Network Working Group S. Das Internet-Draft Independent Inventor Intended status: Informational 17 November 2026 Expires: 21 May 2027 AI Safety and AI Governance for Social Media Algorithms During Elections: Protected Algorithm Verification and Controlled Output Finality, with a Detailed Account of Thirteen Embodiments draft-das-electoral-algorithmic-integrity-finality-00 Abstract Computation is not authority to make an output effective. Social media platforms, search engines, advertising systems and recommendation engines decide what very large numbers of people see, in what order and how widely it spreads. During an election, a referendum or a declared emergency, an undisclosed or unapproved change to such an algorithm could change what voters see without any visible change in the content itself. Outside observers have no practical way to check which algorithm was actually running when a given output reached the public, and proprietary models cannot simply be opened to inspection. This document describes a technical architecture that verifies the algorithmic process and not the opinions expressed. A protected measurement step identifies the algorithm instance that is actually running and derives a fingerprint of its active configuration. The fingerprint is compared with an approved reference that applies to the protected period. An independent validator binds the verified algorithmic state to evidence about a specific pending output and issues a signed validation receipt. A separately scoped, non-bearer authorization may then allow an enforcement point to release the output. Before release, the enforcement point independently reconstructs the actual pending operation and checks the receipt, the authorization and its own protected state. If the running algorithm differs from the approved one, the output is withheld or routed through an explicitly authorized fallback. The document describes thirteen embodiments in detail: separation of organic speech from algorithmic amplification, dynamic extraction of the algorithm fingerprint, approved reference registers, server and edge enforcement, transitions into and out of election and emergency periods, correspondence between execution and output, coverage of alternative delivery routes, distributed execution reconstruction, device-side display enforcement, degraded-mode recovery, continuous drift detection, confidential verification, and atomic consumption of authorization. For each embodiment it states what the embodiment Das Expires 21 May 2027 [Page 1] Internet-Draft Algorithmic Integrity and Finality November 2026 teaches, how it works, where its enforcement point sits, how it fails, and what it does not claim. A cross-reference table connects each embodiment to a public Zenodo record that summarizes the same disclosure. This document is an individual, informational technical proposal. It makes no allegation about any platform, party or election, it is not a legal opinion or a certification, and it does not claim that any law requires this architecture. Source, Patent Basis and Notices This note is to be removed before publishing as an RFC. Based on an Indian provisional patent application filed 25 October 2026, and disclosed with the Mothership (PCT/IB2026/055615) on 16 July 2026. The Mothership application was filed on 4 June 2026 and is the foundational framework. The algorithm-integrity measurement builds on the earlier algorithm-fingerprint application PCT/ IB2026/053385. The 25 October specification is a 481-page document with 13 separately headed embodiments (1A and 1 to 12). Where to read the full PCT filing: search "Sangam Das" on WIPO PATENTSCOPE (https://patentscope.wipo.int/); search PCT/IB2026/055615 (pctib202655615); or read the PCT file directly in Technical Disclosure Commons at https://www.tdcommons.org/dpubs_series/12542/. Public cross-reference: the Zenodo record "Algorithmic Manipulation by Social Media and the Protection of Elections: AI Safety and AI Governance for the European Democracy Shield and Other Democracies", DOI 10.5281/zenodo.24300261, https://zenodo.org/records/24300261, published 17 November 2026 under CC BY-NC 4.0. Section 7 maps each statement of that record to the embodiments in this document. AI-assisted drafting notice: this text, including the policy- relevance material, was prepared with AI assistance and may contain mistakes. Readers must cross-check it against the original public disclosures, the patent filing copy and the official legal texts before drawing any conclusion. The application number of the 25 October filing is not stated in the supplied specification and is therefore not given here. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Das Expires 21 May 2027 [Page 2] Internet-Draft Algorithmic Integrity and Finality November 2026 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 21 May 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1. The Problem . . . . . . . . . . . . . . . . . . . . . . . 4 1.2. Guiding Principle . . . . . . . . . . . . . . . . . . . . 5 1.3. Purpose and Non-Goals . . . . . . . . . . . . . . . . . . 5 1.4. Relationship to Other Documents . . . . . . . . . . . . . 6 1.5. How to Read This Document . . . . . . . . . . . . . . . . 6 1.5.1. Requirements Language . . . . . . . . . . . . . . . . 6 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 7 3. Problem Statement and Threat Model . . . . . . . . . . . . . 9 3.1. Adversaries and Faults . . . . . . . . . . . . . . . . . 9 3.2. Trust Assumptions . . . . . . . . . . . . . . . . . . . . 10 4. Common Architecture . . . . . . . . . . . . . . . . . . . . . 10 4.1. The Protected Output-Finality Sequence . . . . . . . . . 10 4.2. Three Evidence Objects . . . . . . . . . . . . . . . . . 11 4.3. Roles . . . . . . . . . . . . . . . . . . . . . . . . . . 12 4.4. Separation of Measurement, Validation and Finality . . . 12 5. Candidate Enforcement-Point Requirements . . . . . . . . . . 13 6. The Thirteen Embodiments . . . . . . . . . . . . . . . . . . 15 6.1. Embodiment 1A: Organic Speech Separation . . . . . . . . 15 6.2. Embodiment 1: Dynamic Algorithm Fingerprint Extraction . 17 6.3. Embodiment 2: Approved Reference Register . . . . . . . . 20 6.4. Embodiment 3: Server and Edge Output Finality . . . . . . 21 6.5. Embodiment 4: Election and Emergency Mode Transitions . . 24 Das Expires 21 May 2027 [Page 3] Internet-Draft Algorithmic Integrity and Finality November 2026 6.6. Embodiment 5: Execution-to-Output Correspondence . . . . 26 6.7. Embodiment 6: Alternative-Route and Legacy-Path Coverage . . . . . . . . . . . . . . . . . . . . . . . . 27 6.8. Embodiment 7: Distributed Execution Reconstruction . . . 29 6.9. Embodiment 8: Device-Side Display Finality . . . . . . . 31 6.10. Embodiment 9: Degraded-Mode Recovery . . . . . . . . . . 33 6.11. Embodiment 10: Continuous Algorithm Drift Detection . . . 34 6.12. Embodiment 11: Confidential Verification . . . . . . . . 36 6.13. Embodiment 12: Atomic Finality Consumption . . . . . . . 37 7. Cross-Reference to the Public Zenodo Record . . . . . . . . . 39 8. Relationship to Public Policy Work . . . . . . . . . . . . . 41 9. Security Considerations . . . . . . . . . . . . . . . . . . . 41 10. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 42 11. Notices, Intellectual Property and Non-Endorsement . . . . . 42 12. References . . . . . . . . . . . . . . . . . . . . . . . . . 43 12.1. Normative References . . . . . . . . . . . . . . . . . . 43 12.2. Informative References . . . . . . . . . . . . . . . . . 43 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . 44 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 44 1. Introduction 1.1. The Problem Modern information platforms contain algorithmic systems that select, rank, recommend, amplify, suppress, filter, personalize and present information. These systems are proprietary and change frequently. New model versions are activated, configuration values are tuned, experiments are routed to subsets of users, features are added and fallback models are switched in when something fails. During a period of heightened democratic sensitivity, such as an election, a referendum, a civic consultation or a declared public emergency, the integrity of these systems is a matter of public interest. Three difficulties arise together. * A change to an algorithm can alter what voters see without any visible change in the content, so the change cannot be detected by looking at content alone. * An outside observer, including an election body or a regulator, generally cannot determine which algorithm was running when a particular output was delivered. A statement by the operator that nothing changed is a report and not evidence. * Proprietary algorithms cannot simply be disclosed, and any mechanism that decides which political views are acceptable would be both undesirable and dangerous. Das Expires 21 May 2027 [Page 4] Internet-Draft Algorithmic Integrity and Finality November 2026 The architecture in this document addresses the three difficulties together. It verifies that the algorithm that actually ran is an approved one, that the specific output about to be delivered really came from it, and that nothing was substituted afterwards. It does this without disclosing the algorithm and without evaluating the content or viewpoint of any speech. 1.2. Guiding Principle Computation is not authority to make an output effective. An algorithm may be approved, may have run correctly and may have produced an output, and the output is still not authorized to reach the public until an independent enforcement point has checked, at the actual point of release, that this specific output corresponds to verified execution under current policy. A second principle governs how the architecture treats expression. The creation of expression and the algorithmic amplification of that expression are technically distinct operations. The architecture may verify and control the second without requiring approval of the first. Algorithmic-integrity enforcement is not authority to decide which political opinions may be expressed. 1.3. Purpose and Non-Goals The purpose of this document is to describe the architecture and its thirteen embodiments in enough detail for a reviewer to understand what each one does, where its enforcement point sits, and what it does not do. The following are non-goals. * This document defines no wire format, message encoding, registry or transport. Where an encoding is needed, existing formats are candidates. * It does not decide what content is true, lawful, acceptable or fair. A valid measurement of an algorithm does not establish that the algorithm is neutral. * It does not claim that the architecture prevents every manipulation. It protects the delivery paths that pass through an enforcement point, and it states where that coverage ends. * It makes no experimental, performance, latency or adoption claim. Das Expires 21 May 2027 [Page 5] Internet-Draft Algorithmic Integrity and Finality November 2026 * It does not replace platform risk assessment, human oversight, independent audit, legal process or the work of election bodies. It is intended to give those processes verifiable technical evidence. 1.4. Relationship to Other Documents This document is derived from the 25 October 2026 specification. That specification is distinct from the separate 22 October and 23 October 2026 specifications of the same author, and the number and content of embodiments must not be conflated across them. An earlier Internet-Draft by the same author, [I-D.das-social-media-algorithmic-accountability], addresses the same subject at a higher level. The present document differs in that it describes each of the thirteen embodiments in turn and cross- references them to the public Zenodo record [ZENODO-24300261]. An earlier Zenodo record, [ZENODO-23659488], holds the source specification of that earlier draft. The full text of the Mothership application is available in Technical Disclosure Commons [TDC-12542] and in WIPO PATENTSCOPE [PCT-055615]. 1.5. How to Read This Document Section 4 gives the common architecture that all embodiments share. Section 5 states candidate enforcement-point requirements. Section 6 describes the embodiments in the order used by the source specification, which places Embodiment 1A first because it sets the limits on everything else. Each embodiment section uses the same sub-headings: purpose, what it teaches, how it works, enforcement point, failure handling, limits, and cross-references. Section 7 contains the cross-reference tables. The document is intentionally long, because each embodiment addresses a distinct failure mode and a summary would hide the differences. The terms "MUST", "SHOULD" and "MAY" appear in Section 5 only, and only describe what an implementation would need to do to claim the stated property. They are not requirements on the reader. 1.5.1. Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. Das Expires 21 May 2027 [Page 6] Internet-Draft Algorithmic Integrity and Finality November 2026 2. Terminology The source specification uses its own terms. This document uses industry-standard terms where they exist and gives the source term in the second column of the table. +======================+======================+=====================+ | Term in this | Source term | Meaning | | document | | | +======================+======================+=====================+ | Algorithm | Algorithmic Logic | A representation | | fingerprint | Fingerprint (ALF) | derived from | | | | measured | | | | properties of the | | | | algorithm | | | | configuration | | | | that is actually | | | | active. It need | | | | not contain model | | | | parameters or | | | | reproduce the | | | | computation. | +----------------------+----------------------+---------------------+ | Identity fingerprint | ALF-I | Measured | | | | properties of an | | | | activated | | | | algorithmic | | | | configuration | | | | (which model, | | | | which | | | | configuration, | | | | which | | | | dependencies). | +----------------------+----------------------+---------------------+ | Execution | ALF-E | Protected | | fingerprint | | evidence of | | | | actual execution: | | | | events, stage | | | | associations, | | | | observed | | | | processing | | | | conditions. | +----------------------+----------------------+---------------------+ | Output commitment | ALF-O | A commitment that | | | | binds execution | | | | evidence to one | | | | particular | | | | pending output. | Das Expires 21 May 2027 [Page 7] Internet-Draft Algorithmic Integrity and Finality November 2026 +----------------------+----------------------+---------------------+ | Approved reference | Approved ALF | A registered | | | Reference | fingerprint, | | | | configuration | | | | family or | | | | constraint that | | | | is approved for a | | | | protected period. | +----------------------+----------------------+---------------------+ | Policy epoch | Policy Epoch | A protected | | | | generation number | | | | of the | | | | authorization | | | | state. | +----------------------+----------------------+---------------------+ | Protected validator | Protected Validation | An isolated, | | | Domain | independent | | | | component that | | | | evaluates | | | | evidence and | | | | issues receipts. | +----------------------+----------------------+---------------------+ | Validation receipt | Validation Receipt | A | | | | cryptographically | | | | authenticated | | | | record of a | | | | validation | | | | determination. | +----------------------+----------------------+---------------------+ | Non-bearer | Non-Bearer Finality | A scoped | | authorization | Enablement | permission bound | | | | to one operation, | | | | one enforcement | | | | point and one | | | | policy state. | | | | Possessing it | | | | does not confer | | | | authority | | | | elsewhere. | +----------------------+----------------------+---------------------+ | Enforcement point | Finality Sink | The component at | | | | the real release | | | | boundary that can | | | | prevent an output | | | | from becoming | | | | effective. | +----------------------+----------------------+---------------------+ | Pending operation | Actual Pending | The operation | Das Expires 21 May 2027 [Page 8] Internet-Draft Algorithmic Integrity and Finality November 2026 | | Operation | that is actually | | | | about to occur, | | | | as observed or | | | | reconstructed at | | | | the enforcement | | | | point, not as | | | | described by the | | | | workload. | +----------------------+----------------------+---------------------+ | Non-effective state | Non-Effective State | The state in | | | | which a proposed | | | | output exists but | | | | has not yet been | | | | authorized to | | | | produce its | | | | external effect. | +----------------------+----------------------+---------------------+ | Route coverage | Effectuation-Path | The extent to | | | Coverage | which every path | | | | by which an | | | | output can reach | | | | its target passes | | | | an enforcement | | | | point. | +----------------------+----------------------+---------------------+ | Counter-verification | Independent Counter- | An independent | | | Verification | re-examination of | | | | a disputed | | | | enforcement | | | | decision. | +----------------------+----------------------+---------------------+ Table 1: Terminology Used in This Document 3. Problem Statement and Threat Model 3.1. Adversaries and Faults The architecture considers the following adversaries and faults. It does not assume they are all present in any given deployment. * An unapproved or undisclosed change to a ranking, recommendation or advertising algorithm, made by an insider, a contractor, a compromised build pipeline or an automated update. * A workload that misreports which algorithm or which stages produced an output. Das Expires 21 May 2027 [Page 9] Internet-Draft Algorithmic Integrity and Finality November 2026 * Substitution of an output after it was validated, including through reranking, compression, localization, caching or device- side rendering. * Delivery through a route that bypasses the enforcement point, such as a cache, a legacy endpoint, a direct-origin connection, an experimental branch or a fallback path. * Replay of a valid receipt or authorization, or concurrent claims on one single-use authorization. * Stale policy, a rolled-back policy state, or a node that has not received a revocation. * Compromise or failure of the measurement source, the validator or the approval registry. * Misuse of the enforcement system itself, for example to suppress ordinary speech, or labelling algorithmic amplification as ordinary speech to avoid checks. 3.2. Trust Assumptions The measurement source is outside the sole control of the algorithm it measures. The protected validator and the enforcement point run in protected domains that the workload cannot modify. Cryptographic keys for receipts and authorizations are held in the protected domains. The enforcement point sits at a boundary that a covered output must cross. Every assurance in this document is tied to the component that can provide it, and each embodiment states what happens when its assumptions fail. 4. Common Architecture 4.1. The Protected Output-Finality Sequence All embodiments extend the same seven-step sequence. 1. Measure. The actual execution instance and its active dependencies are measured by a protected measurement source. 2. Fingerprint. An identity fingerprint and an execution fingerprint are established and compared with an approved reference. 3. Check authority. The applicable approval, policy epoch and revocation state are checked. Das Expires 21 May 2027 [Page 10] Internet-Draft Algorithmic Integrity and Finality November 2026 4. Commit. The pending output is committed (output commitment) and independently validated by the protected validator. 5. Authenticate. An authenticated validation receipt and a separately scoped non-bearer authorization are associated with the pending output. 6. Reconstruct and verify. The enforcement point reconstructs the actual pending output and verifies the receipt, the authorization and its own protected local state. 7. Decide. The output is released, withheld, quarantined or sent to bounded recovery, and the outcome is recorded. +-----------+ +-------------+ +--------------+ | Algorithm |-->| Measurement |-->| Approved ref | | (running) | | source | | + policy | +-----------+ +-------------+ | epoch check | | +------+-------+ | pending output | v v +--------------+ +--------------------------------+ | Non-effective|-->| Protected validator | | staging | | binds fingerprint to output | +--------------+ +---------------+----------------+ | validation receipt + scoped, non-bearer authorization v +-------------------------------+ | Enforcement point | | reconstructs actual pending | | output, checks receipt, | | authorization, local state | +-------+---------------+-------+ | | release withhold / fallback Figure 1: Protected Output-Finality Sequence 4.2. Three Evidence Objects The architecture keeps three kinds of evidence apart because they prove different things. Das Expires 21 May 2027 [Page 11] Internet-Draft Algorithmic Integrity and Finality November 2026 * The identity fingerprint shows what configuration was activated. A match does not show that this configuration produced a particular output. * The execution fingerprint shows what actually ran. It must rest on protected observation and is not trustworthy merely because it is signed. * The output commitment binds the first two to the actual pending operation. It must bind what the enforcement point sees, and not only a description supplied by an untrusted workload. 4.3. Roles Measurement source: Observes which algorithmic components were actually activated. Must be outside the sole control of the measured algorithm. Approval authority and registry: Registers approved references, activation, suspension and revocation, with authenticated, versioned state. Protected validator: Evaluates integrity, freshness, completeness and authorization of the evidence, and issues receipts. Enforcement point: Sits where an output becomes effective (server response, gateway, edge, device presentation) and permits or refuses release. Counter-verifier: Independently re-examines a disputed decision using protected evidence. Does not judge content. Policy authority: Defines operating modes, required validation profiles and enforcement scope, with authenticated transitions. 4.4. Separation of Measurement, Validation and Finality Measurement tells the system what ran. Validation decides whether the evidence supports a determination. Finality decides whether this specific output may become effective now. The three are kept in separate protected domains so that a failure or compromise in one does not silently grant the authority of another. A receipt records a determination, and the separately scoped authorization is what the enforcement point accepts as permission. Das Expires 21 May 2027 [Page 12] Internet-Draft Algorithmic Integrity and Finality November 2026 5. Candidate Enforcement-Point Requirements This section restates the architecture as enforcement points. Each item is a property that an implementation would need to provide in order to claim it, and each is supported by one or more embodiments in Section 6. They are candidate requirements for discussion, and they do not define a standard. +=======+==========================================+=============+ | ID | Enforcement point | Embodiments | +=======+==========================================+=============+ | EP-1 | The measurement source MUST be outside | 1, 5, 7, 10 | | | the sole control of the algorithm it | | | | measures, and a signed artifact hash | | | | alone MUST NOT be treated as proof that | | | | the artifact executed. | | +-------+------------------------------------------+-------------+ | EP-2 | An approved reference MUST be registered | 2, 4 | | | with authenticated lifecycle state, a | | | | policy epoch and a freshness bound. | | | | Approval of a configuration MUST NOT be | | | | treated as approval of every output. | | +-------+------------------------------------------+-------------+ | EP-3 | A covered output MUST be held in a non- | 3, 12 | | | effective state until the enforcement | | | | point has verified the release | | | | conditions. | | +-------+------------------------------------------+-------------+ | EP-4 | The enforcement point MUST reconstruct | 3, 5, 8, 12 | | | or independently obtain the actual | | | | pending operation and MUST NOT rely only | | | | on the workload's statement that it is | | | | the one validated. | | +-------+------------------------------------------+-------------+ | EP-5 | A receipt MUST be authenticated and | 3, 4, 5 | | | bound to the pending operation | | | | commitment, the policy epoch and the | | | | operating mode. | | +-------+------------------------------------------+-------------+ | EP-6 | Authorization MUST be separately scoped, | 3, 12 | | | non-bearer, and consumed under sink- | | | | local protected state with replay and | | | | concurrency protection. | | +-------+------------------------------------------+-------------+ | EP-7 | Operating-mode and policy transitions | 4 | | | MUST use authenticated, monotonic | | | | generations, and the enforcement point | | | | MUST evaluate the current mode | | Das Expires 21 May 2027 [Page 13] Internet-Draft Algorithmic Integrity and Finality November 2026 | | immediately before release. | | +-------+------------------------------------------+-------------+ | EP-8 | Every route by which a covered output | 6 | | | can become effective MUST be inside a | | | | declared enforcement perimeter, and | | | | uncovered and unverified routes MUST be | | | | reported as such. | | +-------+------------------------------------------+-------------+ | EP-9 | Validation of a distributed pipeline | 7 | | | MUST check lineage completeness against | | | | a required-stage profile, and missing or | | | | contradictory evidence MUST NOT be | | | | treated as complete. | | +-------+------------------------------------------+-------------+ | EP-10 | Where device-side presentation is | 8 | | | covered, the device-side enforcement | | | | point MUST verify the actual | | | | presentation operation under a stated | | | | visibility-assurance level. | | +-------+------------------------------------------+-------------+ | EP-11 | A fallback MUST be separately | 9 | | | authorized, MUST preserve designated | | | | non-reducible requirements, and MUST NOT | | | | convert missing evidence into proof of | | | | compliance. | | +-------+------------------------------------------+-------------+ | EP-12 | Changes to the active configuration MUST | 10 | | | invalidate the affected approvals and | | | | pending outputs selectively, and | | | | behavioral drift MUST be treated as a | | | | trigger for verification and not as | | | | proof of manipulation. | | +-------+------------------------------------------+-------------+ | EP-13 | Confidential verification MUST bind any | 11 | | | proof to the actual execution evidence | | | | and to the actual pending output. | | +-------+------------------------------------------+-------------+ | EP-14 | Enforcement scope MUST separate direct | 1A | | | publication from algorithmic | | | | amplification, and a failure in one MUST | | | | NOT be inherited by the other. | | +-------+------------------------------------------+-------------+ | EP-15 | A disputed enforcement decision MUST be | 1A | | | open to independent counter-verification | | | | that cannot itself create release | | | | authority. | | +-------+------------------------------------------+-------------+ Das Expires 21 May 2027 [Page 14] Internet-Draft Algorithmic Integrity and Finality November 2026 Table 2: Candidate Enforcement-Point Requirements 6. The Thirteen Embodiments The embodiments are described in the order used by the source specification: Embodiment 1A first, then Embodiments 1 to 12. The "Source" line in each section gives the sections of the 25 October specification on which the description is based. Page numbers refer to the PDF of that specification. 6.1. Embodiment 1A: Organic Speech Separation Source Sections 1A.1 to 1A.40, from PDF page 35 onward. Purpose To prevent a system built to protect algorithmic integrity from turning into an unrestricted tool for controlling ordinary human expression. It distinguishes an act of expression from the later algorithmic distribution of that expression, identifies which operation is actually subject to enforcement, and provides independent review of disputed decisions. What it teaches Three activities are kept apart: expression creation (a person or organization authors or submits information), algorithmic processing (a system selects, ranks, recommends, amplifies, suppresses, filters or personalizes it), and controlled effectuation (a protected system verifies whether a particular algorithmic operation may produce its external effect). The architecture may verify and control the second and third without approving the first. An item of organic content can later take part in an algorithmic distribution operation, and then the two operations remain distinct. The organic origin of content does not exempt later algorithmic recommendation from the applicable checks. How it works 1. Operation classification. A protected classifier labels the actual pending operation, using protected observation or protected interface state, as one of: direct publication, direct communication, ordinary storage, algorithmic recommendation, ranking, amplification, filtering, presentation, or another registered class. An application- supplied label alone cannot exempt an amplification operation from enforcement. The classification is based on the actual operation and the system boundary and never on whether the content supports or criticizes any political actor. Das Expires 21 May 2027 [Page 15] Internet-Draft Algorithmic Integrity and Finality November 2026 2. Enforcement-scope registry. Separate scope records govern publication and algorithmic distribution, each with required validation profiles, enforcement boundaries, a generation and an authenticator. An independently authorized direct- publication operation is not rejected only because a separate recommendation operation lacks valid execution evidence. Where direct publication is itself lawfully subject to a different security requirement, that requirement still applies, so the architecture creates no general right to bypass existing communication security controls. 3. Non-inheritance of failure. If a recommendation operation fails validation, only that operation is denied. The system does not infer that the underlying publication is invalid. Content item X may have separate operations for submission (P1), direct publication (P2), ranking (R1), recommendation (R2) and device presentation (D1). If R2 fails a mandatory check it is withheld, P1 and P2 are evaluated on their own terms, and D1 may require its own authority where it presents R2. 4. Viewpoint-neutral predicates. Enforcement uses technical predicates (was the operation classified correctly, was the right policy selected, does the evidence support the decision) and never content-based predicates. The embodiment states that a finite test set cannot prove an algorithm will never discriminate and that a valid fingerprint does not establish political neutrality. It separates four things: integrity of the measured configuration, correct application of registered predicates, detection of inconsistent enforcement behavior, and broader fairness assessment. The first three can be tested technically, and the fourth needs governance, statistical or legal analysis outside the mechanism. 5. Counter-verification. When an operation is rejected, withheld, quarantined or disputed, an independent counter- verifier reconstructs the decision from protected evidence. It asks whether the right operation was classified, whether the right policy was selected, whether the evidence supports the decision, and whether the operation was wrongly tied to unrelated expressive activity. Its possible results are: original decision supported; scope misapplication detected; technical evidence mismatch; unauthorized predicate detected; indeterminate; or corrective revalidation required. These are technical classifications and not legal findings of unlawful censorship. Das Expires 21 May 2027 [Page 16] Internet-Draft Algorithmic Integrity and Finality November 2026 6. Corrective revalidation. A counter-verification result does not itself create release authority. A corrected output must pass fresh validation and enforcement-point checks, so counter-verification cannot become an override. Enforcement point The classifier and scope registry sit with the validator, and the effect is enforced at the same enforcement points as the other embodiments. Counter-verification is a separate protected process with its own state machine and abuse limits. Failure handling Two opposite attacks are covered. An untrusted workload may label ordinary publication as amplification, so that recommendation- specific controls are wrongly applied. The classifier compares the declared class with evidence of actual processing, routing, source operation and intended effect, and where evidence is insufficient it records the uncertainty and does not treat the subject matter of the expression as proof. Conversely, an operator may label algorithmic recommendation as organic speech to avoid checks, for example by describing a ranked feed as ordinary user publication, and a declarative label does not override protected execution evidence. Abuse of the counter-verification process, and counter-verification of a compromised validator, are also treated. Emergency-period safeguards and independent communication continuity keep direct communication available where the recommendation operation is withheld. Limits This embodiment does not determine the legal status of speech or what a jurisdiction must protect. Counter-verification must not become an unrestricted bypass of algorithmic finality. An untrusted assertion that an operation is protected speech does not by itself establish its technical authorization state. Privacy- preserving counter-verification limits disclosure of personal communications and proprietary algorithms. Cross-reference Supports EP-14 and EP-15. Uses the foundation in Embodiments 1 and 2, the covered boundary in Embodiment 3, the mode state in Embodiment 4, and the route coverage in Embodiment 6. The Zenodo record [ZENODO-24300261] describes it as keeping direct user publication separate from algorithmic amplification. 6.2. Embodiment 1: Dynamic Algorithm Fingerprint Extraction Source Sections 1.1 to 1.26, from PDF page 60 onward. Das Expires 21 May 2027 [Page 17] Internet-Draft Algorithmic Integrity and Finality November 2026 Purpose To measure the algorithm that is actually active, using cryptographic, protected non-cryptographic or hybrid mechanisms, without requiring disclosure of proprietary model internals and without requiring a common recommendation model. What it teaches The actual algorithmic logic performs the computation, and the fingerprint is a separate representation derived from selected properties of that logic or its execution state. An output is a function of model, parameters, features, configuration, user context and runtime state. The fingerprint extraction takes the measured components, the active configuration, the execution identity and a measurement schema. The fingerprint need not contain model parameters, the same fingerprint may accompany different outputs, and a matching fingerprint does not show that an output is fair, accurate, neutral or lawful. How it works 1. Identify the actual execution instance from a protected runtime, workload controller, model loader, process manager or orchestration service. 2. Identify execution-relevant components: executable logic, model artifacts, feature pipelines, ranking configuration, policy configuration and selected runtime dependencies. 3. Obtain protected evidence of which components were actually activated. A deployment manifest alone may be insufficient where the application can substitute a different runtime component. 4. Normalize the measurements into a defined schema with field types, component roles, generations, provenance and encoding rules. 5. Construct the fingerprint, register it with a protected measurement generation and validity conditions, and evaluate whether it corresponds to an approved reference, family or authorized transition. 6. Associate it with execution evidence where output-specific verification is needed, validate integrity, freshness, completeness and authorization, and hand a result to the receipt and finality processes. A measurement event does not by itself release any output. Das Expires 21 May 2027 [Page 18] Internet-Draft Algorithmic Integrity and Finality November 2026 Three extraction variants are described. The cryptographic variant uses hashes and signed measurements. The non- cryptographic variant uses protected internal identifiers and counters. The hybrid variant combines them. Legacy adapters let existing systems participate. Dynamic change detection watches for new model loads, version changes, configuration edits, feature-dependency changes, experimental branches, fallback model selection, component replacement and policy-epoch changes. When one occurs, the affected generation is marked stale, the new activated state is obtained, new measurements are taken, a new fingerprint generation is registered and approval correspondence is re-evaluated. Where the platform allows it, the generation advances atomically with activation. Enforcement point The measurement adapter sits inside or beside the model loader, orchestration layer or runtime, and reports to a protected registration. The measurement source must be outside the sole control of the algorithm. Measurement is kept separate from validation and finality. Failure handling The process distinguishes measurement valid, unavailable, stale, untrusted and incomplete, approval mismatch, policy-epoch mismatch and missing execution association. A failure condition is never reported as successful validation. For covered operations requiring authenticated finality, an unavailable or invalid measurement may block the required authorization. A separately authorized fallback may handle unsupported legacy systems or specified emergencies, and must not silently turn missing evidence into proof of compliance. Limits A signed artifact hash alone does not prove that the artifact actually executed. Where the measurement system cannot reliably observe a change, the deployment must state a lower assurance level, add controls, or reject measurements that fail freshness requirements. Cross-reference Supports EP-1. Feeds Embodiments 2, 5, 7 and 10. The Zenodo record [ZENODO-24300261] describes this as the protected step that identifies the running algorithm and fingerprints its active configuration. Das Expires 21 May 2027 [Page 19] Internet-Draft Algorithmic Integrity and Finality November 2026 6.3. Embodiment 2: Approved Reference Register Source Sections 2.1 to 2.32, from PDF page 77 onward. Purpose To register approved algorithm references and the policy states that apply to protected periods, with authenticated activation, suspension, revocation and version-aware authority. What it teaches Approval of a configuration is a state with a lifecycle and not a permanent fact. A previously approved identity does not authorize every future execution or output, and it stays authoritative only while the applicable policy, freshness, scope and execution- verification conditions hold. Approval of a configuration is not approval of every output or of any political viewpoint. How it works 1. An authorized approval authority submits an approval instruction identifying a reference fingerprint, a permitted configuration family or a constraint. 2. The approval environment verifies the authority's identity and permissions. 3. A structured approval record is built (reference, scope, policy identifier, validity conditions, permitted transitions), registered in protected state and associated with an operating mode, policy epoch, protected interval and deployment scope. 4. A protected approval snapshot is generated and distributed to regional servers, edge nodes, gateways or device-side validators, either as exact-match approval or as approved configuration-family matching. 5. A validator compares an actual fingerprint with a local valid snapshot. A successful match is one input to later output- specific validation and never grants release by itself. Operating modes include normal, pre-election, election-protected, referendum-protected, emergency-protected and recovery. The system does not conclude that an election is under way because an untrusted application supplies a flag. The policy epoch advances through a protected function of the current epoch and an authorized transition, and an old snapshot does not stay authoritative merely because it was once valid. The approval Das Expires 21 May 2027 [Page 20] Internet-Draft Algorithmic Integrity and Finality November 2026 lifecycle has the states proposed, pending verification, approved- inactive, active, suspended, revoked, expired and superseded. A revoked record is not reactivated under the same identity, and a suspended record returns to active only through a newly authorized transition. The lifecycle may be implemented with protected database transactions, monotonic records or append-only authenticated logs. A snapshot carries its identifier, policy and epoch, a commitment to the approval set, effective and expiry times, deployment scope, revocation generation, required assurance level and an authenticator. Enforcement point The approval registry is protected state held by the policy authority. Snapshots are verified by validators and enforcement points before installation into protected local state, which allows bounded-latency local checks. Failure handling A validator rejects a snapshot whose interval has expired, whose epoch is superseded or for which a revocation has been received. Absence of a received revocation is not proof that none occurred, so a distributed deployment must state its consistency and freshness guarantees. Where immediate revocation is needed, an enforcement point may need stronger synchronization than an asynchronously refreshed cache. Where bounded propagation is accepted, a maximum snapshot lifetime is enforced and operations are rejected when it expires. Concurrent policy transitions are handled explicitly. Limits The approval policy must be independently defined, versioned and applied consistently. Approval separate from finality: a valid approval and a matching fingerprint still need output-specific validation and enforcement-point verification. Cross-reference Supports EP-2. Depends on Embodiment 1 and feeds Embodiments 3, 4 and 10. The Zenodo record [ZENODO-24300261] describes the fingerprint as compared with an approved reference for the relevant period. Section 8 discusses a possible relationship to a change log for recommender systems. 6.4. Embodiment 3: Server and Edge Output Finality Source Sections 3.1 to 3.37, from PDF page 96 onward. Das Expires 21 May 2027 [Page 21] Internet-Draft Algorithmic Integrity and Finality November 2026 Purpose To place output-specific validation at server, API-gateway and edge response-release boundaries, so that a computed and approved response is still held until the enforcement point verifies the release conditions. What it teaches An algorithmically generated response can be computationally complete, can come from an approved configuration, and can still be unauthorized for external release. Possession of a receipt or model approval is not sufficient authority. Finality controls can be added to existing server, API and edge infrastructure without replacing it. How it works 1. Non-effective staging. A proposed response is held in a protected output buffer, gateway response queue, non-public object store, withheld network send or equivalent. Its states run from created, staged non-effective, validation pending, validated non-effective, finality pending and release authorized, to release committed, with rejection paths that terminate the operation. 2. Pending response descriptor and commitment. A descriptor of the response is constructed and canonicalized, and a commitment to it is created. The actual execution is bound to the response. 3. Protected validation. The validator checks the binding and issues an authenticated receipt. 4. Separately scoped enablement. A non-bearer authorization is associated with the receipt and the pending operation. 5. Independent verification at the enforcement point. The enforcement point observes the response at the release boundary, commits to the actual descriptor, and denies release unless the receipt is authentic, fresh, consistent with the policy state, matches the commitment, comes with a valid enablement, and the atomic state transition for the authorization succeeds. It must not rely only on the workload's assertion that the response is the same as the one validated. Implementations include a legacy API gateway, a reverse proxy, a service mesh or sidecar and an edge-delivery gateway. Cached responses need a distinction between authorization of the original response and authorization of a later delivery. A cache-delivery Das Expires 21 May 2027 [Page 22] Internet-Draft Algorithmic Integrity and Finality November 2026 descriptor carries the cached object commitment, request context, recipient scope, policy epoch, validity generation and sink identity, and where policy requires a decision per delivery, the gateway performs a fresh check for each. Streaming responses are divided into bounded segments with a stream identifier, sequence number, commitment, previous-segment commitment, policy epoch and authorized sink, and an initial stream authorization is not unlimited authority for later segments. Transformations such as compression, content encoding and localization are handled by validating after transformation, by an authenticated transformation record, or by a commitment over the exact final bytes. A revocation during an in-flight response is addressed. Enforcement point A protected API response gate, network-egress proxy, service-mesh enforcement component, application-server output controller or edge-delivery gateway, positioned where the covered operation can still be prevented. Failure handling The embodiment states that an authorization may be consumed before transmission completes, so an implementation must define whether finality means authorization to submit, handoff to a protected transport or durable delivery commitment. The authorization states are unused, reserved, aborted, release submitted and consumed. If the outcome cannot be determined after a crash, a single-use authorization is not replayed without a protected recovery procedure, and the architecture does not claim that network delivery and authorization consumption are physically atomic when the transport gives no commit guarantee. Security against response substitution, receipt replay and gateway bypass is analyzed, and latency and throughput are discussed without a performance claim. Limits A server-side enforcement point controls a server-side or network- delivery boundary. It does not necessarily control what appears on a user's screen, because a downstream application may modify, suppress or reorder content afterwards. Server-side finality and display finality are distinct (see Embodiment 8). A gateway alone does not prove that every response path is controlled (see Embodiment 6). Cross-reference Supports EP-3 to EP-6. Builds on Embodiments 1 and 2 and is hardened by Embodiments 5, 6 and 12. The Zenodo record [ZENODO-24300261] describes enforcement at various delivery points. Das Expires 21 May 2027 [Page 23] Internet-Draft Algorithmic Integrity and Finality November 2026 6.5. Embodiment 4: Election and Emergency Mode Transitions Source Sections 4.1 to 4.41, from PDF page 119 onward. Purpose To govern changes among ordinary, election-protected, emergency and recovery modes using authenticated policy activation, monotonic authorization generations, distributed revocation and fail-closed control. What it teaches A protected mode is not an application-level flag or a remotely supplied label. It is authenticated, scope-bound, versioned and protected authorization state, and its transitions follow defined authority, ordering, effective-condition and anti-rollback rules. A mode does not decide whether content is legal or acceptable. It only selects which technical verification, evidence and enforcement requirements apply. A previously approved configuration or a previously issued receipt does not stay sufficient after a protected policy transition. How it works 1. Modes. An illustrative set is normal, pre-election, election- protected, referendum-protected, emergency-protected and recovery-restricted. Names are non-limiting, and deployments may define regional, functional or risk-specific modes, each tied to an authenticated policy definition. 2. Transition. An authenticated transition instruction is verified for authority and applicability. Each accepted transition advances a protected generation, and a node rejects a transition older than its established generation unless a defined recovery protocol permits reconstruction. A simple application-controlled counter is not presumed sufficient. The lifecycle runs proposed, authenticated, authorized, prepared, activation pending and active, with superseded, suspended and recovery-restricted branches. 3. Preparation. Policy for a scheduled period can be distributed in advance and pre-installed inactive on enforcement nodes. When an authenticated activation condition is met, the node switches, which reduces activation-time network dependency. It needs a trustworthy time or event source and protection against manipulation of the activation condition. Das Expires 21 May 2027 [Page 24] Internet-Draft Algorithmic Integrity and Finality November 2026 4. Mode binding. Receipts carry the mode and generation, and the enforcement point evaluates the current mode immediately before authorizing release. An operation that began under normal mode and reaches the enforcement point after activation must satisfy the new policy. A narrowly scoped exception for specified previously authorized operations must be authenticated, bounded and verifiable, and a generic assertion that the operation started earlier does not bypass the new policy. 5. Revocation. Revocation during an active period is propagated within a bounded interval that the deployment must state. 6. Emergency modes. An emergency protected mode may impose stronger validation, restrict operations, shorten authorization lifetimes or require additional approval, and it is not automatically a way to switch off ordinary requirements. Any reduction of requirements must be explicitly authorized and constrained. An emergency recovery mode handles unavailable policy infrastructure by rejecting protected outputs until valid state returns, permitting only defined low-risk operations, permitting designated public- service operations under separate authorization, restricting operation to authenticated cached snapshots within a bounded interval, or allowing an authenticated administrative recovery procedure. Enforcement point A protected state registry commits transitions. Local mode-state verification runs at each node, and the sink-local check runs at each enforcement point. Failure handling Rollback control: a software rollback does not restore an older policy generation, and rollback is allowed only where it is authorized, current policy state is preserved, the restored configuration is approved and finality requirements are met. The embodiment also covers recovery from corrupted policy state, transition replay, unauthorized deactivation and split-brain policy state. In a strongly consistent design release may be blocked until the authoritative state is resolved, and in a bounded-consistency design release continues only inside stated freshness and transition bounds. Regional and jurisdictional scope, mobile, desktop, server and edge implementations, legacy integration, and latency and availability considerations are addressed. Das Expires 21 May 2027 [Page 25] Internet-Draft Algorithmic Integrity and Finality November 2026 Limits Central activation is not instantaneous propagation to every disconnected node. The architecture must not claim universal agreement among disconnected nodes without a synchronization protocol. Cross-reference Supports EP-7. Uses Embodiment 2 for policy epochs and Embodiment 9 for degraded-mode recovery. Section 8 discusses a possible relationship to an incident and crisis protocol. 6.6. Embodiment 5: Execution-to-Output Correspondence Source Sections 5.1 to 5.35, from PDF page 140 onward. Purpose To connect independent observation of actual algorithmic execution to a canonical pending operation, using cross-committed evidence and output-specific verification, so that the final ranked response corresponds to the validated execution path. What it teaches Approval of a configuration does not authorize an arbitrary output, and an output-specific cryptographic commitment does not by itself prove the asserted execution history. The identity fingerprint (what was activated), the execution fingerprint (what actually ran) and the output commitment (which output this evidence binds to) are different evidence objects and none stands in for another. How it works 1. A protected execution instance is established and actual execution is observed. 2. Execution evidence is built, and output-producing stages are associated with it. 3. Stage evidence is cross-committed: each stage commitment hashes the execution instance, stage identifier, previous stage commitment, observed input and output commitments, transformation identifier and observation generation, so that the stage records are tamper-evident in sequence. 4. The pending output is observed, and a canonical operation is reconstructed from what is actually pending. Das Expires 21 May 2027 [Page 26] Internet-Draft Algorithmic Integrity and Finality November 2026 5. An execution-to-output binding is constructed, checked for consistency across independent observations, validated, and recorded in a signed receipt with a separately scoped enablement. 6. The enforcement point reconstructs the actual pending operation and verifies the receipt, enablement and local state. Enforcement point The output boundary, with an observer outside the workload. Multi-stage and distributed execution, mobile and desktop output binding and other effectuation boundaries are covered. Failure handling Output substitution after validation: the substituted output produces a different reconstructed commitment and is rejected, without reliance on the validator's earlier statement. Mismatched execution attribution: an attacker associates a response from one execution with the fingerprint of another approved execution, and the system checks the protected execution-instance association, stage lineage and output observation, because a request identifier that an untrusted component can reassign is insufficient. Post- validation transformation such as formatting, compression, localization or reranking: where it changes the authorized operation, renewed validation, an authenticated transformation record or a post-transformation commitment is required, and the enforcement point verifies the post-transformation operation. Limits The security of a hash chain depends on the trustworthiness of the observations and of the stage-recording mechanism. A hash chain alone does not prove correct computation or authorized output generation. Downstream substitution or payload mutation invalidates an authorization that was granted earlier only. Cross-reference Supports EP-4 and EP-5. Builds on Embodiment 1, is extended to distributed pipelines by Embodiment 7, and to the device by Embodiment 8. The Zenodo record [ZENODO-24300261] describes an independent validator that links the verified state to a specific pending output. 6.7. Embodiment 6: Alternative-Route and Legacy-Path Coverage Source Sections 6.1 to 6.43, from PDF page 161 onward. Das Expires 21 May 2027 [Page 27] Internet-Draft Algorithmic Integrity and Finality November 2026 Purpose To find and qualify the multiple effect-capable routes, including legacy, cached, experimental and fallback paths, so that protected finality is not bypassed by delivering the same output another way. What it teaches Validation of an output through one delivery path does not authorize the same or a modified output through every other path. A defined enforcement perimeter separates known covered paths, known uncovered paths and unknown or unverified paths, and a deployment must not infer universal anti-bypass protection because its primary route is protected. The strength of any guarantee depends on how completely routes are discovered and controlled. How it works 1. Each route receives a protected identity. Routes are discovered and entered in a protected effectuation-path registry that records which are covered. 2. Reachability-based coverage check. The delivery infrastructure is modeled as a directed graph: nodes are output-producing services, caches, brokers, gateways and endpoints, edges are possible transfer paths. A covered path must pass a verified enforcement boundary before reaching its protected target. A graph built only from incomplete static documentation is not proof of complete runtime coverage. 3. Route state machine. A route moves from discovered to classification pending and coverage pending, then to verified covered, unverified or known uncovered, and can later be suspended, revoked or superseded. A route is not treated as covered merely because an enforcement adapter was installed. 4. Route-specific validation. The receipt and enablement are bound to the actual route and destination scope, so the same output cannot be released through a different route on the strength of an earlier receipt. Specific routes are treated in turn. Legacy endpoints are protected. Direct-origin bypass is handled by accepting protected-output requests only over authenticated controlled paths, by placing an enforcement point in the origin itself, or by an independent network boundary between origin and destinations. A cache has two checks: generation validation (this object came from a validated execution) and release validation (this delivery is authorized under current policy, destination, route and finality state). An old generation receipt does not authorize Das Expires 21 May 2027 [Page 28] Internet-Draft Algorithmic Integrity and Finality November 2026 unlimited future releases. After a revocation a protected cache may keep the object stored and still refuse to release it. Experimental and A/B routes are identified by the protected observer, because a route label from a compromised experiment controller is insufficient, and output from an unauthorized experimental route is rejected. A fallback route does not inherit the primary route's authorization, and the fallback policy states permitted routes, required output class, validation profile, maximum duration, mode generation and recovery conditions. Streaming, load balancing, service migration and multi-region delivery are covered. Enforcement point One enforcement point per covered route, with route-specific sink- local reconstruction and replay protection. Failure handling Unauthorized route creation, route-registry rollback, split-brain route configuration and route-bound replay are addressed, and failures are classified. Limits The embodiment lists its own limitations. A route registry cannot give complete anti-bypass protection unless the relevant paths are discovered and the enforcement boundaries are really controlled. A compromised infrastructure administrator may create routes outside the perimeter unless infrastructure-level restrictions apply. A gateway cannot stop delivery through a separate uncontrolled direct-origin connection. A cache cannot revoke copies already delivered to external recipients. A receipt cannot prevent bypass where a path never invokes the enforcement point. Non-bypassability holds only inside the defined perimeter. Cross-reference Supports EP-8. Complements Embodiment 3. The Zenodo record [ZENODO-24300261] describes enforcement at various delivery points. 6.8. Embodiment 7: Distributed Execution Reconstruction Source Sections 7.1 to 7.50, from PDF page 184 onward. Purpose To reconstruct execution lineage across distributed algorithmic components and evidence domains, so that the real contributing pipeline is tied to a pending output. Das Expires 21 May 2027 [Page 29] Internet-Draft Algorithmic Integrity and Finality November 2026 What it teaches A collection of distributed trace records does not prove the actual execution history of an output. Distributed computation does not become authority to produce an external effect because its participating services report that they finished. A real pipeline may include models, features, ranking, caches and delivery services in different administrative domains, and an unauthorized reranking step could sit anywhere in it. How it works 1. Each distributed execution has an identity, and each stage is measured and observed, producing a stage execution evidence record. 2. Cross-domain handoff. A source stage's statement that it sent a message does not show that the destination accepted it, and a destination's statement that it received input does not show the source made it. A handoff record therefore ties the source output commitment to the destination input commitment, with the transfer context, evidence references, destination acceptance, handoff generation and policy epoch. Commitments are equal for an unchanged object, and an authorized transformation adds an authenticated transformation relationship. Verification is two-sided. 3. A causal lineage graph is built from protected evidence. The embodiment distinguishes causality from correlation. 4. Lineage completeness is checked against a required-stage profile. The procedure returns incomplete when a required stage, a required handoff or a contributing branch is missing or unresolved, invalid when an unauthorized stage is present or the final output provenance fails, and complete only for the profile. 5. The result is bound to the actual pending output, validated, recorded in a receipt with a scoped enablement, and verified at the enforcement point. Parallel branches, fan-in and merge verification, retries and duplicate execution, speculative execution, cached intermediate results, cross-administrative-domain execution, accelerator and GPU execution, and asynchronous messaging are each addressed. Partial or delayed evidence keeps the pending output non-effective while evidence is collected, through the states execution observed, lineage collection pending, handoff verification pending, completeness verification pending, output binding pending, validated non-effective, finality pending and Das Expires 21 May 2027 [Page 30] Internet-Draft Algorithmic Integrity and Finality November 2026 effectuation authorized. Where mandatory evidence is missing, the output is rejected or an explicitly authorized alternative validation profile is used, and incomplete evidence is never silently classified as complete. Enforcement point Observation at each stage, a distributed validator, and an enforcement point at the final release boundary. Failure handling Attacks covered include stage omission (for example omitting an unauthorized reranking stage from the reported lineage), forged handoff, cross-request evidence substitution, branch substitution, and replay or rollback. Limits Completeness is defined relative to a protected profile and the covered observation perimeter. The system does not claim to reconstruct computation outside that perimeter, and it cannot guarantee detection of every omitted stage outside its observation perimeter. Cross-reference Supports EP-9. Extends Embodiment 5. The Zenodo record [ZENODO-24300261] describes validation tied to a specific pending output. 6.9. Embodiment 8: Device-Side Display Finality Source Sections 8.1 to 8.51, from PDF page 212 onward. Purpose To control selected mobile, desktop and visible-screen presentation boundaries by reconstructing the pending rendering operation and checking authorization before display submission. What it teaches Successful validation of an upstream output does not authorize every device-side presentation derived from it. Receipt of valid algorithmic information is not authority to present an unverified or materially modified result. Three representations are distinguished: the upstream output that was validated, the device presentation representation (content, order, layout) prepared by the device, and the actual pending operation submitted to the protected rendering boundary. They may legitimately differ, so byte-for-byte identity is not required. An explicitly defined authorized transformation relationship is verified instead. Das Expires 21 May 2027 [Page 31] Internet-Draft Algorithmic Integrity and Finality November 2026 How it works 1. The device-side component observes the actual received output and any local transformation, and records the transformation. 2. A presentation descriptor is built covering content identity and ordering, layout and visibility properties, and canonicalized into a presentation commitment. 3. The upstream-to-presentation binding is validated on the device, and an authenticated presentation receipt and separately scoped presentation enablement are produced. 4. An independent device-side enforcement point reconstructs the pending presentation operation and verifies the receipt and enablement against its protected local state before display submission. Four visibility-assurance levels are illustrative and not universally ordered: level A verifies content and ordering prepared by the application, level B verifies selected operations submitted to a protected rendering boundary, level C verifies selected composed properties at an operating-system compositor or equivalent, and level D uses hardware-assisted display-submission checks. Implementations include mobile and desktop applications, browser-based presentation, operating-system compositors and GPU and display submission. Cached device-side content, offline operation (a protected local policy snapshot with a maximum offline duration, a freshness requirement and a revocation- consistency profile), multi-window and overlay conditions, accessibility transformations and localization or responsive layout are treated, and accessibility and localization transformations must remain inside the authorized relationship. Enforcement point A protected compositor, rendering boundary or equivalent privileged display-path component on the device. Failure handling Presentation substitution, unauthorized reordering, post- validation rendering modification, presentation replay, application-to-surface mismatch and concurrent presentation operations are covered, with a presentation state machine and failure classification. If the offline freshness or revocation condition cannot be established, the device may deny the covered presentation or use an explicitly authorized restricted mode. Das Expires 21 May 2027 [Page 32] Internet-Draft Algorithmic Integrity and Finality November 2026 Limits An ordinary monitor or unprivileged application cannot enforce all display paths. A validator that sees only an application's submitted content cannot guarantee that another uncontrolled window will not cover it. A protected rendering descriptor is not identical to the light a display emits, which depends on hardware, brightness, overlays, scaling and external capture, so complete physical-pixel verification must not be claimed. An asynchronously updated local policy cannot give immediate knowledge of remote revocation. Cross-reference Supports EP-10. Complements Embodiment 3. The Zenodo record [ZENODO-24300261] describes device-side display enforcement. 6.10. Embodiment 9: Degraded-Mode Recovery Source Sections 9.1 onward, from PDF page 236 onward. Purpose To keep a covered system useful when a validator, measurement source, registry or network path is unavailable, without letting the fallback quietly become a way to remove the protection. What it teaches Availability pressure is itself an attack surface. A fallback is a separately authorized operating profile with its own scope and duration, and it does not inherit the primary profile's authorization. Missing evidence is never converted into proof of compliance. Some finality requirements are designated non- reducible: they stay in force in every fallback profile. How it works 1. Assurance-preserving fallback profiles are defined in advance by the policy authority. Each states the permitted routes, the output classes it covers, the validation it still requires, its maximum duration, its mode generation and its recovery conditions. 2. Entry into a fallback is an authenticated transition and is recorded. The enforcement point evaluates the active profile before release, in the same way as it evaluates the operating mode in Embodiment 4. Das Expires 21 May 2027 [Page 33] Internet-Draft Algorithmic Integrity and Finality November 2026 3. Non-reducible requirements are checked at the enforcement point regardless of profile. A fallback may narrow what is released (for example to a restricted class of outputs) and may not remove those requirements. 4. Recovery returns to the primary profile only through a defined procedure, after which backlog outputs produced under the fallback are re-evaluated rather than assumed valid. 5. Oscillation protection limits repeated switching between primary and fallback profiles, because an attacker or a flapping dependency could otherwise use the switching itself to find a weaker state. Enforcement point The same enforcement points as Embodiments 3 and 8, extended with a profile check and a record of which profile authorized each release. Failure handling If no valid fallback profile applies, protected outputs are withheld and only operations that are separately authorized as low-risk or public-service continue. A fallback that has exceeded its duration lapses and does not renew itself. Limits A fallback lowers what can be proven for the period it covers. The record of that period is therefore marked with the reduced assurance and is not presented as equivalent to full validation. Cross-reference Supports EP-11. Uses Embodiment 4 for authenticated transitions and Embodiment 6 for fallback routes. The Zenodo record [ZENODO-24300261] describes withholding an output or routing it through an authorized fallback, which corresponds to this embodiment. 6.11. Embodiment 10: Continuous Algorithm Drift Detection Source Sections 10.1 onward, from PDF page 272 onward. Purpose To keep the verified state tied to the algorithm that is running now, so that a change after approval, whether deliberate or accidental, is noticed and its effect on pending and future outputs is contained. Das Expires 21 May 2027 [Page 34] Internet-Draft Algorithmic Integrity and Finality November 2026 What it teaches Approval at one moment does not cover later changes. Detection and enforcement are different functions: noticing that behavior has drifted does not by itself stop an output, and a decision to stop an output needs the enforcement path of the earlier embodiments. Drift is a trigger for verification and is not proof of manipulation, since legitimate updates also change behavior. How it works 1. Each activation of a configuration has an activation generation. A change in any registered dependency (model artifact, feature pipeline, ranking parameter, policy setting) creates a new generation. 2. Dependency invalidation is selective. A change invalidates the approvals, receipts and pending outputs that depended on the changed component, and leaves unrelated ones alone, so that protection does not require stopping everything for every change. 3. Continuous measurement compares the active fingerprint with the approved reference over time, and records both identity changes and changes in execution evidence. 4. Behavioral monitoring may add signals, for example unexpected shifts in distribution of outputs, and these signals start verification. They do not by themselves establish that anything unauthorized happened. 5. The result is passed to the approval and finality processes, which decide whether to suspend an approval, require revalidation or withhold output. Enforcement point Detection runs beside the measurement source. Enforcement remains at the release boundaries of Embodiments 3, 6 and 8. Failure handling Two attacks are treated explicitly. Baseline poisoning: an adversary slowly shifts the reference so that later changes look normal, which is countered by keeping the approved reference under protected authorization and not letting monitoring data rewrite it. Telemetry manipulation: an adversary feeds false monitoring data, which is countered by taking measurements from a source outside the monitored algorithm's control and by cross-checking independent observations. Das Expires 21 May 2027 [Page 35] Internet-Draft Algorithmic Integrity and Finality November 2026 Limits Drift detection cannot establish intent, fairness or neutrality. Behavioral statistics can give false positives and false negatives, and the embodiment treats them as inputs to a decision and never as the decision. Cross-reference Supports EP-12. Depends on Embodiment 1 and acts through Embodiments 2 and 3. Section 8 discusses a possible relationship to logging of changes in recommender systems before an election. 6.12. Embodiment 11: Confidential Verification Source Sections 11.1 onward, from PDF page 310 onward. Purpose To let a verifier gain assurance that the approved algorithm ran and produced the pending output, without being given the proprietary algorithm, personal data or infrastructure details. What it teaches Three kinds of confidentiality are in play: proprietary algorithm confidentiality (C1), personal-data confidentiality (C2) and infrastructure confidentiality (C3). Verification can be designed so that each is preserved. A proof is not the same as an observation, a binding or finality. A zero-knowledge or commitment-based proof can show a statement about hidden values, and it still has to be tied to actual execution evidence, to the actual pending output, and to an enforcement point that acts on it. How it works 1. The prover side commits to the relevant measurements and evidence using cryptographic commitments, and discloses only what the verifier's profile requires. 2. Where the profile allows, an optional zero-knowledge proof shows that the committed configuration matches an approved reference, or that required stages were present, without revealing the underlying values. 3. The proof is bound to the execution evidence by tying it to the same commitments used in Embodiment 5, so that a proof about some other run cannot be reused. 4. The proof is bound to the actual pending output, and the enforcement point checks that binding in the usual way. Das Expires 21 May 2027 [Page 36] Internet-Draft Algorithmic Integrity and Finality November 2026 5. Selective disclosure to an authorized auditor is allowed under separate authorization, with the disclosure itself recorded. Enforcement point The verifier of the proof sits with the protected validator. The release decision still sits with the enforcement point. Failure handling A proof that is valid but unbound is rejected. Failure of the proof system, or a proof outside its stated assumptions, is handled as missing evidence and never as success. Limits Confidentiality has a cost in what an outside party can independently check. A proof shows only what its statement says. The embodiment does not claim that a zero-knowledge proof alone establishes that execution occurred. Cross-reference Supports EP-13. Uses Embodiments 1 and 5. The Zenodo record [ZENODO-24300261] describes verification without opening proprietary models. 6.13. Embodiment 12: Atomic Finality Consumption Source Sections 12.1 onward, from PDF page 347 onward. Purpose To make sure that a single-use authorization produces at most one external effect, even with concurrent requests, retries, crashes and partial failures at the enforcement point. What it teaches Validation, authorization, commitment and the external effect are separate states, and the transitions between them need protection. A valid authorization that can be used twice is a weakness. The embodiment states clearly what can and cannot be guaranteed, and it does not make a universal exactly-once claim. How it works 1. The operation moves through validated, authorized, committed and effectuated states, each recorded in protected state at the enforcement point. 2. Claim and reserve is atomic: the enforcement point claims the authorization for one specific operation in a single protected step, so two concurrent attempts cannot both succeed. Das Expires 21 May 2027 [Page 37] Internet-Draft Algorithmic Integrity and Finality November 2026 3. Fencing tokens or generations stop a stale or restarted component from acting on a claim that has since been taken over. 4. The release is performed, and the final state is recorded. 5. If the outcome is unknown, for example after a crash between release and recording, the operation is placed in an outcome- unknown state. It is not retried blindly, because a retry could cause a second effect. It is resolved by a protected recovery procedure, by checking the downstream state where that is possible, or by an explicit decision. The embodiment separates three guarantees. At-most-once means the effect happens no more than once, and it may happen zero times. Effectively-once means repeated attempts have the same net effect, which needs an idempotent target or deduplication. Transactional exactly-once needs the target to take part in a transaction with the enforcement point. The right choice depends on what the target can support, and the architecture does not claim any of them where the target cannot. Enforcement point The release boundary, with sink-local protected state that cannot be modified by the workload. Failure handling Replay of receipts, duplicate and concurrent claims, expired reservations, crash recovery and split-brain between enforcement- point replicas are covered. An expired reservation does not return an authorization to the unused state unless the outcome is known. Limits Where the network or the target gives no commit guarantee, delivery cannot be made atomic with authorization consumption. The architecture then offers at-most-once or effectively-once, and says so. Cross-reference Supports EP-3, EP-6 and EP-4. Hardens Embodiment 3. The Zenodo record [ZENODO-24300261] describes an authorization that permits release only after the enforcement point has checked the pending operation. Das Expires 21 May 2027 [Page 38] Internet-Draft Algorithmic Integrity and Finality November 2026 7. Cross-Reference to the Public Zenodo Record The Zenodo record [ZENODO-24300261] is a public summary of the same disclosure. Its text was reviewed in preparing this document. The table maps its themes to the embodiments here. The mapping is the author's reading and may contain mistakes, so readers should check it against the record itself. Das Expires 21 May 2027 [Page 39] Internet-Draft Algorithmic Integrity and Finality November 2026 +===========================================+=============+=========+ | Theme in the Zenodo record | Embodiments | Section | | | | here | +===========================================+=============+=========+ | Verify the algorithmic process | 1A | 6.1 | | and not the opinion expressed; | | | | direct publication kept | | | | separate from amplification | | | +-------------------------------------------+-------------+---------+ | Protected identification of | 1 | 6.2 | | the running algorithm and | | | | fingerprint of its active | | | | configuration | | | +-------------------------------------------+-------------+---------+ | Comparison with an approved | 2, 4 | 6.3, | | reference for the protected | | 6.5 | | period | | | +-------------------------------------------+-------------+---------+ | Independent validator linking | 5, 7 | 6.6, | | verified state to a specific | | 6.8 | | pending output; signed receipt | | | +-------------------------------------------+-------------+---------+ | Separately scoped | 3, 12 | 6.4, | | authorization; enforcement | | 6.13 | | point checks the actual | | | | pending operation | | | +-------------------------------------------+-------------+---------+ | Enforcement at server, | 3, 6, 8 | 6.4, | | gateway, edge and device | | 6.7, | | | | 6.9 | +-------------------------------------------+-------------+---------+ | Withholding or authorized | 9 | 6.10 | | fallback when the algorithm | | | | differs | | | +-------------------------------------------+-------------+---------+ | Changes during a protected | 10 | 6.11 | | period | | | +-------------------------------------------+-------------+---------+ | Verification without opening | 11 | 6.12 | | proprietary models | | | +-------------------------------------------+-------------+---------+ Table 3: Mapping of Zenodo Themes to Embodiments Das Expires 21 May 2027 [Page 40] Internet-Draft Algorithmic Integrity and Finality November 2026 8. Relationship to Public Policy Work This section is an AI-assisted reading and may contain mistakes. It states relationships of subject matter only. It does not claim that any law or policy requires this architecture, that any deployment complies with any law, or that any body endorses the proposal. Public policy work on protecting elections from manipulation through online systems has several recurring themes: transparency about recommender systems, logging of changes to such systems before a vote, testing of systems before and during an election period, and incident and crisis coordination. The Zenodo record [ZENODO-24300261] places the disclosure against public policy work of this kind, and the source documents cited there should be read directly. * Recommender-system transparency: the approved-reference register (Embodiment 2) and the identity fingerprint (Embodiment 1) are technical records of which configuration was active and when, which a transparency process could use as evidence. * Change logging before a vote: Embodiments 1, 2 and 10 produce a protected record of changes to active configurations. Whether any such record satisfies any logging obligation is a legal question that this document does not address. * Testing and red-teaming: Embodiments 1A and 10 distinguish what technical testing can show (integrity, consistent application of registered predicates) from what it cannot (neutrality or fairness), which affects how test results should be described. * Incident and crisis coordination: Embodiment 4 gives authenticated, bounded transitions into protected modes, and Embodiment 9 gives bounded fallback, which a crisis process could reference. The architecture is deliberately limited to the delivery of algorithmic outputs. It does not decide what may be said, and Embodiment 1A is included to keep it that way. 9. Security Considerations Each embodiment in Section 6 lists its own failure handling and limits, and those are part of the security analysis. The following points apply across embodiments. Das Expires 21 May 2027 [Page 41] Internet-Draft Algorithmic Integrity and Finality November 2026 * The architecture is only as strong as its measurement source and its protected domains. If the measurement source is controlled by the measured algorithm, fingerprints say nothing. * Coverage is bounded by the enforcement perimeter. An output delivered by a path outside the perimeter is not protected (Embodiment 6). * Device-side enforcement cannot prove what a user physically sees (Embodiment 8). * Verification of the process is not a finding that the algorithm is fair, neutral or lawful (Embodiments 1, 1A and 10). * The enforcement system can itself be misused to suppress speech or can be wrongly bypassed. Embodiment 1A addresses both and requires that counter-verification cannot create release authority. * Delivery and authorization consumption may not be atomic. Implementations must state which guarantee they give (Embodiment 12). * Privacy: evidence about execution can reveal information about users and infrastructure. Embodiment 11 describes ways to limit it. Deployments should minimize personal data in evidence records. * Availability: stronger checks add latency and failure modes. This document makes no performance claim. 10. IANA Considerations This document has no IANA actions. 11. Notices, Intellectual Property and Non-Endorsement The subject matter of this document is the subject of patent applications by the author, including the application identified in the front-matter note and PCT/IB2026/055615. This document is provided as an individual informational contribution. The author has not made any statement on licensing in this document. The Zenodo record [ZENODO-24300261] is published under CC BY-NC 4.0, which is separate from the terms that apply to this Internet-Draft. Das Expires 21 May 2027 [Page 42] Internet-Draft Algorithmic Integrity and Finality November 2026 This document names no operator, product or country as a source of manipulation. Any reference to public policy work is a reference to subject matter and is not an endorsement or validation by any body, and not a statement of compliance, infringement or non-infringement. This document was prepared with AI assistance and may contain errors. The original patent filings and public disclosures are the authoritative sources. 12. References 12.1. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, . [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, . 12.2. Informative References [I-D.das-social-media-algorithmic-accountability] Das, S., "Social Media Algorithmic Accountability", Work in Progress, Internet-Draft, draft-das-social-media- algorithmic-accountability-00, 2026, . [PCT-055615] Das, S., "WIPO PATENTSCOPE: PCT/IB2026/055615 (filed 4 June 2026)", 4 June 2026, . [TDC-12542] Das, S., "Technical Disclosure Commons: PCT/IB2026/055615 (Mothership application)", 2026, . [ZENODO-23659488] Das, S., "Zenodo record 23659488 (source specification of the earlier social media accountability draft)", 2026, . [ZENODO-24300261] Das, S., "Algorithmic Manipulation by Social Media and the Protection of Elections: AI Safety and AI Governance for Das Expires 21 May 2027 [Page 43] Internet-Draft Algorithmic Integrity and Finality November 2026 the European Democracy Shield and Other Democracies", DOI 10.5281/zenodo.24300261, 17 November 2026, . Acknowledgments The author prepared this draft with AI assistance and is responsible for its content. Author's Address Sangam Das Independent Inventor Balasore Phone: +91-9861363532 Email: info@sangamdas.com URI: https://www.linkedin.com/in/sangam-das-41b4a148/ Das Expires 21 May 2027 [Page 44]