| Internet-Draft | Algorithmic Integrity and Finality | November 2026 |
| Das | Expires 21 May 2027 | [Page] |
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 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.¶
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.¶
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 21 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.¶
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.¶
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.¶
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.¶
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.¶
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.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
The 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 document | Source term | Meaning |
|---|---|---|
| Algorithm fingerprint | Algorithmic Logic Fingerprint (ALF) | A representation 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 fingerprint | ALF-E | Protected 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. |
| Approved reference | Approved ALF Reference | A registered 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 Domain | An isolated, independent component that evaluates evidence and issues receipts. |
| Validation receipt | Validation Receipt | A cryptographically authenticated record of a validation determination. |
| Non-bearer authorization | Non-Bearer Finality Enablement | A scoped 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 Operation | The 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 Coverage | The extent to which every path by which an output can reach its target passes an enforcement point. |
| Counter-verification | Independent Counter-Verification | An independent re-examination of a disputed enforcement decision. |
The architecture considers the following adversaries and faults. It does not assume they are all present in any given deployment.¶
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.¶
All embodiments extend the same seven-step sequence.¶
+-----------+ +-------------+ +--------------+
| 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
The architecture keeps three kinds of evidence apart because they prove different things.¶
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.¶
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 the sole control of the algorithm it measures, and a signed artifact hash alone MUST NOT be treated as proof that the artifact executed. | 1, 5, 7, 10 |
| EP-2 | An approved reference MUST be registered with authenticated lifecycle state, a policy epoch and a freshness bound. Approval of a configuration MUST NOT be treated as approval of every output. | 2, 4 |
| EP-3 | A covered output MUST be held in a non-effective state until the enforcement point has verified the release conditions. | 3, 12 |
| EP-4 | The enforcement point MUST reconstruct or independently obtain the actual pending operation and MUST NOT rely only on the workload's statement that it is the one validated. | 3, 5, 8, 12 |
| EP-5 | A receipt MUST be authenticated and bound to the pending operation commitment, the policy epoch and the operating mode. | 3, 4, 5 |
| EP-6 | Authorization MUST be separately scoped, non-bearer, and consumed under sink-local protected state with replay and concurrency protection. | 3, 12 |
| EP-7 | Operating-mode and policy transitions MUST use authenticated, monotonic generations, and the enforcement point MUST evaluate the current mode immediately before release. | 4 |
| EP-8 | Every route by which a covered output can become effective MUST be inside a declared enforcement perimeter, and uncovered and unverified routes MUST be reported as such. | 6 |
| EP-9 | Validation of a distributed pipeline MUST check lineage completeness against a required-stage profile, and missing or contradictory evidence MUST NOT be treated as complete. | 7 |
| EP-10 | Where device-side presentation is covered, the device-side enforcement point MUST verify the actual presentation operation under a stated visibility-assurance level. | 8 |
| EP-11 | A fallback MUST be separately authorized, MUST preserve designated non-reducible requirements, and MUST NOT convert missing evidence into proof of compliance. | 9 |
| EP-12 | Changes to the active configuration MUST 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. | 10 |
| EP-13 | Confidential verification MUST bind any proof to the actual execution evidence and to the actual pending output. | 11 |
| EP-14 | Enforcement scope MUST separate direct publication from algorithmic amplification, and a failure in one MUST NOT be inherited by the other. | 1A |
| EP-15 | A disputed enforcement decision MUST be open to independent counter-verification that cannot itself create release authority. | 1A |
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.¶
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.¶
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 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.¶
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 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.¶
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 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.¶
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 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.¶
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.¶
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.¶
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.¶
| Theme in the Zenodo record | Embodiments | Section here |
|---|---|---|
| Verify the algorithmic process and not the opinion expressed; direct publication kept separate from amplification | 1A | 6.1 |
| Protected identification of the running algorithm and fingerprint of its active configuration | 1 | 6.2 |
| Comparison with an approved reference for the protected period | 2, 4 | 6.3, 6.5 |
| Independent validator linking verified state to a specific pending output; signed receipt | 5, 7 | 6.6, 6.8 |
| Separately scoped authorization; enforcement point checks the actual pending operation | 3, 12 | 6.4, 6.13 |
| Enforcement at server, gateway, edge and device | 3, 6, 8 | 6.4, 6.7, 6.9 |
| Withholding or authorized fallback when the algorithm differs | 9 | 6.10 |
| Changes during a protected period | 10 | 6.11 |
| Verification without opening proprietary models | 11 | 6.12 |
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.¶
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.¶
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.¶
This document has no IANA actions.¶
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.¶
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.¶
The author prepared this draft with AI assistance and is responsible for its content.¶