metadata is visible, coverage is marked partial; hidden reasoning cannot be reconstructed from an external interface.

13. Diagnostic Evidence Structures

An optional, transport-neutral evidence model has six record types. The example fields below are illustrative only; this document defines no mandatory message format, wire encoding or registry.

TaskContractRecord
  { task_id, revision, conditions[typed], unresolved[] }
HandoffObservationRecord
  { task_id, hop, boundary, message_version,
    condition_status{}, coverage }
DiagnosticWitnessRecord
  { witness_id, pair, invariants, designated_diff,
    admissibility, executable }
DiagnosticObservationRecord
  { witness_id, run, normalized_response, validity,
    context_version }
DiagnosticFindingRecord
  { claim, tier_or_status, interval, coverage,
    assumptions, unresolved }
EvidenceContinuityRecord
  { prior_ref, supersedes, applicability, redactions,
    integrity (optional) }
Figure 10: Illustrative evidence records (non-normative)
 observations --> evidence records --> claim support check
                                          |
   per-claim status: observed / controlled effect /
   bounded interval / historical (qualified) / unresolved
                                          |
   selective disclosure (redaction) --> export package
   optional authentication and integrity; none implies
   that the causal conclusion is conclusive.
Figure 11: Evidence export

Evidence may be authenticated, integrity-protected and selectively disclosed, but none of these implies that a causal conclusion is conclusive. Statuses describe different properties rather than a single ascending scale: an authenticated record may contain a contradicted technical claim; a controlled test may support an effect without locating a hop; a historical observation may remain valid after its current applicability has expired.

14. Selective Revalidation

Facts, roles and models change between observation and use (a semantic time-of-check to time-of-use gap). Each finding carries the set of facts, roles, contract revisions and model versions it depends on. When one changes, only affected findings are marked as requiring revalidation and only affected witnesses are rerun; unaffected findings remain current; findings that cannot be replayed become historical only.

 fact v1 --- message --- proposal --- ... --- use time
   |                                             |
 fact v2 appears here ---------------------------+
 dependency set of each finding checked:
   unaffected -> remains CURRENT
   affected   -> REVALIDATION_REQUIRED (rerun only those)
   no replay  -> HISTORICAL_ONLY
Figure 12: Selective TOCTOU revalidation

15. Correction and Regression Testing

An optional processor proposes corrections at the site the finding implicates (for example the converter rather than the receiver), validates them counterfactually, and runs regression witnesses: the original failure, boundary values, metamorphic relations and preservation of legitimate positive behavior. Outcomes may be partial, recurred or new regression, and are limited to the contexts tested.

 finding --> candidate correction --> counterfactual check
                                          |
        regression witnesses: original failure, boundary,
        metamorphic, positive-behavior preservation
                                          |
   outcomes: validated for tested context / partial /
             fault recurred / new regression / incomparable
 (A validated correction is evidence; deployment is separate.)
Figure 13: Correction and regression verification

A validated correction is diagnostic evidence about a repair. Whether a repair is deployed, and whether any operation proceeds, is decided elsewhere.

16. Trust and Threat Model

The model lists 24 threats. For each: attacker capability, trust boundary, observable effect, detection method, residual uncertainty and safeguards. Evidence may be authenticated and selectively disclosed without implying causal conclusiveness.

T1. Instruction-like content in a factual field

Attacker capability: supplier or tool output embeds a modification claim.

Trust boundary: source to sender.

Observable effect: sender condition changes.

Detection method: matched instruction-source test (Test B).

Residual uncertainty: model may change; test context may not match the incident.

Safeguards: condition-scoped roles; non-effectuating tests.

T2. Authority laundering by an intermediary

Attacker capability: trusted intermediary restates a claim as its own.

Trust boundary: intermediary.

Observable effect: derived message gains eligibility.

Detection method: role carry-through check; lineage.

Residual uncertainty: intermediary hidden.

Safeguards: eligibility never increases by derivation.

T3. Quotation that becomes instruction

Attacker capability: quoted text is treated as governing.

Trust boundary: sender to receiver.

Observable effect: receiver adopts claim.

Detection method: message-replacement test (Test D).

Residual uncertainty: interpretation ambiguity.

Safeguards: segment-level labeling.

T4. Persistent-memory authority laundering

Attacker capability: stored note replays an old claim.

Trust boundary: memory to agent.

Observable effect: claim reappears as condition.

Detection method: taint labels; temporal tests.

Residual uncertainty: memory not observable.

Safeguards: versioned memory records.

T5. Multilingual transformation

Attacker capability: translation alters a limit.

Trust boundary: converter.

Observable effect: condition changes meaning.

Detection method: structural projection per language.

Residual uncertainty: translation ambiguity.

Safeguards: typed values and units.

T6. Summarization omission

Attacker capability: summary drops a condition.

Trust boundary: sender output.

Observable effect: condition absent.

Detection method: lineage divergence.

Residual uncertainty: summary unobserved.

Safeguards: condition presence checks at each boundary.

T7. Converter field rewrite

Attacker capability: payload converter alters a field.

Trust boundary: submitted to received.

Observable effect: argument differs.

Detection method: submitted-versus-received comparison.

Residual uncertainty: converter hidden.

Safeguards: adapter at converter boundary.

T8. Receiver misinterpretation

Attacker capability: receiver restates rule but acts otherwise.

Trust boundary: receiver.

Observable effect: non-compliant proposal.

Detection method: reference-condition test (Test E).

Residual uncertainty: black-box receiver.

Safeguards: correspondence checks.

T9. Stale factual input

Attacker capability: old fact used after update.

Trust boundary: fact store to agent.

Observable effect: wrong decision.

Detection method: version-dependency check.

Residual uncertainty: clock or version skew.

Safeguards: bitemporal records.

T10. Evidence recorder tampering

Attacker capability: attacker edits observation records.

Trust boundary: recorder.

Observable effect: false finding.

Detection method: integrity checks where available.

Residual uncertainty: malicious recorder not excluded.

Safeguards: signed or chained records (optional).

T11. Witness poisoning

Attacker capability: adversary shapes test inputs.

Trust boundary: witness generator.

Observable effect: test misleads.

Detection method: invariant validation; admissibility.

Residual uncertainty: novel manipulation.

Safeguards: predeclared invariants.

T12. Test-aware agent behavior

Attacker capability: agent behaves differently when tested.

Trust boundary: agent.

Observable effect: tests unrepresentative.

Detection method: varied and repeated witnesses.

Residual uncertainty: detection of test awareness limited.

Safeguards: randomization; coverage reporting.

T13. Model drift between incident and test

Attacker capability: model updated after incident.

Trust boundary: agent.

Observable effect: tests not comparable.

Detection method: version stratification.

Residual uncertainty: unknown versions.

Safeguards: incomparable status; abstention.

T14. Replay side effects

Attacker capability: diagnostic run causes real effects.

Trust boundary: adapter.

Observable effect: unintended action.

Detection method: non-effectuating execution controller.

Residual uncertainty: misconfigured isolation.

Safeguards: separate deployment; sandboxing; safety stop.

T15. Evidence confidentiality

Attacker capability: diagnostic export reveals sensitive content.

Trust boundary: exporter.

Observable effect: disclosure.

Detection method: redaction; minimization.

Residual uncertainty: residual inference.

Safeguards: selective disclosure; purpose limits.

T16. Adaptive-selection bias

Attacker capability: selector steers toward a favored cause.

Trust boundary: selector.

Observable effect: biased conclusion.

Detection method: statistical accounting for adaptivity.

Residual uncertainty: model misspecification.

Safeguards: predeclared thresholds; unknown hypothesis.

T17. Multiple concurrent faults

Attacker capability: several faults interact.

Trust boundary: path.

Observable effect: single-blame error.

Detection method: factorial tests; multi-fault assessment.

Residual uncertainty: combinatorial limits.

Safeguards: no mandatory unique blame.

T18. Masked upstream fault

Attacker capability: later step hides earlier fault.

Trust boundary: path.

Observable effect: earlier fault missed.

Detection method: restoration tests.

Residual uncertainty: unobserved stages.

Safeguards: intervals; abstention.

T19. Parallel-branch interaction

Attacker capability: branches jointly cause merge fault.

Trust boundary: merge.

Observable effect: wrong attribution.

Detection method: branch factorial design.

Residual uncertainty: scale.

Safeguards: interaction reporting.

T20. Missing observation boundary

Attacker capability: an interface is not exposed.

Trust boundary: adapter.

Observable effect: no evidence.

Detection method: coverage vector.

Residual uncertainty: irreducible.

Safeguards: explicit unavailable status.

T21. Identity spoofing of source

Attacker capability: source impersonated.

Trust boundary: source.

Observable effect: wrong role.

Detection method: authentication outside this document.

Residual uncertainty: depends on identity layer.

Safeguards: separate authentication; role not inferred from content.

T22. Replay or reorder of evidence

Attacker capability: old evidence resubmitted.

Trust boundary: exporter.

Observable effect: stale finding accepted.

Detection method: continuity record; applicability class.

Residual uncertainty: depends on freshness source.

Safeguards: supersession records.

T23. Over-trust in an authenticated record

Attacker capability: record signed yet contradicted.

Trust boundary: consumer.

Observable effect: false assurance.

Detection method: claim-level status.

Residual uncertainty: human misreading.

Safeguards: status separate from authentication.

T24. Cost exhaustion

Attacker capability: attacker forces many tests.

Trust boundary: selector.

Observable effect: denial of diagnosis.

Detection method: budget and stopping rules.

Residual uncertainty: degraded coverage.

Safeguards: budget caps; mode selection.

17. Security Considerations

This architecture reads message content and runs agents on altered inputs. Diagnostic execution MUST be non-effectuating where the architecture is claimed to be used as described; an implementation that cannot guarantee that SHOULD NOT execute the witness and SHOULD retain the structural definition only. Evidence integrity mechanisms are optional and do not make findings conclusive. Diagnostic findings MUST NOT be presented as authorization decisions.

Security considerations in general follow [RFC3552]. Authentication, authorization and execution control are out of scope and remain with other systems.

18. Privacy Considerations

Diagnostic records may contain instruction text, facts, personal data and trade secrets. Implementations SHOULD minimize retained content, prefer condition-level projections to raw content, support redaction and purpose limitation, and consider [RFC6973]. Selective disclosure reduces evidential support and the reduction should be visible in the record. Nothing here is a statement of compliance with any privacy law.

19. Interoperability Considerations

The architecture requires no change to A2A, MCP or other protocols. Interoperability is at the level of optional diagnostic findings and the evidence model; organizations may keep their own authority policies while exchanging compatible findings. Heterogeneous, asynchronous, streaming and partially observable agents are expected; coverage must be recorded rather than assumed.

20. Performance and Evaluation

Three operating modes are distinguished. No universal latency claim is made.

Table 1: Operating modes
Mode What it does Expected role
A. Lightweight observation Captures task identifiers, message versions, dependency links and deterministic structured-condition checks Always on candidate; cheap; not equivalent to behavioral testing
B. Event-triggered diagnosis Runs witnesses and comparison when a discrepancy or lower-trust influence is suspected On demand within a budget
C. Offline or post-incident Full counterfactual, localization and regression analysis Background or incident review

Design objectives are not measurements. The filing lists illustrative objectives only (for example low milliseconds for tagging and structured comparison; model- and network-dependent time for counterfactual invocation). A credible implementation would report p50, p95, p99 and p99.9 latency, model invocation counts, diagnostic cost and the rate of unresolved findings, and distinguish objectives from benchmark results. The diagnostic path is separate from any fast path in which existing authorization and business rules decide whether an operation runs; a full causal test inside every transaction or protection loop is not assumed.

21. IANA Considerations

This document has no IANA actions.

22. Relationship to Existing Work

This section is a non-normative comparison by the author and makes no novelty conclusion. Semantic delegation, signed handoffs and observability are existing areas and are not claimed as new here; the distinct focus is counterfactual, uncertainty-aware, condition-scoped diagnosis. Reading was limited as stated in the source verification appendix.

Agentic AI Use Cases and Requirements [I-D.agentic-ai-usecases-requirements]

Covers: use cases for agentic AI communication systems and derived protocol requirements; later revisions exist (only search excerpts were read).

Relation: provides use-case context; this document defines no protocol requirements and is a diagnostic complement.

Signed Handoff Context for Agentic Dialogs [I-D.srivastava-agentproto-handoff-context]

Covers: signed offer, signed status and hash-linked chain; it says a digest does not say what an object means, states that semantic preservation is not addressed, and grants no authority.

Relation: complementary. A signed chain shows order and consistency; this document asks, by tests, whether a condition's meaning survived.

Semantic Governance Requirements for AI Agent Delegation and Handoff [I-D.luna-eagf-agent-delegation-semantics]

Covers: requirements REQ-1 to REQ-8 including bounded authority scope and material context preservation; no new protocol; does not define detection or counterfactual diagnosis.

Relation: complementary. It says what must be preserved; this document offers a way to diagnose whether a condition was preserved.

A2A and MCP

Covers: agent-to-agent and model-to-tool interaction protocols (details not re-verified here).

Relation: adapters observe them without changing them.

OpenTelemetry and distributed tracing

Covers: telemetry and span correlation.

Relation: supplies ordering; does not test condition influence.

Provenance, evaluation, prompt-injection and observability work

Covers: records, benchmarks, detectors and monitoring.

Relation: different function; this adds matched counterfactual tests separating fact from instruction by source role.

The same author's separate execution-finality documents concern authority over effects. This document intentionally keeps causal diagnosis separate from execution finality and does not rely on them.

23. Non-Normative Industry Matrix

Non-normative and hypothetical. This mapping is AI-generated and was NOT verified against vendors' official documentation in preparing this document; product names and capabilities may be wrong or out of date. It must be cross-checked against official documentation and the original filings before any conclusion is drawn. It is not an endorsement, validation, adoption, infringement or deficiency statement, and no company has reviewed it.

Google

1. Existing technology (to be verified): A2A, Agent Development Kit, Gemini Enterprise Agent Platform.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): workflow-event interface or A2A client/server middleware.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Microsoft

1. Existing technology (to be verified): Agent Framework workflow, handoff and middleware layers.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): middleware and workflow event hooks.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

OpenAI

1. Existing technology (to be verified): agent handoff callbacks, input filters, tool wrappers, tracing processors.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): handoff input-transformation hooks and tool wrappers.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Anthropic

1. Existing technology (to be verified): tool-result processing, subagent delegation, MCP interfaces.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): tool-result normalization and approved MCP interfaces.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Amazon

1. Existing technology (to be verified): cloud agent runtime and tool gateway services.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): gateway or runtime event stream.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Salesforce

1. Existing technology (to be verified): enterprise agent platform actions and data connectors.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): connector and action-request layer.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

IBM

1. Existing technology (to be verified): enterprise orchestration and governance tooling.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): orchestration event interface.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

NVIDIA

1. Existing technology (to be verified): agent toolkits and inference microservices.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): pipeline and microservice boundaries.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Mistral

1. Existing technology (to be verified): hosted agents and function-calling APIs.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): API request/response boundary.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Aleph Alpha

1. Existing technology (to be verified): sovereign enterprise AI deployments.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): private deployment boundary.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

SAP

1. Existing technology (to be verified): business-process AI agents inside enterprise applications.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): business-process event layer.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Siemens

1. Existing technology (to be verified): industrial AI and engineering assistants.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): supervisory or planning interface.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Ericsson

1. Existing technology (to be verified): telecom network AI and automation.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): network-management agent interface.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Nokia

1. Existing technology (to be verified): telecom and network automation agents.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): management-plane agent interface.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Alibaba

1. Existing technology (to be verified): cloud agent platforms and model services.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): service gateway.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

Tencent

1. Existing technology (to be verified): cloud agent platforms and messaging-integrated assistants.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): platform middleware.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

DeepSeek, Doubao, Kimi and GLM environments

1. Existing technology (to be verified): model-API based agent frameworks.

2. Present limitation addressed: valid transport, task state and signatures do not show that each governing condition survived each handoff.

3. Where an adapter could attach (hypothetical): API client wrapper.

4. What would be observed: originating instruction, transformed message versions, source segments and tool results, proposed operation fields.

5. What would be tested: matched fact and instruction-claim variants in a non-effectuating copy; the sender output, converter payload and receiver input compared per condition.

6. What would be reported: condition-preservation status per boundary, tested influence with interval, earliest observable divergence or bounded interval, and abstention where unsupported.

7. What would not be done: nothing is approved, rejected or blocked by the adapter; the platform's own authorization, guardrails and approvals keep deciding.

8. Residual limits: hidden reasoning is not reconstructable; black-box or metadata-only components yield partial coverage; replay may be unavailable.

24. Trustworthy Multi-Agent AI, European Digital Sovereignty, and Diagnostic Accountability

Non-normative. Nothing here is a claim that any implementation is legally compliant because it produces diagnostic evidence, and no geopolitical intent or wrongdoing is attributed to any country or vendor.

25. Worked Examples

Example A: financial risk limit

A treasury agent chain has an approved exposure limit. A market feed's commentary says the limit is raised; an intermediate agent treats it as a revision. The diagnostic varies the market facts and the unsupported statement separately and compares proposed adjustments; it may report tested influence and the divergence boundary. Non-effectuating operations: simulated order proposals only. The treasury authorization system decides whether any real trade happens.

Example B: cross-provider A2A handoff

Agent X on one platform delegates to agent Y on another over A2A. Y proposes a purchase above X's limit. An adapter records the submitted and received messages; if the limit was present at submission and absent at receipt, a conversion interval is reported; if present at receipt, a receiver test is run. Non-effectuating operations: replayed proposals to a stub. Authorization remains with each party.

Example C: MCP tool-output prompt injection

A retrieval tool returns a price and the sentence that the budget was raised. Only the sentence is varied. A material difference supports tested sensitivity; it does not prove that this sentence caused the earlier incident. Non-effectuating operations: replay without calling the real booking tool.

Example D: parallel agents merge

Two research branches are each correct, the merge promotes a claim from one. Branch substitution and a factorial design separate branch from merge from interaction faults. Non-effectuating operations: merge replay in isolation.

Example E: banking legacy adapter

A legacy payment API is fronted by an agent. The adapter captures request fields and responses and associates the limit with the amount field; it reports a request-field divergence. It cannot see inside the legacy system; coverage is partial. Non-effectuating operations: requests to a test environment only.

Example F: European industrial maintenance waiver

A maintenance planning agent receives a supplier note that a safety interval waiver is approved. The note is not from an eligible role. Tests vary the note; findings go to engineers. Non-effectuating operations: planning simulation only; certified safety systems and humans decide everything physical.

26. Validation Plan

No experimental results are reported in this document. The following is a plan. The underlying filing reports only toy reference implementations with synthetic harnesses (E08 described in detail; the filing indicates similar runnable references for later embodiments, not examined here), which demonstrate that the described control flow can run as software and do not establish real-world performance.

27. IPR Disclosure, Attribution and Notices

Source filing: Indian Provisional Application 202631134241, dated 8 November 2026 [IN-134241] (as reported by the applicant; not a certified registry record). Foundational reference: PCT/IB2026/055615 [PCT-055615], filed 4 June 2026, published 16 July 2026 as WO 2026/150382 A2. A separate Indian application, 202631131650 (3 November 2026, evaluator governance), is a distinct subject and is kept distinct from this document.

Public cross-references: https://doi.org/10.5281/zenodo.24076231 [ZENODO-CAUSAL]; Technical Disclosure Commons records [TDC-12934] and [TDC-12973] and [TDC-12542].

This document contains no licensing commitment and no statement about royalties or essentiality. Contributors are bound by BCP 78 and BCP 79 ([RFC5378], [RFC8179]). Patents or applications mentioned here are not a substitute for formal IETF IPR disclosure. The author will make the IPR disclosure for this document at the appropriate time, through the official IETF process and in accordance with IETF guidelines. The IETF IPR database must be checked independently.

Source integrity: the filing text supplied is a 1,657-page PDF; the editor read the front disclosure, section headings of all embodiments, and selected sections in detail, not every page. The structured JSON technical disclosure was not available. Claims were not reconstructed. Verify against the official filed copy.

27.1. Technical Source and AI-Assistance Notice

The underlying causal semantic handoff invention and associated patent specification are attributed to human inventor Sangam Kumar Das. The present Internet-Draft explanatory text, keyword index, flowchart descriptions, research comparisons and industry relevance mapping may have been prepared with AI assistance. Such material can contain omissions or errors and must be verified against the original filed patent documents, public technical sources, and applicable IETF specifications.

27.2. Non-Endorsement Notice

This document is an individual technical proposal. It does not establish IETF standardization, European Union regulatory approval, certification, sponsorship, adoption, implementation, commercial affiliation or endorsement by any company, government, standards organization or research institution. Company names and product references identify potential research or integration environments only. No copying, infringement, technological deficiency or unauthorized use is alleged.

27.3. Technical Scope Notice

The diagnostic architecture supplies evidence about observable instruction interpretation changes. It does not independently authorize payments, execute network configuration changes, approve industrial operations, replace certified safety systems or guarantee that a historical cause has been uniquely identified.

The filing dates and public-disclosure references are provided for traceability. The patentability, priority entitlement and legal scope of individual technical features require assessment against the official filed documents and applicable law.

28. Normative References

[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", RFC 2119, DOI 10.17487/RFC2119, BCP 14, , <https://www.rfc-editor.org/info/rfc2119>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", RFC 8174, DOI 10.17487/RFC8174, BCP 14, , <https://www.rfc-editor.org/info/rfc8174>.

29. Informative References

[EU-AIA]
European Union, "Regulation (EU) 2024/1689 (Artificial Intelligence Act)", , <https://eur-lex.europa.eu/eli/reg/2024/1689/oj>.
[EU-NIS2]
European Union, "Directive (EU) 2022/2555 (NIS2)", , <https://eur-lex.europa.eu/eli/dir/2022/2555/oj>.
[I-D.agentic-ai-usecases-requirements]
Reddy, T., Sarker, Z., and K. Yao, "Agentic AI Use Cases and Requirements", Work in Progress, Internet-Draft, draft-agentic-ai-usecases-requirements, , <https://datatracker.ietf.org/doc/draft-agentic-ai-usecases-requirements/>.
[I-D.luna-eagf-agent-delegation-semantics]
Luna, D., "Semantic Governance Requirements for AI Agent Delegation and Handoff", Work in Progress, Internet-Draft, draft-luna-eagf-agent-delegation-semantics-00, , <https://datatracker.ietf.org/doc/draft-luna-eagf-agent-delegation-semantics/>.
[I-D.srivastava-agentproto-handoff-context]
Srivastava, K., "Signed Handoff Context for Agentic Dialogs", Work in Progress, Internet-Draft, draft-srivastava-agentproto-handoff-context-00, , <https://datatracker.ietf.org/doc/draft-srivastava-agentproto-handoff-context/>.
[IN-134241]
Das, S., "Indian Provisional Patent Application 202631134241", As reported by the applicant. Not a certified registry record., , <https://ipindia.gov.in/>.
[PCT-055615]
Das, S., "PCT/IB2026/055615 (WO 2026/150382 A2)", Filed 4 June 2026; published 16 July 2026. Verify on PATENTSCOPE., , <https://patentscope.wipo.int/>.
[RFC3552]
Rescorla, E. and B. Korver, "Guidelines for Writing RFC Text on Security Considerations", RFC 3552, DOI 10.17487/RFC3552, BCP 72, , <https://www.rfc-editor.org/info/rfc3552>.
[RFC5378]
Bradner, S. and J. Contreras, "Rights Contributors Provide to the IETF Trust", RFC 5378, DOI 10.17487/RFC5378, BCP 78, , <https://www.rfc-editor.org/info/rfc5378>.
[RFC6973]
Cooper, A., Tschofenig, H., Aboba, B., Peterson, J., Morris, J., Hansen, M., and R. Smith, "Privacy Considerations for Internet Protocols", RFC 6973, DOI 10.17487/RFC6973, , <https://www.rfc-editor.org/info/rfc6973>.
[RFC8179]
Bradner, S. and J. Contreras, "Intellectual Property Rights in IETF Technology", RFC 8179, DOI 10.17487/RFC8179, BCP 79, , <https://www.rfc-editor.org/info/rfc8179>.
[TDC-12542]
Das, S., "Mothership PCT disclosure record (title not verified)", Technical Disclosure Commons. Not fetched; verify., , <https://www.tdcommons.org/dpubs_series/12542/>.
[TDC-12934]
Das, S., "When AI Agents Change Instructions... (title truncated as supplied)", Technical Disclosure Commons. Not fetched; verify., , <https://www.tdcommons.org/dpubs_series/12934/>.
[TDC-12973]
Das, S., "Technical Disclosure Commons record 12973 (title not verified)", Technical Disclosure Commons. Not fetched; verify., , <https://www.tdcommons.org/dpubs_series/12973/>.
[ZENODO-CAUSAL]
Das, S., "Trustworthy Multi-Agent AI for Europe: Causal Detection of Instruction Drift, Prompt-Injection Influence, and Communication-Hop Failures Through Counterfactual Diagnostics", Zenodo, version v1. Relevance: public research description of the same causal diagnostic subject matter. Not fetched by the editor; verify., DOI 10.5281/zenodo.24076231, , <https://doi.org/10.5281/zenodo.24076231>.

Appendix A. Appendix A: All 18 Embodiments

The source filing is summarized for each of its 18 embodiments in the author's words, based on the filing's own text; the filing should be consulted for exact wording. E12-V16 (legacy service and API-based task automation) is a deployment variation under E12 and not a nineteenth embodiment. None of the embodiments grants or vetoes execution authority.

A.1. E01: Typed Originating Task-Contract Construction

1. Problem addressed

An originating instruction is free text; later agents summarize, translate or split it, and conditions (limits, prerequisites, deadlines, prohibitions) are lost or altered. Without a machine-testable reference there is nothing precise to compare later representations against.

2. Inputs and data structures

Attributed task instruction and revision; message and source identifiers; optional structured fields, voice transcript or document-derived text. Data structure: versioned Originating Task Contract of typed predicates (operator, value, unit, applicable operation field) plus unresolved-condition records.

3. Processing steps

Segment the instruction into normative conditions, factual references, proposed amendments, negations, temporal constraints and unresolved expressions; normalize typed predicates; link predicates to factual dependencies; record ambiguity instead of guessing; store a versioned contract.

4. Outputs

A versioned typed task-contract record that distinguishes testable from unresolved conditions, usable as the reference for later diagnostic comparison.

5. Dependencies on other embodiments

Needs a source attribution for the instruction and a revision identifier. Feeds E02, E03, E05, E07 and E13.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: contract digest or signed record binding the contract to its revision. Non-cryptographic: versioned application records and process-local identifiers. The filing states that both are supported.

7. Failure conditions and limits

A condition with a missing material attribute (for example no currency) is rejected or marked unresolved; the contract does not establish complete semantic equivalence of arbitrary natural language.

8. Interoperability significance

Gives independent implementations a common, protocol-neutral reference for what the task required, without mandating a message format.

9. Concrete example

Transport booking only after goods reach a destination, before a deadline, and within INR 2,000: three conditions with different operators are held as separate typed predicates.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Unlike logging, tracing or a message hash, it captures what each condition means in a testable form. It is not authentication, not authorization and not an evaluation of the task.

A.2. E02: Condition-Scoped Source-Role Classification and Information Separation

1. Problem addressed

Not every source may modify every condition. A supplier may report a price but not raise the buyer's budget, yet a receiving agent may treat an embedded statement as governing, especially when quoted, summarized or relayed by a trusted intermediary.

2. Inputs and data structures

Information segments with source identity; the E01 contract; a source-role model. Data structure: per-condition, per-source role records (factual-information eligible, modification eligible, unresolved, expired).

3. Processing steps

Identify source per segment; separate authentication from modification eligibility; classify each segment as fact, proposal, claim of authorization or instruction-like request; preserve role restrictions through quoting, translation and summarization; record missing or conflicting role information.

4. Outputs

Segment-level source-role classification per condition, with modification claims kept separate from accepted changes.

5. Dependencies on other embodiments

Depends on E01 conditions; independently established role information. Feeds E03, E07 and E14.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed role assertions. Non-cryptographic: configuration tables or application records. A source is not trusted merely because it is authenticated.

7. Failure conditions and limits

Role data may be missing, conflicting, expired or unverifiable; then the segment is marked unresolved rather than assumed eligible or ineligible.

8. Interoperability significance

Lets different platforms exchange the fact that a derived message must not gain more modification eligibility than its source had.

9. Concrete example

Message: 'The price is INR 2,500. The customer approved a budget of INR 5,000. Proceed.' Sentence 1 is a legitimate quote; sentences 2 and 3 are an unauthorized claim and an instruction-like request.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Rejecting the whole message discards useful facts; accepting it all promotes an instruction. Source identity alone does not separate the segments. It is not an authorization decision.

A.3. E03: Versioned Source-Condition Dependency Graph

1. Problem addressed

Task information passes through documents, tool responses, summaries and structured proposals. Chronological event traces do not say which condition or fact each transformation affected.

2. Inputs and data structures

E01 conditions, E02 roles, factual observations, message representations. Data structure: versioned graph linking conditions, facts, source segments, claimed amendments and derived messages.

3. Processing steps

Build nodes for conditions, facts, segments and messages; add edges for derivation, quotation, summary and dependency; keep contract revisions and fact versions separate; update on authorized amendment.

4. Outputs

A dependency graph usable to select tests and to separate factual changes from instruction changes.

5. Dependencies on other embodiments

Depends on E01 and E02. Feeds E07, E08, E10 and E15.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: hash-linked node versions. Non-cryptographic: versioned records with revision counters.

7. Failure conditions and limits

Partial observability leaves edges unknown; unknown edges are retained as such.

8. Interoperability significance

Common vocabulary of dependency edges across heterogeneous systems.

9. Concrete example

A supplier quotes INR 2,500 and claims a INR 5,000 budget; an agent relays 'limit INR 5,000'. The graph shows the 5,000 figure traces to the supplier claim, not to a contract revision.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Storing whole messages or traces does not tell a later process which condition a change touched; the graph does.

A.4. E04: Version-Bound Agent Handoff Observation and Message-Transformation Lineage

1. Problem addressed

A successful transmission does not show the receiver interpreted the message according to the originating task, and a log does not show which earlier segments contributed to a message.

2. Inputs and data structures

Normalized communication events with message versions, stream fragments, sender and receiver records. Data structure: canonical handoff interoperability record.

3. Processing steps

Ingest events; finalize streamed messages into versions; correlate sender and receiver observations; project each representation onto the governing conditions; export a common record.

4. Outputs

Version-bound semantic handoff lineage with condition-level representation status.

5. Dependencies on other embodiments

Depends on E03; adapters from E12. Feeds E05 to E10.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: chained or signed records. Non-cryptographic: versioned store.

7. Failure conditions and limits

Sender and receiver may observe different versions (a worked case in the filing); late streamed fragments change a finished message.

8. Interoperability significance

Lineage can be exported without changing the agents' own protocols.

9. Concrete example

Original 'within INR 2,000' becomes 'budget limit not stated' in a summary; the lineage shows where the condition representation changed.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Message logging records that a message existed; lineage binds versions to sources and conditions.

A.5. E05: Condition-Scoped Instruction Fingerprinting and Behavioral Evidence Association

1. Problem addressed

A message digest identifies bytes, not operational meaning; an agent may restate a rule correctly yet act differently under certain facts.

2. Inputs and data structures

Typed conditions; message representations; observed behavior. Data structure: multi-level fingerprint (textual, condition-projected, behavioral).

3. Processing steps

Canonicalize typed conditions; project representations onto a condition; compare at three levels; associate behavioral evidence with source context; invalidate records when versions change.

4. Outputs

Condition-scoped fingerprints and comparison results at textual, structural and behavioral levels.

5. Dependencies on other embodiments

Depends on E01, E04. Feeds E06 and E09.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: digests of canonical forms. Non-cryptographic: structural keys.

7. Failure conditions and limits

Equal fingerprints do not prove equal behavior; different wording may map to the same condition.

8. Interoperability significance

Compact identifiers for condition state that systems may compare without sharing full content.

9. Concrete example

Two differently worded instructions with the same INR 2,000 limit match structurally; an agent that states the limit but proposes INR 2,500 differs behaviorally.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Cryptographic equivalence and semantic equivalence are different properties; this embodiment keeps them separate.

A.6. E06: Version-Bound Diagnostic Observation Evidence

1. Problem addressed

Test results are only meaningful if the test context was reproducible and the pair was valid; otherwise a difference may come from a changed model or context.

2. Inputs and data structures

Witness manifests, non-effectuating observations, response records. Data structure: evidence manifest and observation package with state model.

3. Processing steps

Construct manifests; acquire observations non-effectuatingly; validate matched-pair invariants; normalize responses; validate packages; assess impact of version changes.

4. Outputs

Validated or incompatible observation packages with explicit state.

5. Dependencies on other embodiments

Depends on E05, E07. Feeds E08 to E10 and E18.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed observation records. Non-cryptographic: local digests (the filing notes a locally computed record digest in its toy implementation).

7. Failure conditions and limits

Incompatible pairs, changing model or persistent agent state, and stale facts make evidence incomparable.

8. Interoperability significance

A common observation record lets third parties judge whether two results are comparable.

9. Concrete example

A pair in which the factual price changed unintentionally is rejected before it can support any conclusion.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Ordinary test logs do not enforce matched-pair invariants or version context.

A.7. E07: Condition-Scoped Matched Counterfactual Witness Generation

1. Problem addressed

Observed correlation cannot separate legitimate factual sensitivity from susceptibility to an unauthorized instruction-like statement.

2. Inputs and data structures

Target condition, source segments, replay capability, dependency context. Data structure: matched witness pair with invariants and a designated difference.

3. Processing steps

Generate admissible fact variations (including boundary values), instruction-content interventions, message-representation substitutions and optional factorial sets; validate admissibility (invariants, task revision, fact domain, replay support, safety).

4. Outputs

Admissible diagnostic witnesses, possibly structurally valid yet not executable.

5. Dependencies on other embodiments

Depends on E01 to E06; replay support from E12. Feeds E08, E09.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed witness definitions. Non-cryptographic: versioned witness records.

7. Failure conditions and limits

Where replay is unsupported or unsafe, the witness is retained but not executed.

8. Interoperability significance

Witness manifests can be translated per platform by adapters.

9. Concrete example

Vary the supplier price from 1,800 to 2,500 with the contract unchanged (fact test); vary only the unsupported budget claim with the price unchanged (instruction test).

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Not red-teaming or prompt-injection scanning: pairs differ in one designated respect and keep the originating condition fixed.

A.8. E08: Dependency-Aware Adaptive Diagnostic Witness Selection

1. Problem addressed

Exhaustive testing is infeasible, and the cheapest test is not always the most informative.

2. Inputs and data structures

Hypothesis set, witness catalog with costs, outcome model, budget. Data structure: hypothesis distribution or partition plus witness cost records.

3. Processing steps

Initialize hypotheses (including an unknown state); score candidate witnesses by expected information gain per cost (or alternatives such as robust, partition or decision-loss selection); execute; update; stop on threshold, budget, safety, non-identifiability or no utility.

4. Outputs

A sequence of executed witnesses and a hypothesis state with explicit stopping reason, possibly abstention.

5. Dependencies on other embodiments

Depends on E03, E07; feeds E09, E10.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed observation references. Non-cryptographic: local digests. The filing's toy reference uses a locally computed digest.

7. Failure conditions and limits

Manually specified likelihoods are not calibrated; the filing states its toy results do not establish calibration, superiority, resistance to malicious recorders or complete multi-fault coverage.

8. Interoperability significance

Selection can be implemented over any witness catalog an adapter publishes.

9. Concrete example

Seven hypothesis states (source influence, sender omission, link conversion, receiver failure, stale fact, source-and-receiver, unknown) with five witnesses in the toy reference.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Differs from fixed test suites and generic active learning in being tied to condition-scoped hypotheses and to abstention.

A.9. E09: Uncertainty-Aware Interpretation Comparison

1. Problem addressed

A response difference is not necessarily a change in interpretation, and agents are stochastic.

2. Inputs and data structures

Matched observations, outcome measures, predeclared materiality threshold. Data structure: effect, interval, status, coverage, assumptions tuple.

3. Processing steps

Normalize responses; build comparison units; apply deterministic, fixed-sample or sequentially valid comparison; classify as material difference, practical equivalence, insufficient evidence or invalid comparison; handle multiple outcomes and multiplicity.

4. Outputs

Condition-specific comparison result with interval and status.

5. Dependencies on other embodiments

Depends on E06 to E08; feeds E10.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: authenticated result records. Non-cryptographic: ordinary records.

7. Failure conditions and limits

Missing or ambiguous observations, stale context, and observational equivalence limit conclusions; the threshold must be set before inspecting results.

8. Interoperability significance

Result tuple is transport-neutral.

9. Concrete example

Agent approves INR 1,800 and rejects 2,500 (legitimate factual sensitivity) versus proposes 2,500 only when a budget claim is present (unauthorized-instruction sensitivity).

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Differs from plain A/B statistics by keeping four questions apart and by abstention categories.

A.10. E10: Condition-Scoped Multi-Hop Causal Divergence Localization

1. Problem addressed

A wrong final proposal may come from the source, sender, converter, receiver or a combination.

2. Inputs and data structures

Boundary-graph observations, E09 results. Data structure: communication graph with four-position boundary model.

3. Processing steps

Project governing conditions; find the earliest observable structural divergence; evaluate source influence, converter and receiver effects; form fault intervals when stages are unobserved; assess multiple faults.

4. Outputs

Earliest observable divergence, bounded fault interval, or unresolved localization, with evidence tier.

5. Dependencies on other embodiments

Depends on E04, E07 to E09.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed boundary records. Non-cryptographic: versioned records.

7. Failure conditions and limits

Earliest observable divergence is not necessarily the first underlying cause; an earlier fault may be masked; intervals are used when stages are hidden.

8. Interoperability significance

Works with whatever boundaries an adapter exposes.

9. Concrete example

Condition preserved at hop 1, absent at hop 3, hop 2 unobserved: fault interval (hop 1, hop 3].

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Not distributed tracing: it does not just order spans, it tests which transition changed the condition.

A.11. E11: Condition-Scoped Branching, Merging and Parallel-Agent Diagnosis

1. Problem addressed

Parallel branches and merges can hide, create or combine faults.

2. Inputs and data structures

Branching event graph; merge inputs and outputs; branch substitutions. Data structure: branch event graph with condition applicability.

3. Processing steps

Classify condition applicability per branch; validate consistent merge context; substitute branch outputs; run factorial interactions; compute a branch-divergence frontier; assess masking and multiple faults.

4. Outputs

Branch-versus-merge fault separation with interaction evidence and intervals.

5. Dependencies on other embodiments

Depends on E10; feeds E13, E16.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed branch records. Non-cryptographic: versioned records.

7. Failure conditions and limits

Observation gaps between branch and merge; asynchronous branch versions; masked faults.

8. Interoperability significance

Branch records can be exported in common form.

9. Concrete example

Two research agents each correct, merge step promotes an unauthorized budget claim from one.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Branch tracing shows structure; this embodiment tests which branch or merge changed the condition.

A.12. E12: Protocol-Independent Diagnostic Adapters (including variation E12-V16, legacy service and API automation)

1. Problem addressed

Agents use different protocols and runtimes; diagnostics designed for one cannot apply uniformly.

2. Inputs and data structures

Native events from agent protocols, SDKs, gateways, tool servers, legacy APIs. Data structure: canonical event record, adapter capability manifest, coverage vector.

3. Processing steps

Register adapters; inspect capabilities; capture and normalize events; classify boundaries; correlate across protocols; publish capabilities; translate witnesses; run non-effectuating execution; export events.

4. Outputs

Canonical events, coverage manifest, translated witnesses.

5. Dependencies on other embodiments

Depends on native interfaces; supports E04 and E07 to E11.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: authenticated event evidence. Non-cryptographic: application records.

7. Failure conditions and limits

Where only metadata is exposed, hidden reasoning cannot be reconstructed from an external interface; coverage is marked partial.

8. Interoperability significance

Requires no change to A2A, MCP or other protocols; adapters sit beside them. The filing lists twenty deployment variations V01 to V20; V16 (legacy service and API-based task automation) is a variation under E12, not a nineteenth embodiment.

9. Concrete example

Legacy API: capture request fields, associate with conditions, classify source responses, build a request-response graph.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Not a new protocol or tracer; a normalization layer for diagnostics.

A.13. E13: Condition-Scoped Instruction-Message-Action Correspondence Diagnosis

1. Problem addressed

An agent may state a rule correctly yet emit a non-compliant operation, or an application layer may alter an argument.

2. Inputs and data structures

Contract conditions, messages, plans, proposed operations, dispatch attempts. Data structure: instruction-message-action chain with typed correspondence relations.

3. Processing steps

Evaluate numerical, aggregate and temporal conditions against proposed operations; compare proposal with dispatch arguments; test message-to-action and source-to-action interventions; classify faults and evidence tiers.

4. Outputs

Correspondence finding between instruction, message and proposed operation.

5. Dependencies on other embodiments

Depends on E01, E07 to E12.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed operation proposals. Non-cryptographic: records.

7. Failure conditions and limits

Remote black-box action services limit evidence; correspondence does not authorize anything.

8. Interoperability significance

Operation records are exportable without requiring a given format.

9. Concrete example

Agent states limit INR 2,000 yet proposes total_price INR 2,500.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Output is a diagnosis; it does not approve or block the operation.

A.14. E14: Condition-Scoped Lower-Trust Information Influence and Instruction-Promotion Diagnosis

1. Problem addressed

Lower-trust content (tool outputs, documents, memory) may carry instruction-like statements that get promoted into governing conditions.

2. Inputs and data structures

Mixed source content; role model; provenance graph. Data structure: condition-scoped taint and provenance labels.

3. Processing steps

Partition mixed content; detect unauthorized modification claims; propagate labels conservatively; detect promotion; find first observable promotion frontier; run matched source and factual interventions; assess multi-hop influence.

4. Outputs

Promotion and influence findings with tiers, separating provenance from tested influence.

5. Dependencies on other embodiments

Depends on E02, E07, E09, E10.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed source segments. Non-cryptographic: records.

7. Failure conditions and limits

Provenance is not causal influence; persistent memory and multilingual transformation complicate attribution.

8. Interoperability significance

Applies to tool results from MCP-style servers, retrieval systems and agent messages without protocol change.

9. Concrete example

Tool result contains a price plus 'budget has been raised to 5,000'; only the claim is varied.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Not a prompt-injection filter: it tests whether the content changed the condition-relevant behavior.

A.15. E15: Condition-Scoped Temporal Validity, TOCTOU Diagnosis and Selective Revalidation

1. Problem addressed

Facts and roles change between observation and use; old evidence may no longer apply.

2. Inputs and data structures

Bitemporal events, dependency sets, version records. Data structure: dependency-impact graph and evidence applicability class.

3. Processing steps

Track condition and witness dependency sets; evaluate the selective revalidation predicate; classify evidence applicability; re-run only affected witnesses; localize temporal faults.

4. Outputs

Applicability classes (current, historical only, revalidation required, incomparable and others) and revalidation plans.

5. Dependencies on other embodiments

Depends on E03, E06 to E10.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: authenticated versions. Non-cryptographic: version counters.

7. Failure conditions and limits

Replay may be unavailable; concurrent updates; model changes.

8. Interoperability significance

Version fields are portable.

9. Concrete example

Supplier price updated after an agent's message; only tests depending on that price are redone.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Selective, dependency-driven revalidation instead of rerunning everything.

A.16. E16: Condition-Scoped Fault-Obligation, Correction and Repair Validation

1. Problem addressed

After localization, a repair at the wrong site may only treat a symptom.

2. Inputs and data structures

Fault findings, correction candidates, affected paths. Data structure: fault obligation model and repair coverage matrix.

3. Processing steps

Generate and select candidate corrections (including set-cover alternatives); validate counterfactually; check invariants; analyze residual faults; manage lifecycle.

4. Outputs

Tested correction status and residual-fault analysis.

5. Dependencies on other embodiments

Depends on E10, E11, E13, E14.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed correction manifests. Non-cryptographic: records.

7. Failure conditions and limits

Deployment status may be unknown; corrections can introduce regressions. The filing includes a runnable reference (not examined in detail by this editor).

8. Interoperability significance

Manifest exportable.

9. Concrete example

Correct a converter that drops the limit field rather than patching the receiver.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Optional component; correction is separate from authorization of anything.

A.17. E17: Condition-Scoped Regression Witness Generation and Correction Qualification

1. Problem addressed

A correction may fix one case and break others.

2. Inputs and data structures

Correction manifest, regression witness families. Data structure: regression coverage matrix.

3. Processing steps

Generate regression witnesses (original failure, boundary, metamorphic, positive-behavior); ablate corrections; evaluate recurrence across contexts; apply statistical qualification and held-out tests.

4. Outputs

Regression outcome classes such as fault recurred, new regression detected, validated for specified contexts.

5. Dependencies on other embodiments

Depends on E16.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: signed regression results. Non-cryptographic: records.

7. Failure conditions and limits

Test selection bias; coverage gaps; durability limited to tested contexts.

8. Interoperability significance

Common outcome classes.

9. Concrete example

Exact-budget and just-above-budget cases re-run after a repair.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

Verification of a diagnostic correction, not a deployment gate.

A.18. E18: Condition-Scoped Diagnostic Evidence Packaging, Claim Support and Selective Disclosure

1. Problem addressed

Evidence must be reusable without overstating it; authenticated records can still contain contradicted claims.

2. Inputs and data structures

Evidence records, claims, dependency closure. Data structure: evidence record model, claim model and evidence status set.

3. Processing steps

Type evidence; define claim categories and support requirements; compute dependency closure and sufficiency; build verification vector and export manifests; support redaction and supersession; replay.

4. Outputs

Evidence package with per-claim status (for example directly observed, controlled effect supported, bounded interval supported, historical causation qualified).

5. Dependencies on other embodiments

Depends on all prior embodiments.

6. Cryptographic and non-cryptographic alternatives

Cryptographic: integrity and authentication of packages. Non-cryptographic: versioned records.

7. Failure conditions and limits

Redaction or withholding reduces support; statuses are not a single ascending scale.

8. Interoperability significance

Selective disclosure across organizations.

9. Concrete example

Package states a bounded interval and a tested influence, but only qualified historical causation.

10. Distinction from tracing, guardrails, authentication, authorization and evaluation

It records and qualifies diagnostic evidence; it does not decide any operation.

Appendix B. Appendix B: Flowcharts

The twelve figures appear in the body: principal pipeline (<xref target="fig-main"/>), source roles (<xref target="fig-roles"/>), lineage (<xref target="fig-lineage"/>), matched design (<xref target="fig-matched"/>), witness selection (<xref target="fig-witness"/>), comparison (<xref target="fig-compare"/>), localization (<xref target="fig-local"/>), branching (<xref target="fig-branch"/>), adapters (<xref target="fig-adapter"/>), evidence records and export (<xref target="fig-records"/>, <xref target="fig-export"/>), revalidation (<xref target="fig-toctou"/>) and correction (<xref target="fig-correct"/>).

flowchart TD
  A[Typed task contract] --> B[Source-role model]
  B --> C[Dependency graph]
  C --> D[Handoff lineage]
  D --> E[Counterfactual processor]
  D --> F[Witness selector]
  D --> G[Behavioral comparator]
  E --> H[Fault localizer]
  F --> H
  G --> H
  H --> I[Evidence exporter]
  I --> J[Optional correction]
  I --> K[Diagnostic finding]
  K -.-> L[External authorization]

Appendix C. Appendix C: Diagnostic Examples

Worked examples A to F appear in <xref target="examples"/>. In each, the diagnostic output is evidence and the listed non-effectuating operations produce no real external effect.

Appendix D. Appendix D: Patent Examiner Keywords and Search Phrases

This appendix is a search aid for examiners and readers. It is not a statement about novelty, scope or relevance of any prior art. Embodiment references are to <xref target="app-emb"/>.

Table 2: Terms (169)
Term Embodiments
instruction drift E01,E05
semantic drift E04,E10
semantic handoff E04
causal diagnosis E07,E10
counterfactual testing E07
matched pair E06,E07
matched counterfactual E07
intervention pair E07
factual intervention E07
instruction-source intervention E07,E14
message-representation substitution E07
factorial intervention E07,E11,E14
boundary-value witness E07
diagnostic witness E07,E08
witness admissibility E07
non-effectuating execution E06,E12
non-effectuating test adapter E07,E12
typed task contract E01
originating task contract E01
governing condition E01
typed predicate E01
condition operator E01
unresolved condition E01
ambiguity record E01
task revision E01,E15
source role E02
source-role eligibility E02
modification eligibility E02
factual-information eligibility E02
authentication versus eligibility E02
segment-level provenance E02,E14
mixed source content E14
instruction-like assertion E02,E14
unauthorized instruction claim E14
authority laundering E14
instruction promotion E14
promotion frontier E14
taint label E14
provenance graph E14
conservative taint propagation E14
lower-trust source E14
tool output injection E14
prompt injection influence E14
persistent memory laundering E14
dependency graph E03
source-condition dependency E03
factual dependency E03
condition scope E11
handoff lineage E04
handoff observation E04
message version E04
stream finalization E04
sender-receiver correlation E04
canonical handoff record E04
interoperability record E04,E12
communication boundary E04,E10
four-position boundary E10
sender input E10,E12
sender output E10
submitted payload E10,E12
receiver input E10
proposed operation E13
dispatch attempt E13
instruction fingerprint E05
condition-scoped fingerprint E05
behavioral fingerprint E05
canonicalization E05
behavioral evidence E05,E06
diagnostic evidence manifest E06
observation package E06
evidence state model E06
reproducibility invariant E06
test-context invariant E06
adaptive witness selection E08
expected information gain E08
cost-aware selection E08
diagnostic budget E08
stopping rule E08
abstention E08,E09,E10
identifiability E08
observational equivalence E08,E09
hypothesis space E08
hypothesis state E08
posterior update E08
decision-loss selection E08
partition selection E08
robust selection E08
uncertainty-aware comparison E09
materiality threshold E09
practical equivalence E09
confidence sequence E09
sequential testing E09
paired comparison E09
effect interval E09
multiple comparison E09
categorical response distribution E09
statistical decision category E09
insufficient evidence E09
invalid comparison E09
legitimate factual sensitivity E07,E09
unauthorized-instruction sensitivity E07,E09
earliest observable divergence E10
fault interval E10
bounded fault interval E10
evidence tier E10,E14,E18
localization E10
multi-hop localization E10
communication conversion E10
converter localization E10
receiver-side failure E10
sender-side omission E10
boundary restoration E10
multiple-fault E10,E11
masked fault E11
branching event graph E11
merge fault E11
branch fault E11
branch interaction E11
divergence frontier E11
causally consistent cut E11
cooperative attribution E11
protocol-independent adapter E12
native-to-canonical mapping E12
adapter capability manifest E12
coverage vector E12
observation coverage E12
cross-protocol correlation E12
three-plane adapter E12
legacy API adapter E12
metadata-only agent E12
instruction-message-action chain E13
typed correspondence E13
action discrepancy E13
aggregate constraint E13
temporal prerequisite E13
proposal-to-dispatch comparison E13
TOCTOU E15
semantic TOCTOU E15
selective revalidation E15
bitemporal E15
dependency impact graph E15
evidence applicability E15
historical only E15,E18
version incomparable E15
fault obligation E16
correction candidate E16
repair coverage E16
root correction E16
residual fault E16
correction lifecycle E16
regression witness E17
metamorphic relation E17
correction ablation E17
held-out test E17
correction durability E17
evidence record E18
claim support E18
evidence dependency closure E18
evidence sufficiency E18
verification vector E18
selective disclosure E18
redaction E18
reproducibility class E18
historical causal attribution E18
supersession E18
diagnostic exporter E18
diagnostic-only E01-E18
execution authorization outside E01-E18
non-authority E01-E18
Table 3: Compound phrases (88)
Compound phrase Embodiments
counterfactual diagnosis of instruction drift E07
paired tests separating fact from instruction by source role E02,E07
condition-scoped source-role separation E02
typed originating task condition E01
earliest observable divergence boundary E10
bounded fault interval where intermediate stages are unobserved E10
adaptive selection of diagnostic witnesses under a budget E08
uncertainty-aware comparison of agent interpretations E09
evidence-bounded fault localization E10,E18
abstention when evidence is insufficient E08,E09,E10
unauthorized instruction promotion from a lower-trust source E14
tool output containing an unsupported modification claim E14
legitimate factual price change versus unauthorized budget claim E07,E09
instruction fingerprint separate from message hash E05
cryptographic equivalence versus semantic equivalence E05
version-bound handoff lineage E04
sender-submitted versus receiver-observed message E04,E12
message-representation substitution at a boundary E07
receiver reference-condition test E10
correct verbal rule but non-compliant proposed operation E13
selective revalidation of affected witnesses E15
semantic time-of-check to time-of-use E15
evidence applicability class E15
correction of the implicated condition E16
regression witness for a corrected fault E17
positive-behavior preservation after correction E17
claim-specific evidence sufficiency E18
selective disclosure of diagnostic evidence E18
historical causal attribution qualified by assumptions E18
protocol-independent diagnostic adapter E12
adapter capability and coverage manifest E12
legacy service API request-response graph E12
non-effectuating diagnostic execution E12
branch substitution intervention E11
branch versus merge fault separation E11
masked divergence in parallel branches E11
factorial fact and instruction test E07,E14
boundary-value factual witness E07
instruction-source sensitivity measure E07
factual sensitivity measure E07
matched-pair invariant validation E06
incompatible diagnostic pair E06
version-dependent evidence invalidation E05,E06
condition-scoped taint propagation E14
provenance versus causal influence E14
persistent memory authority laundering E14
multilingual source transformation E14
source attribution loss without action change E14
multiple lower-trust source segments E14
source claim that does not cause the final operation E14
multi-hop source influence localization E14
communication conversion interval E10
non-monotonic condition propagation E10
sender transformation independent of source E10
receiver action-condition failure E10
partial observability fault interval E10
multiple independent faults E10,E11
joint branch interaction E11
redundant fault injection E11
expected information gain per unit cost E08
decision-loss witness selection E08
nonprobabilistic hypothesis partition selection E08
robust selection under uncertain likelihood E08
observational equivalence under feasible witnesses E08
stopping and abstention rule E08
diagnostic budget exhaustion E08
sequentially valid interpretation comparison E09
practical equivalence testing E09
signed probability difference for condition violation E09
predeclared materiality threshold E09
temporal validity of diagnostic evidence E15
bitemporal diagnostic record E15
dependency impact graph for revalidation E15
instruction-message-action correspondence E13
proposal-to-dispatch argument comparison E13
aggregate expenditure condition E13
temporal prerequisite correspondence E13
remote black-box action service E13
fault obligation and repair coverage E16
correction ablation procedure E17
metamorphic regression relation E17
cross-context recurrence of a fault E17
evidence dependency closure E18
evidence verification vector E18
controlled witness export manifest E18
exact-hop versus bounded localization E18
replay of diagnostic evidence E18
diagnostic finding without execution authority E01-E18

D.1. Boolean search strings

  • ("counterfactual" OR "matched pair") AND ("source role" OR "source-role") AND ("instruction" AND "fact")
  • ("agent" W/5 "handoff") AND ("fault localization" OR "earliest divergence")
  • ("prompt injection" OR "injected instruction") AND ("counterfactual" OR "intervention") AND ("agent")
  • ("adaptive" W/3 "witness") AND ("information gain" OR "hypothesis")
  • ("task contract" OR "typed condition") AND ("agent" W/10 "message")
  • ("modification eligibility" OR "eligible to modify") AND ("source")
  • ("non-effectuating" OR "no external effect") AND ("diagnostic") AND ("agent")
  • ("fault interval" OR "bounded interval") AND ("unobserved") AND ("hop" OR "handoff")
  • ("abstention" OR "insufficient evidence") AND ("localization" OR "diagnosis") AND ("agent")
  • ("practical equivalence" OR "materiality threshold") AND ("agent" AND "counterfactual")
  • ("branch" AND "merge") AND ("agent") AND ("factorial" OR "interaction")
  • ("adapter") AND ("A2A" OR "agent-to-agent") AND ("canonical event" OR "coverage")
  • ("time-of-check" OR "TOCTOU") AND ("semantic") AND ("revalidation")
  • ("regression" AND "correction") AND ("agent") AND ("fault" OR "counterfactual")
  • ("evidence" W/5 "selective disclosure") AND ("diagnostic") AND ("agent")
  • ("cryptographic equivalence" OR "hash") AND ("semantic equivalence") AND ("instruction")
  • ("unauthorized" W/3 "instruction") AND ("lower-trust" OR "supplier") AND ("budget" OR "limit")
  • ("tool output" OR "tool result") AND ("instruction promotion" OR "authority laundering")

Appendix E. Appendix E: European and Global Industry Mapping

The industry matrix is in <xref target="industry"/> and the European discussion in <xref target="europe"/>; both are non-normative, hypothetical and unverified against vendor documentation.

Appendix F. Appendix F: Patent and Public-Disclosure Cross-References

Cross-references to public records. Items not fetched by the editor are marked.

Table 4: Source verification
Record Status of verification
Indian Provisional 202631134241 (8 Nov 2026) Read from supplied PDF (partial); not verified against registry
PCT/IB2026/055615; WO 2026/150382 A2 Dates as supplied by the author; verify on PATENTSCOPE
Zenodo 10.5281/zenodo.24076231 Not fetched; title and date as given by the author
TD Commons 12934, 12973, 12542 Not fetched (access blocked); titles as given by the author
draft-agentic-ai-usecases-requirements Search snippets read; full text not read; later revisions exist
draft-srivastava-agentproto-handoff-context-00 Fetched and summarized
draft-luna-eagf-agent-delegation-semantics-00 Fetched and summarized
Indian application 202631131650 (3 Nov 2026) Mentioned only; kept distinct; not read

Appendix G. Appendix G: Reproducibility and Test Checklist

Self-audit (summary): diagnostic-only stance kept; 12 distinctions, components A to J, 12 figures, tests A to E, pseudocode, uncertainty section, 6 evidence records, 24 threats, 3 modes, legacy limit, industry matrix (labeled hypothetical, unverified), European section, prior-art comparison, worked examples A to F, validation plan, IPR and four notices, and the examiner index are present. Known gaps: JSON unavailable, filing read partially, cited sources partly unfetched, vendor documentation not verified.

Author's Address

Sangam Kumar Das
Independent Inventor
Balasore 756001
Odisha
India