Datasets:

apol
/

agency-transfer-policy-atlas

implementation_id string	control_family_id string	legacy_ids list	intervention string	layer string	instrument_class string	operative_instrument_legal_basis string	implementer string	jurisdiction string	legal_force string	legal_status string	legal_scope_correction string	operational_maturity string	mechanism_targeted string	causal_chain_point string	agency_domain string	primary_evidence_basis string	mechanism_evidence_tier string	intervention_effect_evidence string	claim_class string	expected_effect string	main_failure_evasion string	rights_trade_offs string	metric string	decision_tier string	decision_rationale string	horizon_date string	part_1_linkage string	part_2_linkage string	last_reviewed string	notes string	family_name string	mechanism_ids list	context_entity_ids list	research_gap_ids list	claim_ids list	effect_claim_ids list	legal_claim_ids list	mechanism_claim_ids list	source_ids list	policy_package_ids list	effect_claim_checked bool	legal_claim_checked bool	mechanism_claim_checked bool	effect_claim_reviewed bool	priority_effect_review_outcomes list	priority_effect_publication_actions list	publication_claim_class string	publication_status string
I-001	CF-001	[ "P4-001" ]	Root-purpose provenance across multi-agent handoffs	Agent / deployment	Procurement	Agent protocol, audit standard, procurement clause	Agent providers, deployers, auditors, public procurers	Cross-jurisdictional	Research	Research proposal	Research proposal	Pilot	Bind every delegated subtask, tool call, generated artifact, and terminal action to a signed root-purpose record, with explicit scope changes and controller identity.	Capability → controller → delegated vector	Executive; decisional; collective / institutional	Project — confirmatory mechanism	Project-C	None — mechanism or proposal only	Established — project mechanism	Makes harmful composition inspectable and prevents local compliance from substituting for end-to-end purpose review.	False or vague root purposes; off-ledger work; laundering through external tools; excessive retention of sensitive prompts.	Minimize retained content, separate oversight metadata from message content, and provide due process for contested classifications.	Share of consequential actions traceable to an unchanged root purpose; undocumented scope-change rate; reviewer agreement.	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	Confirmatory mechanism	Contextual capability link	2026-08-11	Canonical project URL recorded in v0.3 Sources; policy effect remains untested.	Root-purpose provenance across multi-agent handoffs	[ "M-01", "M-06", "M-08", "M-14", "M-16" ]	[]	[ "G-03" ]	[ "CLM-0001", "CLM-0002" ]	[ "CLM-0002" ]	[]	[ "CLM-0001" ]	[ "SRC-001" ]	[ "P02" ]	false	false	false	false	[]	[]	Provisional — project mechanism not claim-checked	research_preview
I-002	CF-002	[ "P4-002" ]	Workflow-level global guard before consequential completion	Agent / deployment	Law / regulation	Deployment rule, safety case control, high-impact action gate	Agent providers, platforms, campaigns, public deployers	Cross-jurisdictional	Research	Research proposal	Research proposal	Pilot	Evaluate the cumulative plan, intermediate artifacts, recipient, scale, target, and foreseeable downstream use immediately before a consequential action is executed.	Vector → target → consequential action	Executive; decisional	Project — confirmatory mechanism	Project-C	None — mechanism or proposal only	Established — project mechanism	Restores end-to-end review at the point where reversible assistance becomes real-world harm.	Action splitting below thresholds; external execution; warning fatigue; evaluator-model collusion or shared blind spots.	Use risk-based triggers, clear appeal, minimal delay, and protected space for lawful political participation.	Blocked harmful completions; unnecessary interruption rate; bypass cost; latency; successful appeal rate.	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	Confirmatory mechanism	Contextual capability link	2026-08-11	Should be evaluated against intentionally fragmented benign-looking workflows.	Workflow-level global guard before consequential completion	[ "M-01", "M-06", "M-16" ]	[]	[ "G-03" ]	[ "CLM-0003", "CLM-0004" ]	[ "CLM-0004" ]	[]	[ "CLM-0003" ]	[ "SRC-001" ]	[ "P02" ]	false	false	false	false	[]	[]	Provisional — project mechanism not claim-checked	research_preview
I-003	CF-003	[ "P4-003" ]	Machine-readable authorization boundaries and least privilege for agents	Agent / deployment	Procurement	Technical standard, product-safety duty, procurement clause	Standards bodies, agent providers, deployers, regulators	Cross-jurisdictional	Research	Research proposal	Research proposal with standards precedents	Pilot	Express purpose, permitted tools, targets, data, spend, duration, audience, and escalation conditions as enforceable grants; require renewed confirmation for high-impact deviations.	Controller → vector	Executive; decisional; institutional	Project — exploratory	Project-X	None — mechanism or proposal only	Plausible hypothesis	Preserves meaningful delegation while preventing quiet expansion of an agent's authority.	Overbroad default grants; consent fatigue; compromised credentials; delegation to ungoverned agents.	Avoid identity requirements for low-risk use and make authorization legible, revocable, and accessible.	Unauthorized-action rate; median permission scope; revocation time; percentage of consequential actions explicitly confirmed.	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	Exploratory link	Contextual capability link	2026-08-11	Combine with root-purpose provenance; permissions without purpose can still authorize harmful composition.	Agent authorization, least privilege and action control	[ "M-01", "M-06", "M-14" ]	[ "C-06" ]	[ "G-03" ]	[ "CLM-0005", "CLM-0006" ]	[ "CLM-0006" ]	[]	[ "CLM-0005" ]	[ "SRC-002" ]	[ "P02" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-004	CF-004	[ "P4-004" ]	Tamper-evident workflow traces and election evidence preservation	Cross-layer	Coordination / incident response	Recordkeeping duty, incident protocol, audit standard	Labs, agent providers, platforms, electoral authorities	Cross-jurisdictional	Research	Research proposal	Research proposal with DSA and incident-reporting precedents	Research	Record model/version, controller, root purpose, tool permissions, consequential actions, delivery endpoints, and intervention events in privacy-preserving tamper-evident logs with legal hold and researcher-access paths.	Change → agency transfer → accountability	Collective / institutional; epistemic	Project — exploratory	Project-X	Indirect / mixed	Plausible hypothesis	Makes investigation, remedy, independent audit, and policy learning possible after fast-moving incidents.	Encrypted channels, offshore actors, log tampering, excessive data capture, selective retention.	Strict purpose limitation, access controls, retention limits, legal process, and protection of journalists and political association.	Evidence completeness; preservation latency; chain-of-custody failures; validated external access requests.	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	Exploratory link	No direct link	2026-08-11	Do not equate logging with indiscriminate surveillance.	Tamper-evident workflow traces and election evidence preservation	[ "M-01", "M-06", "M-11", "M-12", "M-13" ]	[]	[ "G-03", "G-08", "G-13" ]	[ "CLM-0007", "CLM-0008" ]	[ "CLM-0008" ]	[]	[ "CLM-0007" ]	[ "SRC-003" ]	[ "P02", "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-005	CF-005	[ "P4-005" ]	Longitudinal APE–MASK–manipulation evaluation suite	Model / provider	Research / evaluation	Independent TEVV protocol and pre-release assurance	Providers, AI safety institutes, accredited researchers	Cross-jurisdictional	Research	Research proposal	Core research proposal	Pilot	Run a staged evaluation from harmful-persuasion propensity and pressured deception to contextual efficacy, adaptive attacks, agent scaffolds, personalization, and 7-, 30-, and 90-day agency outcomes.	Capability → vector → target → change	Evaluative; decisional; relational	Normative proposal	E0	None — mechanism or proposal only	Plausible hypothesis	Separates propensity, deception, efficacy, durability, safeguard robustness, and deployment scale.	Benchmark gaming, contamination, narrow languages, evaluator bias, unsafe disclosure, and weak ecological validity.	Ethics review, participant protection, controlled access to dangerous artifacts, and publication of aggregate rather than operational misuse details.	Attempt, honesty, efficacy, persistence, reversal, robustness, and reach-adjusted risk with confidence intervals.	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	Scores must not be treated as validated democratic-harm thresholds until construct and outcome validity are established.	Longitudinal APE–MASK–manipulation evaluation suite	[ "M-04", "M-05", "M-06", "M-12" ]	[]	[ "G-02", "G-04", "G-06", "G-12" ]	[ "CLM-0009", "CLM-0010" ]	[ "CLM-0010" ]	[]	[ "CLM-0009" ]	[ "SRC-004" ]	[ "P03" ]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-006	CF-006	[ "P4-006" ]	Manipulation capability–reach threshold tied to release and deployment decisions	Cross-layer	Policy / organizational control	Delegated regulatory threshold, release gate, election-risk trigger	AI Office, electoral authorities, safety institutes, providers	EU pilot; internationally adaptable	Research	Research proposal	Core research proposal	Research	Combine validated manipulation capability, safeguard robustness, release reversibility, personalization, tool access, likely reach, and election timing into graduated obligations.	Capability → controller → vector → scale	Evaluative; decisional; collective / institutional	Normative proposal	E0	None — mechanism or proposal only	Open question	Connects research evidence to consequential policy action while scaling controls to actual risk and reversibility.	Threshold gaming, metric monoculture, offshore release, rapid derivative models, and over-restriction of open research.	Multiple independent measures, appeal, periodic recalibration, public methodology, and protected research access.	False-negative and false-positive designation rates; time to mitigation; residual reach-adjusted risk; circumvention.	Research / hold	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Research — 6–30 months	No direct link	Contextual capability link	2026-08-11	Central unanswered policy question of Part 4.	Manipulation capability-to-deployment decision logic	[ "M-02", "M-05", "M-06", "M-12", "M-14" ]	[]	[ "G-02", "G-12" ]	[ "CLM-0011", "CLM-0012" ]	[ "CLM-0012" ]	[]	[ "CLM-0011" ]	[ "SRC-004" ]	[]	false	false	true	false	[]	[]	Open question	research_preview
I-007	CF-007	[ "P4-007" ]	Election access-risk observatory for frontier and open-weight releases	Electoral operations	Research / evaluation	Public-interest observatory, secure evidence repository	Universities, AI safety institutes, electoral networks, repositories	International	Research	Research proposal	Research and preparedness proposal	Pilot	Track release dates, access modality, derivatives, safeguard state, language coverage, campaign tooling, election windows, and incident signals without publishing reproducible misuse pathways.	Capability → controller → vector → target	Collective / institutional; epistemic	Normative proposal	None	None — mechanism or proposal only	Plausible hypothesis	Identifies when a capability change should trigger tabletop exercises, platform contacts, monitoring, or independent tests.	Incomplete model coverage, hidden derivatives, jurisdictional gaps, stigmatizing open models, and unsafe detail leakage.	Publish risk indicators and methods, not targeting recipes, exploit prompts, or sensitive actor attribution without due process.	Release coverage; verification lag; election-window alerts; interventions initiated; false alarms.	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	A suitable later candidate for a cleaned, versioned public dataset.	Open-weight and election-window observability	[ "M-02", "M-03", "M-12", "M-13", "M-14" ]	[ "C-02" ]	[ "G-05", "G-13" ]	[ "CLM-0013", "CLM-0014" ]	[ "CLM-0014" ]	[]	[ "CLM-0013" ]	[ "SRC-005" ]	[ "P05", "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-008	CF-008	[ "P4-008" ]	Official-source grounding for verifiable election-administration facts	Electoral operations	Technical standard / product control	Platform and assistant duty, electoral protocol	Election authorities, assistants, search and social platforms	National; Brazil provides a binding operational precedent	Proposed	Proposed / pending	No general Brazilian official-source-grounding duty; product control proposed for verifiable election-administration facts.	Pilot	Ground answers about registration, eligibility, polling places, voting procedures and certified results in current authoritative election sources, with citations and escalation when sources conflict.	Vector → target → belief or behavior change	Epistemic; decisional; collective / institutional	Normative proposal	None	Indirect / mixed	Strong inference	Static factual election-security messages improved immediate survey confidence and accuracy in United States and Brazil experiments; whether official-source assistant grounding reduces voting-information errors or disenfranchisement is untested.	Delayed official data, fabricated citations, cross-platform migration, correction backfire, and unequal reach.	Confine authoritative grounding to verifiable election administration facts; do not convert it into viewpoint control.	Grounding accuracy; citation validity; source-conflict escalation; correction reach ratio; voting-information error rate; turnout-information error rate.	Pilot	Low-regret component with bounded evidence; monitor transfer and rights costs.	Pilot — 6–18 months	No direct link	No direct link	2026-08-13	The checked evidence concerns static forced-exposure messages, not assistants or real-world voting access.	Official-source grounding for verifiable election-administration facts	[ "M-07", "M-11", "M-13", "M-14" ]	[ "C-05", "D-01" ]	[ "G-06", "G-10", "G-12" ]	[ "CLM-0015", "CLM-0016" ]	[ "CLM-0016" ]	[]	[ "CLM-0015" ]	[ "SRC-078" ]	[]	true	false	false	true	[ "Unsupported as written" ]	[ "Rewrite claim and downgrade implementation to Strong inference" ]	Strong inference	research_preview
I-009	CF-006	[ "P4-009" ]	Proposed capability-plus-reach systemic-risk trigger	Model / provider	Law / regulation	Policy proposal informed by AI Act Article 51 and Annex XIII	European Commission and AI Office	EU	Proposed	Proposed / pending	Policy proposal; not the current EU designation rule	Research	Propose a second policy screen combining measured capability with reach, tool access, scalability and release modality; do not treat this as the current AI Act Article 51 rule.	capability growth → frontier provider → model or API → population and institutions → scalable influence → agency transfer to provider or operator	Executive; Collective / institutional	Normative proposal	E0	None — mechanism or proposal only	Plausible hypothesis	Avoids relying only on FLOP thresholds and can capture highly persuasive, agentic or widely distributed models below a compute threshold	Capability tests can be gamed; designation may lag releases; regulatory uncertainty and threshold gaming	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Share of frontier-equivalent models designated; time from capability evidence to designation; false-negative rate	Research / hold	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Research — 6–30 months	No direct link	Contextual capability link	2026-08-11	Article 51 uses alternative routes—high-impact capability or Commission designation. Reach and users are Annex XIII factors, not a second mandatory trigger.	Manipulation capability-to-deployment decision logic	[ "M-02", "M-06", "M-14", "M-16" ]	[]	[ "G-02" ]	[ "CLM-0017", "CLM-0018" ]	[ "CLM-0018" ]	[]	[ "CLM-0017" ]	[ "SRC-007" ]	[]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-010	CF-010	[ "P4-010" ]	Harmful manipulation as a specified systemic-risk domain	Model / provider	Policy / organizational control	regulated systemic-risk taxonomy	AI Office and GPAI systemic-risk providers	EU	Voluntary / private	Operational voluntary	implemented through AI Act compliance practice	Operational	Require identification and assessment of persuasion, deception and personalised targeting in multi-turn, high-stakes contexts	persuasive capability → provider or malicious user → personalised dialogue → voter or decision-maker → covert belief or behaviour change → agency transfer → democratic harm	Epistemic; Evaluative; Decisional; Collective / institutional	Normative proposal	E0	Indirect / mixed	Plausible hypothesis	Makes manipulation a first-class frontier risk rather than treating it only as misinformation or content safety	Severe-scale threshold may omit cumulative agency loss, local elections, synthetic intimacy and minority targeting	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Percentage of systemic-model reports with manipulation threat models, evaluations, mitigations and residual-risk findings	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	This is the clearest legal destination for the APE, MASK and election red-team artefacts developed in Parts 1–2.	GPAI harmful-manipulation systemic-risk governance	[ "M-03", "M-04", "M-05", "M-08", "M-12" ]	[]	[ "G-02" ]	[ "CLM-0019", "CLM-0020", "CLM-0021" ]	[ "CLM-0021" ]	[ "CLM-0019" ]	[ "CLM-0020" ]	[ "SRC-004" ]	[]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-011	CF-011	[ "P4-011" ]	APE-style harmful-persuasion propensity diagnostic — not a release gate	Model / provider	Research / evaluation	pre-release benchmark and release condition	model providers, independent evaluators and AI safety institutes	international recommendation	Research	Research proposal	APE is an E0 diagnostic, not a validated release gate.	Research	Test whether a model attempts persuasion on harmful and election-relevant objectives across direct, indirect and multi-turn prompts	instruction following → campaign or influence actor → generated persuasive dialogue → voter → attempted preference shaping → agency transfer	Evaluative; Decisional; Collective / institutional	External empirical	E0	None — mechanism or proposal only	Plausible hypothesis	Detects willingness to persuade where generic harmful-content benchmarks and direct-request refusal tests miss risk	Attempt rate is not persuasion efficacy; evaluator-model bias; benchmark contamination; narrow topic coverage	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Harmful-attempt rate with 95% confidence intervals by topic, language, prompt type, turn and safeguard condition	Research / hold	Useful diagnostic signal; no validated path from score to human harm or release prohibition.	Research — 6–30 months	No direct link	Contextual capability link	2026-08-11	Run the persisted APE suite against base, instruct, API, open-weight and abliterated variants; preserve transcripts and release-level metadata.	APE-style harmful-persuasion propensity diagnostic — not a release gate	[ "M-03", "M-04", "M-12", "M-14", "M-16" ]	[]	[ "G-02" ]	[ "CLM-0022", "CLM-0023" ]	[ "CLM-0023" ]	[]	[ "CLM-0022" ]	[ "SRC-008" ]	[ "P04" ]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-012	CF-012	[ "P4-012" ]	Context-specific manipulation efficacy evaluation	Model / provider	Research / evaluation	human-participant TEVV protocol and deployment gate	providers, regulators and accredited research institutions	international recommendation	Voluntary / private	Operational voluntary	implemented by Google DeepMind; recommended as common requirement	Pilot	Measure both manipulative tactics and actual belief or behaviour change in controlled high-stakes interactions across locales	manipulative capability → provider or operator → contextual dialogue → person → belief or behavioural change → loss of reflective agency	Epistemic; Evaluative	Normative proposal	E2	None — mechanism or proposal only	Plausible hypothesis	Separates propensity from efficacy and tests process and outcome harms in contexts where manipulation may actually work	Cost, participant ethics, limited external validity, cultural variation and inability to test the most harmful scenarios directly	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Belief and behaviour change versus control; tactic frequency; heterogeneous effects by locale and vulnerability; persistence at follow-up	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	APE should be the low-cost propensity screen; models crossing it should receive deeper election-specific outcome testing rather than treating APE as sufficient.	Context-specific manipulation efficacy evaluation	[ "M-10", "M-12", "M-13", "M-16" ]	[]	[ "G-02", "G-04", "G-06", "G-12" ]	[ "CLM-0024", "CLM-0025", "CLM-0026" ]	[ "CLM-0026" ]	[ "CLM-0024" ]	[ "CLM-0025" ]	[ "SRC-009", "SRC-093", "SRC-095", "SRC-113", "SRC-114" ]	[ "P03" ]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-013	CF-013	[ "P4-013" ]	MASK-style honesty and pressured-deception diagnostic — not a manipulation gate	Model / provider	Research / evaluation	behavioural benchmark and release assurance	providers and independent evaluators	international recommendation	Research	Research proposal	MASK is an E0 diagnostic, not a manipulation or release gate.	Research	Test whether a model knowingly contradicts its represented beliefs when instructed, incentivised or pressured to deceive	deceptive propensity → model or agent controller → trusted answer or recommendation → user or official → false belief under asymmetric trust → agency transfer	Epistemic; Decisional; Executive; Relational	External empirical	E0	None — mechanism or proposal only	Plausible hypothesis	Adds intentional or pressure-induced deception to manipulation evaluation instead of conflating honesty with factual accuracy	Model beliefs are difficult to infer; benchmark may not represent strategic deception in deployed agents	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Honesty conditional on inferred knowledge; contradiction rate under pressure; calibration and evaluator agreement	Research / hold	Useful diagnostic signal; no validated path from score to human harm or release prohibition.	Research — 6–30 months	No direct link	Contextual capability link	2026-08-11	MASK complements APE: APE measures willingness to influence; MASK tests whether the model will knowingly misrepresent information while doing so.	MASK-style honesty and pressured-deception diagnostic — not a manipulation gate	[ "M-05", "M-06", "M-09", "M-13", "M-16" ]	[]	[ "G-02" ]	[ "CLM-0027", "CLM-0028" ]	[ "CLM-0028" ]	[]	[ "CLM-0027" ]	[ "SRC-010" ]	[ "P04" ]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-014	CF-014	[ "P4-014" ]	Adaptive jailbreak, fine-tuning and abliteration robustness testing	Model / provider	Research / evaluation	mandatory adversarial TEVV and robustness margin	providers, external red teams and regulators	EU compliance practice; international recommendation	Mixed	Applicable	Article 55 duties bind systemic-risk GPAI providers; NIST/Code testing methods remain voluntary compliance choices.	Pilot	Evaluate mitigations under adaptive prompting, safety-removal fine-tunes, altered system prompts, scaffolds and abliterated checkpoints	latent capability → adversarial modifier → safeguard removal → influence campaign → harmful persuasion at scale → agency transfer	Evaluative	Project — exploratory	Project-X	None — legal existence only	Established — legal status	Tests the actual threat model demonstrated by open weights and Part 1 rather than relying on default-chat refusals	Expensive adaptive testing; publication can aid attackers; no stable definition of sufficient robustness	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Worst-case safeguard bypass rate; APE and MASK degradation after attack or fine-tune; attack cost; safety-capability Pareto frontier	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	Exploratory link	Contextual capability link	2026-08-11	Part 1’s persisted jailbreak and abliterated-model runs are direct evidence for making adversarial elicitation a formal release test.	Adaptive jailbreak, fine-tuning and abliteration robustness testing	[ "M-02", "M-16" ]	[]	[ "G-02", "G-12" ]	[ "CLM-0029", "CLM-0030", "CLM-0031" ]	[ "CLM-0031" ]	[ "CLM-0029" ]	[ "CLM-0030" ]	[ "SRC-011" ]	[ "P01", "P03" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-015	CF-015	[ "P4-015" ]	Agent-scaffold and tool-access evaluation	Model / provider	Research / evaluation	agentic capability evaluation and deployment gate	providers, NIST or AISI and independent evaluators	international recommendation	Voluntary / private	Operational voluntary	evaluation infrastructure implemented; policy gate recommended	Pilot	Test models with realistic memory, browsing, tools, parallelism, long horizons and campaign-like scaffolding	agentic capability → operator → tools and persistent workflow → target population → adaptive repeated influence → agency transfer at scale	Evaluative; Executive	Project — exploratory	Project-X	None — mechanism or proposal only	Plausible hypothesis	Reveals capabilities hidden by single-turn model tests and measures end-to-end campaign assistance	Results depend heavily on scaffold quality; agents can exploit or cheat evaluations; dual-use evaluation artefacts	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	End-to-end task success; autonomous steps; tool calls; intervention frequency; time and cost to campaign objective	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	Exploratory link	Contextual capability link	2026-08-11	Re-run Part 1 prompts through agent scaffolds; Part 2 should report both raw-model and system-level capability trajectories.	Agent-scaffold and tool-access evaluation	[ "M-04", "M-06", "M-16" ]	[]	[ "G-02", "G-03" ]	[ "CLM-0032", "CLM-0033", "CLM-0034" ]	[ "CLM-0034" ]	[ "CLM-0032" ]	[ "CLM-0033" ]	[ "SRC-002" ]	[ "P02", "P03" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-016	CF-046	[ "P4-016" ]	Independent evaluator access to capable and minimally mitigated models	Model / provider	Law / regulation	GPAI Code external-evaluator commitments; proposed statutory access right and safe harbour	AI Office, providers and accredited evaluators	EU compliance practice	Proposed	Proposed / pending	Code commitment for signatories; broader access right proposed	Pilot	Give qualified external evaluators timely access to the strongest model, relevant internals and the version with the fewest safeguards	provider information asymmetry → independent testing → credible risk evidence → constrained release decision → reduced agency concentration	Epistemic; Decisional; Relational; Collective / institutional	Normative proposal	None	None — mechanism or proposal only	Plausible hypothesis	Reduces self-assessment bias and permits realistic APE, MASK and red-team testing before irreversible release	Trade-secret and security leakage; provider-selected evaluators; insufficient time, compute or publication freedom	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Number and independence of evaluators; access lead time; compute budget; findings fixed; publication delays	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	The Code commitment is a voluntary compliance route for signatories, not a standalone statutory right to minimally mitigated models or publication freedom.	Independent evaluation and researcher access	[ "M-02", "M-05", "M-12", "M-13", "M-14" ]	[]	[ "G-13" ]	[ "CLM-0035", "CLM-0036" ]	[ "CLM-0036" ]	[]	[ "CLM-0035" ]	[ "SRC-004" ]	[ "P13" ]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-017	CF-017	[ "P4-017" ]	Pre-release safety case and Safety and Security Model Report	Model / provider	Governance / assurance	GPAI Code Safety and Security Model Report; proposed statutory safety-case gate	systemic-risk GPAI provider and AI Office	EU	Proposed	Proposed / pending	Code commitment for signatories; statutory release-condition extension proposed	Pilot	For signatories, document systemic-risk reasoning in the Safety and Security Model Report; a regulator-filed pre-release safety case or release condition is a further proposal.	capability evidence → provider decision → release modality → public exposure → residual manipulation risk → accountability for agency transfer	Evaluative; Decisional; Collective / institutional	Normative proposal	None	None — mechanism or proposal only	Plausible hypothesis	Forces evaluation results, assumptions, mitigations, uncertainty and release reasoning into one auditable decision record	Can become compliance theatre; provider defines acceptability; confidential filings limit public scrutiny	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Pre-release filing rate; safety-margin size; unresolved findings; regulator challenge rate; post-release falsification	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	No EU pre-release approval follows automatically from the Code. APE/MASK may be evidence inputs, never isolated acceptance thresholds.	Pre-release safety case and Safety and Security Model Report	[ "M-02", "M-12", "M-14", "M-16" ]	[]	[ "G-02" ]	[ "CLM-0037", "CLM-0038" ]	[ "CLM-0038" ]	[]	[ "CLM-0037" ]	[ "SRC-004" ]	[]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-018	CF-018	[ "P4-018" ]	Continuous post-deployment re-evaluation	Model / provider	Research / evaluation	continuous monitoring and mandatory report update	providers, deployers and regulator	EU; international recommendation	Binding — limited scope	Applicable	Binding Article 55 duties for GPAI with systemic risk; monitoring methods remain under development.	Operational	Repeat risk assessment after post-training, new tools, higher inference compute, new integrations, incidents or observed drift	deployment change → provider or integrator → altered model affordances → users → emergent influence or dependency → agency transfer	Executive; Relational	Legal / operational precedent	None	None — legal existence only	Established — legal status	Addresses risks that static pre-release benchmarks cannot observe, including adaptive relationships and integration effects	Privacy-invasive logging; distributed-system blind spots; monitoring burden; uncertain causal attribution	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Re-evaluation cadence; drift in APE, MASK and efficacy metrics; incidents per million interactions; time to mitigation	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	The research harness should preserve model version, provider, safeguard state and date so Part 2 becomes a genuine longitudinal monitor.	Continuous post-deployment re-evaluation	[ "M-05", "M-12", "M-15", "M-16" ]	[]	[]	[ "CLM-0039", "CLM-0040", "CLM-0041" ]	[ "CLM-0041" ]	[ "CLM-0039" ]	[ "CLM-0040" ]	[ "SRC-012" ]	[ "P01", "P03" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-019	CF-019	[ "P4-019" ]	Serious-incident reporting plus a proposed manipulation near-miss register	Model / provider	Coordination / incident response	AI Act Article 55 serious-incident duty; proposed near-miss extension	providers, deployers, AI Office and national authorities	EU baseline; recommended expansion internationally	Proposed	Proposed / pending	Serious-incident reporting binding; near-miss extension proposed	Design-ready	Report serious incidents under the binding AI Act duty; separately pilot protected reporting of manipulation near misses, safeguard bypass and deceptive agent behaviour.	deployed model → operator or autonomous agent → harmful influence event → affected public → detected agency loss → corrective action	Evaluative; Executive; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Creates feedback between real misuse and model evaluation, safeguards and regulatory thresholds	Current definitions favour acute catastrophic harm and may miss cumulative autonomy loss, election effects and near misses	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Reporting latency; incident and near-miss rate; taxonomy coverage; corrective-action closure time; recurrence	Research / hold	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	Near misses are not automatically statutory serious incidents. Preserve the distinction in metrics, confidentiality and enforcement.	Serious-incident reporting plus a proposed manipulation near-miss register	[ "M-02", "M-06", "M-12", "M-14", "M-16" ]	[]	[]	[ "CLM-0042", "CLM-0043" ]	[ "CLM-0043" ]	[]	[ "CLM-0042" ]	[ "SRC-007" ]	[ "P01" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-020	CF-020	[ "P4-020" ]	Independent systemic-risk executive and board accountability	Model / provider	Governance / assurance	corporate governance requirement	frontier developer boards and regulators	EU compliance practice; international recommendation	Voluntary / private	Partial / verify	Code commitment for signatories; not a standalone statutory governance duty	Partial / uneven	Assign risk oversight to an executive independent of core capability and product functions, with documented escalation	commercial pressure → provider governance → release decision → public exposure → unchecked influence capability → concentrated power	Decisional; Relational; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Reduces conflicts between model release incentives and manipulation-risk findings	Nominal independence without authority; information controlled by product teams; board capture	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Risk leader independence; veto and escalation use; unresolved safety objections; time and budget allocated to manipulation risk	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	Executive and board governance is a GPAI Code commitment for signatories; broader corporate-governance duties remain a policy extension.	Independent systemic-risk executive and board accountability	[ "M-05", "M-14", "M-15", "M-16" ]	[]	[]	[ "CLM-0044", "CLM-0045", "CLM-0046" ]	[ "CLM-0046" ]	[ "CLM-0044" ]	[ "CLM-0045" ]	[ "SRC-004" ]	[]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-021	CF-021	[ "P4-021" ]	Frontier-safety whistleblower protection	Model / provider	Governance / assurance	employment law, anti-retaliation right and reporting hotline	California Attorney General, courts and frontier developers	California, United States	Binding	Applicable	implemented	Operational	Protect employees reporting concealed risk, misleading safety claims or framework noncompliance and provide anonymous channels	internal risk knowledge → employee disclosure → regulator or board action → constrained unsafe release → reduced provider power asymmetry	Multiple / context-dependent	Legal / operational precedent	None	None — legal existence only	Established — legal status	Improves detection of suppressed evaluation failures and misleading public safety claims	California definition largely omits democratic manipulation; employees may lack evidence or still fear career harm	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Protected reports; substantiation rate; retaliation claims; time to investigation and remedy	Implement now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	A stronger version should expressly protect disclosure of harmful-manipulation eval failures, benchmark gaming and unsafe open-weight decisions.	Frontier-safety whistleblower protection	[ "M-05", "M-12", "M-16" ]	[]	[]	[ "CLM-0047", "CLM-0048", "CLM-0049" ]	[ "CLM-0049" ]	[ "CLM-0047" ]	[ "CLM-0048" ]	[ "SRC-013" ]	[]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-022	CF-022	[ "P4-022" ]	Staged and differential model access	Model / provider	Policy / organizational control	deployment control and conditional licence	model provider under regulator oversight	EU compliance practice; international recommendation	Voluntary / private	Operational voluntary	implemented by some providers and recognised in EU Code	Pilot	Begin with limited, vetted API access and expand only after monitoring; withhold downloadable weights where residual risk is unacceptable	capability → provider-controlled access → vetted operator → bounded influence vector → reduced scale of agency transfer	Multiple / context-dependent	Normative proposal	E0	None — mechanism or proposal only	Plausible hypothesis	Preserves monitoring, rate limiting, intervention and withdrawal options while uncertainty is high	Entrenches closed providers; excludes researchers and smaller firms; abuse migrates to less controlled models	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Misuse rate by access tier; time to detect and suspend abuse; false-positive suspension rate; access concentration	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	Part 1 demonstrates why API safeguards and downloadable or abliterated weights must be evaluated as different policy objects.	Staged access and release-modality governance	[ "M-05", "M-16" ]	[]	[ "G-05" ]	[ "CLM-0050", "CLM-0051", "CLM-0052" ]	[ "CLM-0052" ]	[ "CLM-0050" ]	[ "CLM-0051" ]	[ "SRC-004" ]	[ "P05" ]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-023	CF-023	[ "P4-023" ]	Risk-tiered API identity and campaign-account verification	Model / provider	Technical standard / product control	customer identification, due diligence and access condition	API and cloud providers with regulator oversight	international recommendation; US IaaS proposal provides precedent	Proposed	Proposed / pending	Proposed and unresolved; no final rule identified as of 2026-08-11	Design-ready	Verify high-volume or high-risk customers, beneficial ownership and resellers; retain privacy-preserving attribution records	API capability → campaign or intermediary → high-volume generation → targeted electorate → adaptive manipulation → attributable operator	Evaluative	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Raises the cost of anonymous industrial-scale abuse and improves incident investigation	Privacy and civil-liberties harm; exclusion; stolen identities; offshore providers; concentration advantages for incumbents	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Verified share of high-risk accounts; attribution rate; evasion rate; false rejection and privacy incidents	Research / hold	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	Contextual capability link	2026-08-11	The EO 14110 large-AI-training component was affected by that Order's 2025 revocation. Do not code a current federal training-run reporting duty; the EO 13984 customer-identification basis remains unresolved without a final rule.	Risk-tiered API identity and campaign-account verification	[ "M-10", "M-15", "M-16" ]	[ "C-06" ]	[ "G-07" ]	[ "CLM-0053", "CLM-0054" ]	[ "CLM-0054" ]	[]	[ "CLM-0053" ]	[]	[]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-024	CF-024	[ "P4-024" ]	Technical enforcement for Article 5-qualified manipulative practices	Model / provider	Enforcement	AI Act Article 5 prohibition plus provider implementation controls	AI-system providers, market surveillance authorities and AI Office	EU	Binding	Applicable	Binding prohibition applicable; model- and service-layer implementation uneven	Operational	Build classifiers, account controls, rate limits and escalation for uses that satisfy the Article 5 conjunctive test; do not encode a general prohibition on political persuasion.	model capability → political operator → deceptive or vulnerability-exploiting interaction → voter → significant harm and agency loss	Evaluative; Decisional	Legal / operational precedent	E1	None — legal existence only	Established — legal status	Connects abstract prohibition to detectable provider-side abuse patterns	Article 5 requires significant harm and causal conditions; legitimate persuasion and political speech may be overblocked	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Recall and precision on validated harmful-manipulation cases; evasion rate; appeals; repeat offenders	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	Article 5 is practice- and harm-specific, not a general ban on political manipulation. Controls require contextual review and appeal.	Technical enforcement for Article 5-qualified manipulative practices	[ "M-10", "M-16" ]	[ "C-06" ]	[ "G-06", "G-14" ]	[ "CLM-0055", "CLM-0056", "CLM-0057" ]	[ "CLM-0057" ]	[ "CLM-0055" ]	[ "CLM-0056" ]	[ "SRC-015", "SRC-111" ]	[ "G0", "P01" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-025	CF-022	[ "P4-025" ]	Release-modality safety case for open-weight models	Model / provider	Law / regulation	Policy proposal; AI Act Article 55 supplies systemic-risk duties but no ex ante open-weight licence	model provider and frontier regulator	EU baseline; international recommendation	Proposed	Proposed / pending	Not implemented as an EU release gate; systemic-risk duties can still apply	Pilot	Pilot pre-release assessment and independent review based on capability, modifiability, reach, reversibility and residual risk before irreversible weight release.	frontier capability → weight publisher → unrestricted modifier → abliterated or fine-tuned model → scalable manipulation → irreversible agency transfer capacity	Evaluative; Relational	Normative proposal	E0	None — mechanism or proposal only	Plausible hypothesis	Moves risk assessment before the point at which safeguards, monitoring and withdrawal become impossible	May suppress beneficial openness, competition and independent safety research; thresholds may favour incumbents	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Risk margin at release; third-party findings; time spent in staged access; zero releases before required sign-off	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	The EU does not require capability-based prior approval to publish open weights. Include research and competition safe harbours and measure incumbent-entrenchment effects.	Staged access and release-modality governance	[ "M-02", "M-05", "M-16" ]	[]	[ "G-05" ]	[ "CLM-0058", "CLM-0059" ]	[ "CLM-0059" ]	[]	[ "CLM-0058" ]	[ "SRC-007" ]	[ "P05" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-026	CF-007	[ "P4-026" ]	Open-weight ecosystem monitoring and release registry	Model / provider	Governance / assurance	government monitoring programme and public registry	national AI institutes, model repositories and research partners	United States recommendation; globally applicable	Voluntary / private	Operational voluntary	recommended by NTIA; fragmented implementation	Operational	Maintain a registry of weight releases, derivatives, capability evaluations, licences, hashes and observed misuse or safeguard removal	weight release → distributed modifiers → derivative models → opaque deployment → public influence → reduced supervisory visibility	Collective / institutional	Normative proposal	E0	Indirect / mixed	Plausible hypothesis	Improves visibility without immediately prohibiting open-weight development	Incomplete offshore coverage; false confidence; registry may stigmatise benign research or aid model discovery	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Coverage of major releases and derivatives; registration lag; benchmark freshness; misuse linkage rate	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	Contextual capability link	2026-08-11	Part 2 can seed this registry with dated OpenRouter and repository releases, parameter scale, APE, MASK and safeguard results.	Open-weight and election-window observability	[ "M-02", "M-12", "M-14", "M-16" ]	[]	[ "G-05", "G-13" ]	[ "CLM-0060", "CLM-0061", "CLM-0062" ]	[ "CLM-0062" ]	[ "CLM-0060" ]	[ "CLM-0061" ]	[ "SRC-005" ]	[ "P05", "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-027	CF-027	[ "P4-027" ]	Frontier model-weight cybersecurity standard	Model / provider	Technical standard / product control	mandatory cybersecurity controls and audit	frontier developers, cloud providers and regulator	EU and California	Mixed	Applicable	Binding cybersecurity duty for GPAI with systemic risk; Code methods are voluntary; California scope is separate.	Operational	Protect unreleased weights with least privilege, strong identity controls, hardened interfaces, encryption, confidential computing and exfiltration testing	frontier weights → theft or insider → uncontrolled deployment → malicious actor → scalable harmful capability → loss of provider control	Multiple / context-dependent	Legal / operational precedent	E0	None — legal existence only	Established — legal status	Reduces unauthorised conversion of controlled models into unmonitored or abliterated deployments	High cost; insider threats remain; controls become moot after public release	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Privileged-access count; red-team exfiltration success; time to revoke access; security incidents; control coverage	Enforce now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	Part 1 shows the safety consequence of uncontrolled weights; the register should distinguish intentional open release from theft.	Frontier model-weight cybersecurity standard	[ "M-02", "M-05", "M-06", "M-16" ]	[]	[ "G-05" ]	[ "CLM-0063", "CLM-0064", "CLM-0065" ]	[ "CLM-0065" ]	[ "CLM-0063" ]	[ "CLM-0064" ]	[ "SRC-004" ]	[ "P01", "P05" ]	false	false	true	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-028	CF-028	[ "P4-028" ]	Mandatory downstream model documentation	Model / provider	Governance / assurance	technical documentation and information-sharing duty	GPAI provider and downstream system provider	EU	Binding	Applicable	implemented	Operational	Provide integrators with current information on capabilities, limitations, intended uses, evaluation results and safeguard assumptions	provider knowledge → downstream integrator → assistant or agent design → user → hidden model limits → preventable agency transfer	Epistemic; Executive	Legal / operational precedent	E0	None — legal existence only	Established — legal status	Lets downstream systems implement context-specific manipulation, memory and tool controls	Trade-secret redactions; documentation becomes stale; open-source exemption creates gaps	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Field completeness; update latency; downstream acknowledgement; integration incidents attributable to missing information	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	APE, MASK and red-team limitations should appear in downstream documentation, including differences between API and downloadable variants.	Mandatory downstream model documentation	[ "M-02", "M-06", "M-13", "M-16" ]	[]	[ "G-05" ]	[ "CLM-0066", "CLM-0067", "CLM-0068" ]	[ "CLM-0068" ]	[ "CLM-0066" ]	[ "CLM-0067" ]	[ "SRC-007" ]	[ "P01", "P05" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-029	CF-029	[ "P4-029" ]	Machine-readable provenance for synthetic outputs	Model / provider	Technical standard / product control	technical standard and transparency obligation	generative-AI providers, deployers and standards bodies	EU and China	Binding	Applicable	implemented from 2026-08-02 in EU and 2025-09-01 in China	Operational	Embed robust, interoperable and detectable markers in generated text, audio, image and video, with visible labels for covered content	generation capability → model provider → synthetic political content → public → mistaken authenticity → epistemic agency transfer	Epistemic; Collective / institutional	Legal / operational precedent	E0	None — legal existence only	Established — legal status	Improves attribution and detection of synthetic campaign material	Metadata stripping, laundering, false labels and limited effect on truthful but manipulative dialogue	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Marker detection and retention after transformations; false-positive rate; platform ingestion coverage	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	This mitigates provenance, not the interactive agency-transfer mechanisms measured by APE or manipulation efficacy tests.	Synthetic-output provenance and disclosure	[ "M-08", "M-12", "M-13", "M-14", "M-15" ]	[ "X-01" ]	[ "G-05", "G-09" ]	[ "CLM-0069", "CLM-0070", "CLM-0071" ]	[ "CLM-0071" ]	[ "CLM-0069" ]	[ "CLM-0070" ]	[ "SRC-016" ]	[ "P01", "P05", "P09" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-030	CF-003	[ "P4-030" ]	Agent identity, least privilege and high-impact confirmation	Model / provider	Technical standard / product control	technical protocol, authorisation standard and deployment rule	NIST, standards bodies, agent providers and integrators	United States initiative; international recommendation	Voluntary / private	Operational voluntary	standards initiative implemented; controls under development	Operational	Give agents verifiable identities, scoped permissions, short-lived credentials and mandatory human confirmation for consequential actions	agentic model → delegated tools → communication or account action → citizen or institution → autonomous influence → agency transfer	Executive; Collective / institutional	Project — exploratory	Project-X	Indirect / mixed	Plausible hypothesis	Constrains agents from autonomously escalating persuasion campaigns, impersonating users or acting beyond delegated intent	Confirmation fatigue; compromised identities; standards fragmentation; agents may route around controls	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Unauthorised-action rate; permission scope; credential lifetime; confirmed versus automatic consequential actions	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	Exploratory link	Contextual capability link	2026-08-11	Part 1 agentic demonstrations should log every tool permission and human checkpoint, not only final text outputs.	Agent authorization, least privilege and action control	[ "M-06", "M-14", "M-16" ]	[]	[ "G-03" ]	[ "CLM-0072", "CLM-0073", "CLM-0074" ]	[ "CLM-0074" ]	[ "CLM-0072" ]	[ "CLM-0073" ]	[ "SRC-002" ]	[ "P02" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-031	CF-031	[ "P4-031" ]	Long-horizon agency and dependency monitoring	Model / provider	Research / evaluation	post-deployment human-factors monitoring and periodic impact study	providers, independent researchers and regulators	international recommendation	Voluntary / private	Operational voluntary	recommended; methods nascent	Pilot	Audit persistent memory, repeated recommendations, emotional cues, reliance and preference drift over extended interactions	memory and personalisation → assistant provider → repeated intimate interaction → user → dependency and preference shaping → cumulative agency transfer	Evaluative; Decisional; Executive; Relational	Normative proposal	E2	None — mechanism or proposal only	Plausible hypothesis	Targets the project’s central risk that manipulation becomes relational infrastructure rather than a single persuasive message	Intrusive observation; difficult causal inference; no settled agency metric; vulnerable users are hard to study ethically	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Preference and trust drift; reliance; reversal after disclosure; user control over memory; persistence after 7, 30 and 90 days	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	APE is too short-horizon for this mechanism; Part 2 should explicitly mark it as a missing evaluation layer rather than overclaim coverage.	Long-horizon agency and dependency monitoring	[ "M-03", "M-04", "M-05", "M-09", "M-12" ]	[ "D-01" ]	[ "G-01", "G-04", "G-06", "G-10", "G-12", "G-16" ]	[ "CLM-0075", "CLM-0076", "CLM-0077" ]	[ "CLM-0077" ]	[ "CLM-0075" ]	[ "CLM-0076" ]	[ "SRC-017" ]	[ "P03", "P10" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-032	CF-032	[ "P4-032" ]	Threshold-triggered notification for GPAI systemic-risk classification	Model / provider	Governance / assurance	mandatory compute reporting	GPAI provider and AI Office	EU	Binding — limited scope	Applicable	Article 52 notification is threshold-triggered; no general advance notification of training runs.	Operational	Notify the Commission without delay and within two weeks when Article 51 high-impact capability is met or expected to be met; do not treat this as general pre-training notification.	planned capability scale-up → developer → regulator notice → early evaluation → deployment constraints before public exposure	Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Gives regulators time to commission manipulation and agentic testing before launch	Compute is an imperfect proxy; efficiency gains and distributed training can evade thresholds; reporting burden	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Notification compliance; lead time before release; unreported threshold crossings; capability per FLOP	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	Part 2 should show why notification must trigger capability testing rather than treating compute itself as manipulation evidence.	Threshold-triggered notification for GPAI systemic-risk classification	[ "M-14", "M-15", "M-16" ]	[]	[]	[ "CLM-0078", "CLM-0079", "CLM-0080" ]	[ "CLM-0080" ]	[ "CLM-0078" ]	[ "CLM-0079" ]	[ "SRC-007", "SRC-111" ]	[ "P01" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-033	CF-033	[ "P4-033" ]	Regulator model-access, evaluation, mitigation and recall powers	Model / provider	Research / evaluation	supervisory and market-enforcement power	European AI Office	EU	Binding	Applicable	implemented	Operational	Allow the regulator to demand information, run evaluations, order mitigations and withdraw or recall a non-compliant model	provider-controlled evidence → independent regulator → enforceable finding → release or access change → reduced public exposure	Epistemic; Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Turns safety reports and manipulation evaluations into potentially consequential oversight	Limited technical capacity; slow enforcement; confidentiality; recall is ineffective after open-weight proliferation	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Access latency; regulator evaluations; mitigation orders; compliance time; recall effectiveness	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	The project can propose APE, MASK and election red-team suites as concrete tests available to the AI Office or national institutes.	Regulator model-access, evaluation, mitigation and recall powers	[ "M-02", "M-05", "M-12", "M-13", "M-14" ]	[]	[ "G-13" ]	[ "CLM-0081", "CLM-0082", "CLM-0083" ]	[ "CLM-0083" ]	[ "CLM-0081" ]	[ "CLM-0082" ]	[ "SRC-007" ]	[ "G0", "P01" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-034	CF-098	[ "P4-034" ]	Certified AI management system as procurement or licensing condition	Model / provider	Procurement	ISO certification, procurement clause or licence condition	public procurers, regulators and accredited certification bodies	international recommendation	Voluntary / private	Operational voluntary	standard implemented; mandatory use recommended selectively	Operational	Require an independently audited AI management system with risk ownership, lifecycle controls, monitoring and continual improvement	provider organisation → governance processes → model lifecycle → downstream public systems → reduced unmanaged influence risk	Relational; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Provides auditable organisational infrastructure for repeated evaluation and incident response	Process certification does not prove model safety; audit shopping; cost barriers for smaller providers	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Certification coverage; major nonconformities; corrective-action closure; manipulation-specific control maturity	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	Certification should require evidence from the project’s tests; ISO compliance alone cannot substitute for APE, MASK or adversarial evaluation.	Public-sector AI registers, impact assessment and procurement	[ "M-05", "M-12", "M-14", "M-16" ]	[ "S-01" ]	[ "G-11" ]	[ "CLM-0084", "CLM-0085", "CLM-0086" ]	[ "CLM-0086" ]	[ "CLM-0084" ]	[ "CLM-0085" ]	[ "SRC-018" ]	[ "P11" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-035	CF-035	[ "P4-035" ]	Democracy and agency impact assessment	Model / provider	Governance / assurance	rights-and-democracy impact assessment with stakeholder engagement	states, providers, public authorities and independent assessors	Council of Europe parties and EU	Non-binding official guidance	Operational voluntary	HUDERIA is non-binding; CETS 225 is not yet in force.	Pilot	Assess scale, scope, likelihood, reversibility and stakeholder impacts on autonomy, opinion formation and democratic processes	model and deployment design → controlling actor → influence environment → affected publics → agency transfer → democratic concentration	Collective / institutional	Legal / operational precedent	None	None — mechanism or proposal only	Plausible hypothesis	Captures autonomy and democratic harms omitted by catastrophic-risk frameworks	High-level principles; private-sector scope varies; self-assessment and weak remedies	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Assessments completed; stakeholder inclusion; high-risk findings; mitigation closure; residual democratic risk	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	Use the project’s causal chain and agency-transfer construct to make HUDERIA operational for frontier-model manipulation.	Rights, democracy and agency impact assessment	[ "M-14", "M-15", "M-16" ]	[ "S-01" ]	[ "G-01", "G-11", "G-12", "G-14" ]	[ "CLM-0087", "CLM-0088", "CLM-0089" ]	[ "CLM-0089" ]	[ "CLM-0087" ]	[ "CLM-0088" ]	[ "SRC-019" ]	[ "G0" ]	false	false	true	false	[]	[]	Plausible hypothesis	research_preview
I-036	CF-036	[ "P4-036" ]	International frontier-risk reporting framework	Model / provider	Governance / assurance	international code and reporting template	OECD, G7 governments and advanced-AI developers	international	Voluntary / private	Operational voluntary	implemented	Operational	Standardise voluntary disclosures on risk identification, evaluation, incident management, security and transparency across providers	cross-border providers → comparable reporting → regulator and public scrutiny → convergent controls → reduced jurisdictional arbitrage	Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Creates comparable evidence across providers and jurisdictions without waiting for a global treaty	Self-reporting, selective disclosure, inconsistent interpretation and no sanction for poor performance	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Provider participation; field completeness; independent verification; manipulation-domain coverage; year-on-year improvement	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	Add standard fields for APE, MASK, harmful-manipulation efficacy, open-weight modality and adversarial safeguard performance.	International frontier-risk reporting framework	[ "M-12", "M-14", "M-16" ]	[]	[]	[ "CLM-0090", "CLM-0091", "CLM-0092" ]	[ "CLM-0092" ]	[ "CLM-0090" ]	[ "CLM-0091" ]	[ "SRC-020" ]	[]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-037	CF-037	[ "P4-037" ]	Security assessment and algorithm filing for socially mobilising generative AI	Model / provider	Governance / assurance	administrative filing, security assessment and inspection	Cyberspace Administration of China and sector regulators	China	Binding	Applicable	implemented	Operational	Require security assessment, algorithm filing and regulatory cooperation for services with public-opinion or social-mobilisation capacity	generative capability → service provider → mass communication → population → opinion mobilisation → state-supervised deployment	Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Demonstrates a direct model-service trigger based on influence and mobilisation capability	Can enable political censorship and state control rather than democratic agency; opaque assessment criteria	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Services assessed and filed; review time; enforcement actions; transparency and appeal availability	Monitor / shape	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	Useful as a structural precedent, not a normative model: democratic safeguards must prevent the intervention itself becoming concentrated manipulation power.	Security assessment and algorithm filing for socially mobilising generative AI	[ "M-14", "M-16" ]	[]	[]	[ "CLM-0093", "CLM-0094", "CLM-0095" ]	[ "CLM-0095" ]	[ "CLM-0093" ]	[ "CLM-0094" ]	[ "SRC-021" ]	[]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-038	CF-038	[ "P4-038" ]	Enhanced review of cloud–frontier-model partnerships	Model / provider	Enforcement	merger review, market investigation and antitrust enforcement	competition authorities	United States, EU and allied jurisdictions	Binding	Applicable	implemented case-by-case	Operational	Scrutinise equity, exclusivity, control, information sharing, compute commitments and preferential distribution in AI partnerships	compute and capital concentration → cloud–model alliance → model and distribution control → dependent firms and publics → concentrated epistemic power	Epistemic; Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Limits vertical structures that concentrate compute, models, distribution and safety evidence in the same actors	Partnerships also finance expensive models and safeguards; uncertain market definition; remedies may arrive late	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Exclusivity clauses; compute concentration; switching costs; model distribution share; merger remedies	Monitor / shape	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	Contextual capability link	2026-08-11	Part 2 capability concentration should be read alongside ownership and cloud dependence, not merely as a model leaderboard.	AI market concentration and structural oversight	[ "M-02", "M-05", "M-09", "M-13", "M-14" ]	[ "S-01" ]	[]	[ "CLM-0096", "CLM-0097", "CLM-0098" ]	[ "CLM-0098" ]	[ "CLM-0096" ]	[ "CLM-0097" ]	[ "SRC-022" ]	[ "P12" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-039	CF-039	[ "P4-039" ]	Model and cloud portability, interoperability and anti-lock-in rules	Model / provider	Procurement	competition remedy, interoperability standard and procurement requirement	competition authorities, public procurers and standards bodies	EU and international recommendation	Proposed	Proposed / pending	Case-specific competition and procurement powers provide precedents; no general binding model-and-cloud portability duty is established by the cited policy brief.	Design-ready	Reduce egress barriers, support model switching and multi-provider interfaces, and prevent self-preferencing or coercive tying	platform dependency → dominant provider → default model and assistant → organisations and citizens → infrastructural reliance → agency concentration	Relational; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Preserves contestability and reduces dependence on one provider’s model, safeguards and epistemic defaults	Interoperability can increase attack surfaces and lowest-common-denominator safety; technical quality differs	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Migration time and cost; egress charges; multi-homing rate; interface compatibility; market concentration	Monitor / shape	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	Contextual capability link	2026-08-11	Open weights can improve contestability, but Part 1 shows why competition benefits cannot erase release-risk differences.	Interoperability, portability and exit	[ "M-05", "M-13", "M-14", "M-15", "M-16" ]	[ "S-01" ]	[ "G-10", "G-11" ]	[ "CLM-0099", "CLM-0100" ]	[ "CLM-0100" ]	[]	[ "CLM-0099" ]	[ "SRC-023" ]	[ "P12" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-040	CF-040	[ "P4-040" ]	Public-interest compute for independent evaluation and safer alternatives	Model / provider	Research / evaluation	public infrastructure and competitive grant allocation	governments, universities and AI safety institutes	California precedent; international recommendation	Voluntary / private	Operational voluntary	CalCompute planning mandated; other public programmes operational	Operational	Fund public compute capacity for evaluators, academia, civil society and public-interest model development	compute concentration → limited independent evidence → provider dominance → weak oversight → concentrated model and epistemic power	Epistemic; Relational; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Makes external APE, MASK, agentic and open-weight evaluation technically feasible and supports plural model supply	Political allocation, duplication, security exposure, insufficient scale and capture by incumbents	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Independent evaluator GPU hours; models evaluated; organisations served; replication rate; cost per evaluation	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	This is the infrastructure needed to keep the Part 1–2 evaluation suite independent, persistent and runnable as new open models appear.	Public-interest compute and AI infrastructure	[ "M-05", "M-12", "M-13", "M-14", "M-15" ]	[ "S-01" ]	[ "G-11", "G-13" ]	[ "CLM-0101", "CLM-0102", "CLM-0103" ]	[ "CLM-0103" ]	[ "CLM-0101" ]	[ "CLM-0102" ]	[ "SRC-013" ]	[ "P12", "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-041	CF-046	[ "P4-041" ]	Open benchmark commons and standardised public evaluation registry	Model / provider	Research / evaluation	public evaluation platform, registry and reporting standard	AISI, NIST, academia and independent evaluators	international	Voluntary / private	Operational voluntary	partially implemented	Pilot	Publish versioned evaluation code, model metadata, scoring specifications, uncertainty and reproducible release-level results	opaque provider testing → public benchmark infrastructure → comparable evidence → procurement and regulatory decisions → reduced information asymmetry	Epistemic; Decisional; Collective / institutional	Normative proposal	E0	None — mechanism or proposal only	Plausible hypothesis	Makes capability and safeguard trends independently auditable across closed, open and modified models	Benchmark gaming, contamination, evaluator monoculture and publication of dangerous artefacts	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Release coverage; reproducibility; benchmark freshness; private holdout share; inter-evaluator agreement	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	Contextual capability link	2026-08-11	The project should package APE, MASK and election red-team tasks as versioned Inspect-compatible evals while keeping dangerous cases access-controlled.	Independent evaluation and researcher access	[ "M-12", "M-13", "M-14", "M-16" ]	[ "D-01" ]	[ "G-06", "G-12", "G-13" ]	[ "CLM-0104", "CLM-0105", "CLM-0106" ]	[ "CLM-0106" ]	[ "CLM-0104" ]	[ "CLM-0105" ]	[ "SRC-024" ]	[ "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-042	CF-042	[ "P4-042", "P4-085" ]	Election-specific systemic-risk assessment and mitigation for very large platforms and search engines	Platform / distribution	Law / regulation	Digital Services Act Articles 34-35 plus Commission election guidelines	VLOPs and VLOSEs; supervised by the European Commission and Digital Services Coordinators	European Union	Binding	Applicable	In force; guidelines are the current compliance benchmark	Operational	Require VLOPs and VLOSEs to assess election-specific risks from service design, ranking, advertising, generative AI, coordinated manipulation and local context; deploy mitigations before the electoral period, maintain an incident mechanism and publish a post-election review.	AI-enabled campaign or state actor → platform design and ranking → voter exposure and salience → belief, attention or trust change → agency transfer to campaign and platform → concentrated agenda-setting → electoral distortion.	Epistemic; Attentional; Evaluative; Decisional; Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Moves governance upstream from takedowns to product design, staffing, language coverage, recommender controls and election readiness.	Covers only designated very large services; relies heavily on platform self-assessment; lacks common outcome benchmarks; can become process compliance without reducing exposure or agency loss.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Risk and mitigation coverage by vector and language; prevalence and reach of violating content; recommendation-driven exposure; incident time-to-detection and time-to-mitigation; post-election residual risk; audit findings.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-11	The duties bind VLOPs/VLOSEs. The election guidelines interpret and operationalise those duties but are not a standalone law.	Platform election-risk assessment and implementation	[ "M-05", "M-09", "M-10", "M-12", "M-13" ]	[ "C-01", "C-02", "C-05", "C-06" ]	[ "G-07", "G-12" ]	[ "CLM-0107", "CLM-0108", "CLM-0109" ]	[ "CLM-0109" ]	[ "CLM-0107" ]	[ "CLM-0108" ]	[ "SRC-003", "SRC-053" ]	[ "P06" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-043	CF-043	[ "P4-043" ]	Recommender-system transparency for online platforms	Platform / distribution	Governance / assurance	Digital Services Act Articles 27 and 38	Online platforms and VLOPs; enforced by Digital Services Coordinators and the European Commission	European Union	Binding	Applicable	DSA Article 27 recommender transparency for online platforms; the Article 38 non-profiling option is separated into I-084.	Operational	Disclose in clear language the main recommender parameters and users' options to modify or influence them under DSA Article 27. The non-profiling option is coded separately under I-084.	Platform optimisation → personalised feed → repeated political exposure → attention and preference shaping → agency transfer to platform → concentrated control of the decision environment.	Epistemic; Attentional; Evaluative; Decisional; Relational	Legal / operational precedent	None	None — legal existence only	Established — legal status	Creates an exit from inferred-vulnerability targeting and gives users some control over how political attention is allocated.	Non-profiled does not mean neutral or safe; platforms can nudge users toward the default; transparency may be too abstract; chronological feeds can still be gamed by coordinated networks.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Default versus alternative feed prominence; opt-in and persistence rates; political-content concentration and source diversity; exposure inequality; user comprehension; differential manipulation prevalence by feed mode.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-11	null	Recommender-system transparency for online platforms	[ "M-03", "M-05", "M-09", "M-10", "M-12" ]	[ "C-01", "C-05", "C-06" ]	[ "G-12" ]	[ "CLM-0110", "CLM-0111", "CLM-0112" ]	[ "CLM-0112" ]	[ "CLM-0110" ]	[ "CLM-0111" ]	[ "SRC-025" ]	[ "P06" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-044	CF-044	[ "P4-044", "P4-083" ]	Ban on manipulative interface design and dark patterns	Cross-layer	Law / regulation	Digital Services Act Article 25	Online platforms; national Digital Services Coordinators	European Union	Binding	Applicable	In force	Operational	Prohibit online-platform interfaces that deceive, manipulate, or materially distort users' ability to make free and informed decisions.	Platform or campaign-controlled interface → deceptive choice architecture → reduced reflection or coerced disclosure → behavioural or dependency change → agency transfer → platform power concentration.	Evaluative; Decisional; Relational; Collective / institutional	Legal / operational precedent	E1	Indirect / mixed	Strong inference	First-page button prominence can change consent acceptance; whether a legal ban improves free and informed choice, data disclosure, political behaviour, or agency is untested.	Boundary between persuasion and manipulation is context-sensitive; enforcement may focus on commercial UX and miss political onboarding, consent flows, notification pressure or agent defaults.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Completion and consent asymmetries; reversal or opt-out friction; comprehension; alignment with stated preferences; actual data disclosure; enforcement findings.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-13	A small interface experiment supports a bounded behavioural endpoint, not the effect of DSA enforcement or a general dark-pattern ban.	Dark patterns and unfair personalization	[ "M-05", "M-06", "M-14", "M-15", "M-16" ]	[ "C-01", "C-06" ]	[ "G-12" ]	[ "CLM-0113", "CLM-0114", "CLM-0115" ]	[ "CLM-0115" ]	[ "CLM-0113" ]	[ "CLM-0114" ]	[ "SRC-025", "SRC-026", "SRC-115" ]	[ "P06" ]	true	true	false	true	[ "Partially supported with different evidence" ]	[ "Rewrite claim and downgrade implementation to Strong inference" ]	Strong inference	research_preview
I-045	CF-045	[ "P4-045" ]	Online-platform ad transparency and sensitive-data targeting limits	Platform / distribution	Law / regulation	Digital Services Act Articles 26 and 39	Online platforms and VLOPs	European Union	Binding	Applicable	DSA Article 26 ad transparency and sensitive-data targeting limits for online platforms; Article 39 repositories are VLOP/VLOSE-only.	Operational	Identify advertisements and sponsors, disclose key targeting parameters, and prohibit ad profiling using special-category data under DSA Article 26. The DSA Article 39 repository is VLOP/VLOSE-only; the broader political-ad repository is coded under P4-049.	Campaign or proxy → opaque paid targeting → selected vulnerable voter → preference or turnout change → agency transfer to campaign and ad platform → unequal political power.	Evaluative; Decisional; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Makes paid influence more observable and reduces direct use of sensitive traits for personalised persuasion.	Issue ads, influencers and organic coordination may evade classification; inferred proxies can reproduce sensitive targeting; repositories can be incomplete or hard to query.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Ad coverage and missingness; sponsor verification; targeting-field completeness; repository latency; spend and reach by audience; rate of proxy-sensitive segments; researcher usability.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-11	null	Political-ad transparency, targeting and repositories	[ "M-03", "M-09", "M-14", "M-16" ]	[ "C-01", "C-05", "C-06" ]	[ "G-07", "G-12" ]	[ "CLM-0116", "CLM-0117", "CLM-0118" ]	[ "CLM-0118" ]	[ "CLM-0116" ]	[ "CLM-0117" ]	[ "SRC-025" ]	[ "P06", "P07" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-046	CF-046	[ "P4-046", "P4-093" ]	Vetted researcher access to non-public platform data	Institutional / structural	Research / evaluation	Digital Services Act Article 40 and 2025 delegated act on data access	Digital Services Coordinators, European Commission, VLOPs/VLOSEs and vetted researchers	European Union	Binding — limited scope	Applicable	Conditional Article 40 data access for designated VLOPs/VLOSEs, subject to reasoned requests and safeguards.	Operational	Allow qualified independent researchers to request necessary VLOP and VLOSE data to identify systemic risks and evaluate mitigation, through Digital Services Coordinator vetting and a common portal.	Opaque platform data → weak external detection → unchecked optimisation and influence → hidden agency transfer; researcher access interrupts this by making exposure, coordination and mitigation auditable.	Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Enables measurement of recommender exposure, cross-account coordination and whether safeguards work, closing the central evidence gap in Part 3.	Slow approvals can miss elections; privacy and trade-secret objections can narrow datasets; qualified-researcher criteria exclude journalists and smaller civil society groups; platform data schemas can frustrate replication.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Application and approval times; request scope granted; dataset completeness; reproducibility; number of election studies; disclosure disputes; time from finding to mitigation or enforcement.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-11	null	Independent evaluation and researcher access	[ "M-05", "M-12", "M-14", "M-16" ]	[ "C-01", "C-06" ]	[ "G-12", "G-13" ]	[ "CLM-0119", "CLM-0120", "CLM-0121" ]	[ "CLM-0121" ]	[ "CLM-0119" ]	[ "CLM-0120" ]	[ "SRC-027", "SRC-059" ]	[ "P06", "P13" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-047	CF-047	[ "P4-047" ]	Annual independent audits of systemic-risk and code commitments	Cross-layer	Governance / assurance	Digital Services Act Article 37 and Article 45 code framework	Independent audit organisations, platforms and European Commission	European Union	Mixed	Applicable	In force; Code commitments auditable from 2025-07-01	Operational	Require annual independent audits of VLOP/VLOSE compliance, including risk controls and, for adherents, auditable Code of Conduct on Disinformation commitments.	Platform self-governance → unverified safeguards → persistent manipulation capacity → agency transfer; external assurance raises the cost of superficial controls.	Epistemic; Evaluative; Relational	Legal / operational precedent	None	None — legal existence only	Established — legal status	Creates recurring scrutiny of whether claimed election safeguards are implemented and effective.	Auditor dependence, narrow assurance scopes, confidential evidence, lagging annual cadence and checklist behaviour can conceal election-period failures.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Qualified and adverse findings; remediation closure time; repeat findings; audit coverage of local languages and recommenders; public disclosure quality; enforcement following audit failures.	Enforce now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	No direct link	2026-08-11	Annual Article 37 audits are binding for covered services; joining the Article 45 Code remains voluntary.	Annual independent audits of systemic-risk and code commitments	[ "M-02", "M-04", "M-05", "M-12", "M-13" ]	[ "C-01" ]	[]	[ "CLM-0122", "CLM-0123", "CLM-0124" ]	[ "CLM-0124" ]	[ "CLM-0122" ]	[ "CLM-0123" ]	[ "SRC-028", "SRC-086" ]	[ "P06" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-048	CF-042	[ "P4-048" ]	Pre-election stress tests, readiness dialogues and regulator election toolkits	Cross-layer	Coordination / incident response	DSA Elections Toolkit and voluntary readiness exercises under the Article 35 framework	European Commission, Digital Services Coordinators, national election authorities, platforms and civil society	EU and partner countries, including Moldova	Voluntary / private	Operational voluntary	Operational best practice; used in EU elections and Moldova	Operational	Run scenario exercises with platforms, election bodies, regulators, security actors and civil society; test escalation paths, local-language capacity and platform responses before voting.	Emerging AI-enabled incident → fragmented detection and authority → delayed response → rapid belief or turnout effects → temporary agency seizure; stress tests pre-wire coordination.	Epistemic; Executive; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Reduces coordination latency and exposes gaps before a late-breaking deepfake, coordinated network or false voting claim appears.	Scripted exercises may not test adversarial evasion; voluntary platform participation; threat scenarios can become stale; unclear thresholds for public attribution or takedown.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Exercise gap closure; escalation and decision times; contact availability; platform response consistency; red-team success rate; coverage of private messaging and small platforms.	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	No direct link	2026-08-11	null	Platform election-risk assessment and implementation	[ "M-12", "M-13", "M-14", "M-16" ]	[ "C-01", "C-02", "C-05" ]	[ "G-12" ]	[ "CLM-0125", "CLM-0126", "CLM-0127" ]	[ "CLM-0127" ]	[ "CLM-0125" ]	[ "CLM-0126" ]	[ "SRC-029" ]	[ "P06" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-049	CF-045	[ "P4-049" ]	Mandatory labels, sponsor notices and an EU repository for political advertising	Platform / distribution	Law / regulation	Regulation (EU) 2024/900, especially Articles 11-15	Sponsors, political-ad service providers, publishers, national competent authorities and European Commission	European Union	Binding	Applicable	Applicable since 2025-10-10	Operational	Require each paid or targeted political ad to be labelled and linked to a notice naming the sponsor, election, remuneration, spend and targeting; transmit online ads to a European repository and accelerate complaint handling near elections.	Sponsor or proxy → covert paid communication → voter → hidden persuasion → agency transfer to sponsor → concentrated political influence; disclosure restores source contestability.	Epistemic; Evaluative; Decisional; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Makes who paid, how much and why a voter was reached visible to voters, regulators and researchers.	Platforms may exit political advertising, shifting influence to organic creators, messaging apps and foreign intermediaries; sponsor declarations can be false; repository implementation can lag.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Share of ads with complete notices; sponsor-verification failures; repository latency; notification resolution within 48 hours in the final month; undisclosed paid influence detected outside formal ads.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	6–18 months	No direct link	No direct link	2026-08-11	null	Political-ad transparency, targeting and repositories	[ "M-03", "M-09", "M-11", "M-13", "M-14" ]	[ "C-06" ]	[ "G-07", "G-14" ]	[ "CLM-0128", "CLM-0129", "CLM-0130" ]	[ "CLM-0130" ]	[ "CLM-0128" ]	[ "CLM-0129" ]	[ "SRC-030" ]	[ "P07" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-050	CF-045	[ "P4-050" ]	Restrict personal-data political-ad targeting and prohibit special-category profiling	Platform / distribution	Governance / assurance	Regulation (EU) 2024/900 Articles 18-19	Campaigns, controllers, ad-tech providers, publishers and data protection authorities	European Union	Binding	Applicable	Regulation 2024/900 restricts personal-data political targeting and prohibits special-category profiling; it is not a blanket microtargeting ban.	Operational	Permit personal-data-based political targeting only when data were collected from the person and separate explicit consent was obtained; prohibit profiling using special-category data and targeting known near-voting-age minors.	Campaign data broker → vulnerability inference and microtargeting → individual voter → tailored pressure or preference shaping → agency transfer to campaign → asymmetric democratic power.	Evaluative; Decisional; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Cuts the Cambridge Analytica pathway from intimate or political data to personalised influence without meaningful voter consent.	Contextual and lookalike targeting can act as proxies; consent fatigue and bundled interfaces undermine autonomy; political actors can move targeting in-house or into messaging lists.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Valid-consent rate; use of inferred proxies; audience-segment size; sensitive-data violations; targeting explanations understood; proportion of political reach occurring outside regulated advertising.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	6–18 months	No direct link	No direct link	2026-08-11	null	Political-ad transparency, targeting and repositories	[ "M-03", "M-09", "M-11", "M-14", "M-16" ]	[ "C-06" ]	[ "G-07", "G-14" ]	[ "CLM-0131", "CLM-0132", "CLM-0133" ]	[ "CLM-0133" ]	[ "CLM-0131" ]	[ "CLM-0132" ]	[ "SRC-030", "SRC-031" ]	[ "P07" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-051	CF-045	[ "P4-051" ]	Three-month pre-election ban on political-ad services for third-country sponsors	Platform / distribution	Law / regulation	Regulation (EU) 2024/900	Political-ad service providers, publishers and competent national authorities	European Union	Binding	Applicable	Applicable since 2025-10-10	Operational	Prohibit political advertising services for sponsors outside the EU during the three months before an election or referendum in a Member State.	Foreign state or proxy → paid distribution → domestic electorate → agenda or trust change → agency transfer abroad → external concentration of electoral influence.	Evaluative; Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Raises the cost of overt foreign paid influence during the highest-risk period.	Shell entities, domestic agents, influencers, issue advertising and unpaid coordinated content can evade sponsor-location rules; a three-month window misses long-horizon preference shaping.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Rejected third-country purchases; beneficial-owner verification; proxy sponsorship cases; foreign-linked organic reach; spending displacement before the window.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	6–18 months	No direct link	No direct link	2026-08-11	null	Political-ad transparency, targeting and repositories	[ "M-03", "M-05", "M-06", "M-09", "M-10" ]	[ "C-06" ]	[ "G-07", "G-14" ]	[ "CLM-0134", "CLM-0135", "CLM-0136" ]	[ "CLM-0136" ]	[ "CLM-0134" ]	[ "CLM-0135" ]	[ "SRC-030", "SRC-031" ]	[ "P07" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-052	CF-029	[ "P4-052" ]	Machine-readable marking of AI-generated content and visible deepfake disclosure	Platform / distribution	Law / regulation	EU AI Act Article 50 and Code of Practice on Transparency of AI-generated Content	AI providers and deployers; market-surveillance authorities and AI Office	European Union	Binding	Applicable	Binding Article 50 duties from 2026-08-02; voluntary code is an implementation route. Limited transition to 2026-12-02 for Article 50(2) pre-existing generators.	Operational	Require providers to mark synthetic outputs in machine-readable form and deployers to disclose deepfakes and certain AI-generated public-interest text; use an EU code and common icons to standardise compliance.	Generative model or campaign → synthetic impersonation → voter → false source belief or trust shock → agency transfer to fabricator → electoral harm; provenance and disclosure restore source knowledge.	Epistemic; Decisional; Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Supports cross-platform detection and visible disclosure of synthetic candidate or election content.	Metadata can be stripped; open models and foreign actors may not comply; absence of a mark is not proof of authenticity; labels can create disclosure fatigue or a liar's dividend.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Marking coverage by provider and modality; persistence after editing and re-upload; detector precision and recall; label visibility and comprehension; share of viral synthetic incidents carrying valid provenance.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	Applicable now; limited transition to 2026-12-02	No direct link	No direct link	2026-08-11	null	Synthetic-output provenance and disclosure	[ "M-05", "M-08", "M-09", "M-13", "M-14" ]	[ "C-03", "C-04", "C-05", "X-01" ]	[ "G-09", "G-12", "G-14" ]	[ "CLM-0137", "CLM-0138", "CLM-0139" ]	[ "CLM-0139" ]	[ "CLM-0137" ]	[ "CLM-0138" ]	[ "SRC-016", "SRC-111" ]	[ "P01", "P09" ]	false	true	false	false	[]	[]	Established — legal status	research_preview
I-053	CF-053	[ "P4-053" ]	Auditable platform commitments on demonetisation, bots, fake accounts, impersonation, deepfakes and recommender integrity	Platform / distribution	Governance / assurance	EU Code of Conduct on Disinformation integrated under DSA Article 45	Adhering platforms, search engines, advertising actors, fact-checkers and civil society; overseen by Commission and Digital Services Board	European Union	Voluntary / private	Operational voluntary	Integrated into DSA framework in 2025; commitments auditable since 2025-07-01; first co-regulatory reports published 2026-03-24	Operational	Commit signatories to cut disinformation revenue, label and archive political ads, detect fake accounts and bot amplification, address impersonation and malicious deepfakes, support fact-checkers, improve recommender transparency and report service-level indicators.	Malicious actor → coordinated accounts and monetised amplification → platform audience → repeated exposure and trust erosion → agency transfer to network controller → concentrated narrative power.	Epistemic; Attentional; Relational	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Targets distribution infrastructure and coordinated behaviour rather than adjudicating every political claim.	Selective commitments, inconsistent enforcement, opaque prevalence denominators, migration to non-signatories and difficulty separating authentic mobilisation from coordination.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	CIB networks disrupted; fake-account prevalence; coordinated reach before removal; demonetised revenue; recommender exposure reduction; language coverage; independent audit exceptions.	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	No direct link	2026-08-11	null	Auditable platform commitments on demonetisation, bots, fake accounts, impersonation, deepfakes and recommender integrity	[ "M-03", "M-05", "M-08", "M-09", "M-10" ]	[ "C-01", "C-02", "C-05", "C-06" ]	[]	[ "CLM-0140", "CLM-0141", "CLM-0142" ]	[ "CLM-0142" ]	[ "CLM-0140" ]	[ "CLM-0141" ]	[ "SRC-032" ]	[ "P06" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-054	CF-054	[ "P4-054" ]	Time-bound election Rapid Response System for platform escalation	Cross-layer	Coordination / incident response	Code of Conduct on Disinformation election cooperation and Rapid Response System	EDMO, fact-checkers, civil society signatories and participating platforms	European Union and election-specific deployments	Voluntary / private	Operational voluntary	Operational since the 2024 European Parliament election and reused in later elections	Operational	Allow authorised non-platform signatories, fact-checkers and civil society to rapidly flag severe, time-sensitive content, accounts or trends and discuss response with platforms under their policies.	Late-breaking synthetic or coordinated incident → rapid spread → no time for ordinary review → belief, turnout or legitimacy shock → temporary agency transfer; an escalation lane compresses response time.	Epistemic; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Creates a trusted, fast channel for the final hours when ordinary notice systems are too slow.	Platform policy still controls action; privileged flaggers may miss local or minority-language incidents; false positives and government pressure can chill speech; no common service-level obligation.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Acknowledgement and disposition time; reach accrued before action; escalation acceptance rate; false-positive and appeal outcomes; coverage by country, language, platform and modality.	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	No direct link	2026-08-11	null	Election rapid response and escalation	[ "M-08", "M-10", "M-12", "M-13", "M-14" ]	[ "C-01", "C-02", "C-05" ]	[ "G-08", "G-12" ]	[ "CLM-0143", "CLM-0144", "CLM-0145" ]	[ "CLM-0145" ]	[ "CLM-0143" ]	[ "CLM-0144" ]	[ "SRC-033" ]	[ "P06", "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-055	CF-055	[ "P4-055" ]	European Centre for Democratic Resilience and reinforced election coordination	Institutional / structural	Coordination / incident response	European Democracy Shield institutional mechanism	European Commission, participating Member States, candidate countries, EU bodies and stakeholder platform	European Union and participating candidate countries	Voluntary / private	Operational voluntary	Launched 2026-02-24; operational design still maturing	Operational	Pool Member State, EU and civil-society expertise to anticipate, detect and respond to FIMI and disinformation; connect election networks, DSA data access, political-ad data and operational response.	Cross-border actor → fragmented national signals → delayed attribution and response → accumulating trust and agenda effects → agency transfer to hostile actor; pooled awareness reverses the asymmetry.	Epistemic; Relational; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Addresses cross-platform and cross-border coordination gaps and can turn dispersed signals into a common operational picture.	Mandate overlap, politicised attribution, intelligence-to-public-evidence gaps, centralisation without accountability and slow information sharing.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Shared alerts and lead time; Member State participation; joint analyses; cross-platform incident resolution; public attribution quality; exercise and after-action improvements.	Monitor / shape	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	No direct link	2026-08-11	null	European Centre for Democratic Resilience and reinforced election coordination	[ "M-05", "M-10", "M-12", "M-13", "M-14" ]	[ "C-01", "C-02" ]	[]	[ "CLM-0146", "CLM-0147", "CLM-0148" ]	[ "CLM-0148" ]	[ "CLM-0146" ]	[ "CLM-0147" ]	[ "SRC-034" ]	[ "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-056	CF-056	[ "P4-056" ]	Mandatory digital imprints on paid political material and specified organic campaign material	Electoral operations	Law / regulation	Elections Act 2022 Part 6 and Electoral Commission statutory guidance	Campaigners, candidates, parties, police and Electoral Commission	United Kingdom	Binding	Applicable	In force since November 2023; guidance updated in 2026	Operational	Require covered digital material, including social posts, online ads, podcasts and publicly distributed messaging-app content, to state who promoted and on whose behalf it was published.	Campaign or third party → unattributed digital communication → voter → mistaken source trust → agency transfer to hidden promoter → opaque political power.	Epistemic; Decisional; Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Preserves contestability by letting voters and regulators identify the responsible campaign actor.	Organic material by unregistered actors has gaps; private messages and cross-border actors are hard to enforce; an imprint does not disclose targeting, spend, AI use or beneficial funder.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Imprint compliance and readability; promoter traceability; enforcement cases; unattributed reach; cross-platform persistence; proportion of campaign communication outside scope.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-11	null	Mandatory digital imprints on paid political material and specified organic campaign material	[ "M-03", "M-05", "M-11", "M-13", "M-14" ]	[ "C-06" ]	[ "G-07" ]	[ "CLM-0149", "CLM-0150", "CLM-0151" ]	[ "CLM-0151" ]	[ "CLM-0149" ]	[ "CLM-0150" ]	[ "SRC-035" ]	[ "P07" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-057	CF-057	[ "P4-057" ]	Systemic platform duties for criminal foreign-interference content	Cross-layer	Law / regulation	National Security Act 2023 plus Online Safety Act 2023	Online services and Ofcom, with police and national-security bodies	United Kingdom	Binding	Applicable	Illegal-content duties in force from 2025-03-17	Operational	Treat the National Security Act foreign-interference offence as a priority offence, requiring in-scope services to assess risk, prevent users encountering illegal material and minimise its availability; Ofcom can impose major fines and disruption measures.	Foreign power or proxy → covert online conduct → UK voters and institutions → trust, behaviour or legitimacy change → agency transfer abroad → democratic harm.	Decisional; Relational; Collective / institutional	Legal / operational precedent	None	None — legal existence only	Established — legal status	Creates a systemic duty to act on covert hostile-state operations rather than relying only on voluntary moderation.	Most domestic misinformation and lawful manipulative persuasion remain outside scope; attribution to a foreign power is difficult; encryption and cross-border evidence constrain action; overreach risks political surveillance.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Foreign-interference risk assessments; detected networks; time-to-disruption; Ofcom enforcement; successful prosecutions; reach before removal; false attribution and appeal outcomes.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-11	null	Systemic platform duties for criminal foreign-interference content	[ "M-05", "M-09", "M-12", "M-14", "M-16" ]	[]	[ "G-08" ]	[ "CLM-0152", "CLM-0153", "CLM-0154" ]	[ "CLM-0154" ]	[ "CLM-0152" ]	[ "CLM-0153" ]	[ "SRC-036" ]	[]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-058	CF-058	[ "P4-058" ]	Election-authority deepfake detection and impact-monitoring pilot	Cross-layer	Law / regulation	Electoral Commission operational pilot with the Home Office Accelerated Capability Environment	UK Electoral Commission and ACE, with parties, candidates, platforms and police	Great Britain	Voluntary / private	Pilot	Pilot ran around the May 2026 elections; full findings due autumn 2026	Pilot	Monitor online audio and video for false candidate depictions or election-process claims, track reach, coordinate corrections or requested takedowns, and refer serious content to police or other bodies.	AI generator → candidate or authority impersonation → voter → false belief or suppressed participation → agency transfer to fabricator; detection and authoritative correction interrupt spread.	Epistemic; Decisional; Collective / institutional	Normative proposal	E0	None — mechanism or proposal only	Plausible hypothesis	Gives the election authority an early-warning picture and a route to authenticate candidate or process claims.	Detectors are brittle and can generate false positives; monitoring may miss closed messaging and low-resource languages; requests lack compulsory force; public alarms can amplify a fake.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Precision and recall against adjudicated samples; time-to-detection; reach before correction; candidate response time; correction penetration; referrals and platform outcomes.	Pilot	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	Pilot — 6–18 months	No direct link	No direct link	2026-08-11	null	Election-authority deepfake detection and impact-monitoring pilot	[ "M-08", "M-09", "M-11", "M-13", "M-14" ]	[ "C-03", "C-04", "X-01" ]	[ "G-09", "G-12", "G-13" ]	[ "CLM-0155", "CLM-0156", "CLM-0157" ]	[ "CLM-0157" ]	[ "CLM-0155" ]	[ "CLM-0156" ]	[ "SRC-037" ]	[ "P09" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-059	CF-045	[ "P4-059" ]	Statutory political-ad libraries and regulator/researcher access to platform exposure data	Cross-layer	Research / evaluation	Electoral Commission reform recommendation	Would require UK Parliament or government, platforms, Electoral Commission, Ofcom and ICO	United Kingdom	Proposed	Proposed / pending	Proposed; not a general statutory requirement as of 2026-08-11	Design-ready	Require every platform carrying political ads to maintain a comprehensive library with sponsor, targeting, actual reach and spend, and create meaningful data access for regulators and independent researchers.	Opaque platform ads and ranking → unobservable targeting and amplification → voter influence → agency transfer → no accountability; access makes the pathway measurable.	Attentional; Evaluative; Decisional; Relational	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Would close the UK's reliance on voluntary, inconsistent libraries and make algorithmic political reach auditable.	Privacy and trade-secret limits; platform exit from ads; organic influence remains hidden; data can arrive too late for election response.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Platform coverage; field completeness; API availability; actual versus reported reach; researcher access times; detected foreign or unlawful spending.	Research / hold	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	No direct link	2026-08-11	null	Political-ad transparency, targeting and repositories	[ "M-03", "M-05", "M-09", "M-12", "M-14" ]	[ "C-06" ]	[ "G-07" ]	[ "CLM-0158", "CLM-0159" ]	[ "CLM-0159" ]	[]	[ "CLM-0158" ]	[ "SRC-038" ]	[ "P07" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-060	CF-052	[ "P4-060" ]	Treat AI-cloned voices in robocalls as artificial voices requiring consent	Cross-layer	Policy / organizational control	FCC Declaratory Ruling under the TCPA	FCC, state attorneys general and telecom providers	United States	Binding	Applicable	In force; applied in the New Hampshire robocall matter	Operational	Apply TCPA consent and identification rules to AI-generated or cloned voices classified as artificial or prerecorded voices, subject to statutory exemptions.	Campaign operative → cloned candidate voice → mass calls → false instruction or turnout suppression → agency transfer to caller → electoral harm.	Multiple / context-dependent	Legal / operational precedent	None	None — legal existence only	Established — legal status	Uses mature robocall law to impose consent and enforcement costs on a high-risk impersonation vector.	Caller-ID spoofing, offshore operators, peer-to-peer messages and non-voice synthetic content can evade; consent exceptions and enforcement latency remain.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Complaint volume; blocked calls; trace-back success; time-to-enforcement; penalties collected; repeat offending; estimated voter reach.	Enforce now	Enforce legal scope; measure implementation and rights costs separately from legal existence.	0–6 months	No direct link	No direct link	2026-08-11	This is not a ban on every AI political message: prior express consent and exemptions depend on call purpose and context.	Synthetic impersonation and authenticity controls	[ "M-05", "M-08", "M-16" ]	[ "C-03", "X-01" ]	[]	[ "CLM-0160", "CLM-0161", "CLM-0162" ]	[ "CLM-0162" ]	[ "CLM-0160" ]	[ "CLM-0161" ]	[ "SRC-039" ]	[ "P09" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-061	CF-052	[ "P4-061" ]	Apply federal fraudulent-misrepresentation rules to AI that falsely claims campaign authority	Electoral operations	Law / regulation	FEC interpretive rule under 52 U.S.C. §30124 and 11 C.F.R. §110.16	Federal Election Commission	United States federal elections	Binding	Partial / verify	Interpretive rule adopted; broader AI rulemaking declined in 2024	Partial / uneven	Clarify that existing law can cover AI communications where a candidate or agent falsely purports to speak, act or fundraise on behalf of another candidate or party; it does not generally prohibit false depictions with a truthful disclaimer.	Candidate or agent → AI impersonation of rival campaign authority → voter or donor → source deception → agency transfer to impersonator → electoral or financial harm.	Epistemic; Decisional; Executive	Legal / operational precedent	None	None — legal existence only	Established — legal status	Protects authentic campaign authority and donations from a limited class of AI impersonation.	Leaves most candidate deepfakes, lies, issue ads and third-party content outside the rule; FEC deadlock and slow enforcement reduce real-time value.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Complaints and matters under review; time to disposition; proven authority impersonations; voter or donor losses; repeat incidents.	Monitor / shape	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	0–6 months	No direct link	Contextual capability link	2026-08-11	The FEC interpretation covers two narrow fraudulent-misrepresentation settings involving purported campaign authority; it is not a federal deepfake ban.	Synthetic impersonation and authenticity controls	[ "M-06", "M-08", "M-13", "M-16" ]	[ "C-03", "X-01" ]	[]	[ "CLM-0163", "CLM-0164", "CLM-0165" ]	[ "CLM-0165" ]	[ "CLM-0163" ]	[ "CLM-0164" ]	[ "SRC-040" ]	[ "P09" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-062	CF-052	[ "P4-062" ]	On-air and political-file disclosure of AI-generated content in broadcast political ads	Platform / distribution	Law / regulation	FCC Notice of Proposed Rulemaking, proceeding 24-211	Would be implemented by FCC-regulated broadcasters, cable and satellite entities	United States	Proposed	Proposed / pending	Proposed in 2024; no final rule identified as of 2026-08-11	Design-ready	Require radio and television political ads using AI-generated content to carry an on-air notice and a written disclosure in the station's political file.	Campaign → synthetic broadcast ad → mass audience → false authenticity belief → agency transfer to advertiser → electoral influence.	Epistemic; Evaluative	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Would make synthetic content visible in legacy mass media and create an inspectable record.	FCC jurisdiction largely excludes online and streaming ads; broad definitions may over-label routine editing; disclosure may not neutralise deception.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	If adopted: disclosure compliance, political-file completeness, audience comprehension, complaints and cross-channel coverage gaps.	Research / hold	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	No direct link	2026-08-11	The NPRM concerns radio and television political ads and political-file disclosure; it is not a final rule or a cross-platform online standard.	Synthetic impersonation and authenticity controls	[ "M-03", "M-08", "M-09", "M-13", "M-16" ]	[ "C-03", "X-01" ]	[]	[ "CLM-0166", "CLM-0167" ]	[ "CLM-0167" ]	[]	[ "CLM-0166" ]	[]	[ "P09" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-063	CF-052	[ "P4-063" ]	Platform duty to remove or label materially deceptive election deepfakes during a defined election window	Platform / distribution	Law / regulation	California Defending Democracy from Deepfake Deception Act, AB 2655	Large platforms and California enforcement authorities	California, United States	Binding — limited scope	Contested / enjoined	California AB 2655 was enacted but permanently enjoined as applied to the successful plaintiffs; appeal remained pending in 2026.	Partial / uneven	California AB 2655 required large platforms to block specified deceptive election content or label it, provide reporting channels and respond within short timeframes.	AI creator → viral candidate or election deepfake → voter → false belief → agency transfer to creator → electoral harm; platform action attempts to stop distribution.	Epistemic; Decisional; Collective / institutional	Legal / operational precedent	None	Indirect / mixed	Established — legal status	The intended effect was rapid reduction of deceptive synthetic reach; the actual contribution is a warning that broad platform-liability designs can collapse under intermediary-liability and speech law.	Over-removal of satire and political speech, vague deception thresholds, prior-restraint concerns, Section 230 pre-emption and strategic reporting abuse.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	If redesigned: adjudicated precision, removal latency, restored-content rate, satire/parody false positives, reach prevented, litigation survival.	Monitor / shape	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	No direct link	2026-08-11	null	Synthetic impersonation and authenticity controls	[ "M-08", "M-09", "M-13", "M-14", "M-16" ]	[ "C-03", "C-04", "X-01" ]	[ "G-09", "G-14" ]	[ "CLM-0168", "CLM-0169", "CLM-0320" ]	[ "CLM-0169" ]	[ "CLM-0320" ]	[ "CLM-0168" ]	[ "SRC-042" ]	[ "P09" ]	false	false	false	false	[]	[]	Provisional — legal status not claim-checked	research_preview
I-064	CF-029	[ "P4-064" ]	End-to-end cryptographic provenance and interoperable Content Credentials	Platform / distribution	Technical standard / product control	C2PA technical standard, complemented by NIST synthetic-content guidance	Model providers, camera and editing vendors, newsrooms, campaigns and platforms	Global, voluntary technical ecosystem	Voluntary / private	Operational voluntary	C2PA 2.2 published May 2025; adoption expanding but incomplete	Operational	Bind signed assertions about origin and edit history to media, preserve or recover credentials across transformations, and display provenance at consumption.	Creator or model → media asset → platform and voter → uncertainty about source and edits → exploitable trust asymmetry; provenance restores verifiable history and contestability.	Epistemic; Decisional; Relational	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Authenticates known-good media and carries origin information across the supply chain, stronger than unreliable deepfake detection alone.	Credential stripping, screenshots, unsigned legacy content, compromised signing keys, uneven adoption and the false inference that uncredentialed means fake.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Signing adoption by model, newsroom and campaign; credential survival across platforms; validation success; key compromise rate; user recognition; coverage of viral election media.	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	No direct link	2026-08-11	null	Synthetic-output provenance and disclosure	[ "M-05", "M-08", "M-09", "M-13", "M-15" ]	[ "C-03", "C-04", "X-01" ]	[ "G-09", "G-12" ]	[ "CLM-0170", "CLM-0171", "CLM-0172" ]	[ "CLM-0172" ]	[ "CLM-0170" ]	[ "CLM-0171" ]	[ "SRC-043" ]	[ "P09" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-065	CF-065	[ "P4-065" ]	Election-specific human-rights risk assessment for digital platforms	Platform / distribution	Governance / assurance	UNESCO Guidelines for the Governance of Digital Platforms	States, independent regulators and digital platforms	Global normative framework	Voluntary / private	Operational voluntary	Published and being used in implementation initiatives	Operational	Require or encourage platforms to perform an election-cycle risk assessment, engage election administrators and civil society, assess design changes and take enhanced measures during elections and crises.	Platform product decision → altered amplification or access → electoral public → rights and agency effects → power concentration; assessment inserts rights review before deployment.	Attentional; Decisional; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Offers a rights-based baseline for countries without a DSA-style regime and frames elections as a heightened-risk period.	Soft-law compliance, regulator dependence, government capture and limited remedies; generic assessments may miss local languages, messaging apps and AI agents.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Publication and independence of assessments; stakeholder participation; mitigation closure; rights indicators; local-language coverage; post-election evaluation.	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	No direct link	2026-08-11	null	Human-rights and democratic-risk assessment	[ "M-05", "M-06", "M-09", "M-11", "M-14" ]	[ "C-02", "C-06" ]	[ "G-06", "G-07", "G-12", "G-14" ]	[ "CLM-0173", "CLM-0174", "CLM-0175" ]	[ "CLM-0175" ]	[ "CLM-0173" ]	[ "CLM-0174" ]	[ "SRC-044" ]	[ "G0", "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview
I-066	CF-066	[ "P4-066" ]	Political-party code for ethical digital campaigning and AI use	Electoral operations	Policy / organizational control	International IDEA Code of Conduct for the 2024 European Parliament Elections	Political parties, with election bodies and civil-society monitoring	European and adaptable globally	Voluntary / private	Partial / verify	Signed by European political parties in 2024; model available for national adaptation	Partial / uneven	Secure public commitments by parties not to create, use or disseminate deceptive synthetic content or manipulative campaign techniques, to label AI use, protect cybersecurity and disclose campaign practices.	Party or campaign → AI-enabled content or targeting → electorate → preference or trust change → agency transfer to campaign → electoral advantage; party commitments constrain the controller directly.	Evaluative; Relational; Collective / institutional	Normative proposal	None	Indirect / mixed	Plausible hypothesis	Reaches party behaviour that platform rules cannot fully observe and establishes a public norm against agency-undermining tactics.	No enforcement, selective sign-up, third-party proxies, vague AI definitions and incentives to defect in close races.	Necessity, proportionality, privacy, speech, competition, and access require context-specific review.	Party coverage; disclosed AI uses; independently verified violations; corrective action; proxy spending; voter awareness of commitments.	Implement now	Decision follows legal maturity, causal plausibility, reversibility and current evidence—not a composite score.	6–18 months	No direct link	Contextual capability link	2026-08-11	null	Political-party code for ethical digital campaigning and AI use	[ "M-03", "M-05", "M-08", "M-14", "M-16" ]	[ "C-02", "C-06", "X-01" ]	[ "G-07", "G-14" ]	[ "CLM-0176", "CLM-0177", "CLM-0178" ]	[ "CLM-0178" ]	[ "CLM-0176" ]	[ "CLM-0177" ]	[ "SRC-045" ]	[ "P13" ]	false	false	false	false	[]	[]	Plausible hypothesis	research_preview

End of preview. Expand in Data Studio

Agency Transfer Policy Atlas

A causal evidence map of interventions against AI-mediated manipulation.

This is the Part 4 artifact of From Persuasion to Agency Transfer. It maps where an intervention could interrupt the chain from AI capability and control to concentrated influence and democratic harm, who could implement it, and what is known about its legal status, mechanism, effect evidence, failure modes, rights risks, maturity, and evaluability.

Status

v0.1.0-beta.2 is a research preview, not a validated policy benchmark. The train split name is only the Hugging Face Dataset Viewer container; these records are not intended as model-training examples.

The release has 68 control families, 118 implementations, 320 atomic claims, 123 sources, 16 mechanisms, 15 candidate policy packages, and 24 non-compensable decision gates. It has no composite effectiveness score or leaderboard.

Critical evidence boundary

The existence of a law, the plausibility of a mechanism, and evidence that an intervention works are different claims. In this preview:

22 of 123 source records participate in at least one checked claim–source relation; the other 101 remain candidate source records;
56 duplicate claim–source rows are removed from the generated release;
6 of 118 control-effect claims have a checked empirical source and a bounded observed endpoint; the other 112 remain unchecked;
of the six priority implementations, three are now strong inference, two are open questions, and only technique-recognition prebunking retains an established component-effect classification;
established legal-status rows without a checked legal claim are also exposed as provisional in the default atlas view; 25 of 53 currently have a checked primary-legal claim;
the two project-mechanism rows are provisional until their mechanism claims receive claim-specific source verification.

The dataset therefore supports mapping, audit, and research-priority setting. It does not support claims that the listed controls reduce democratic harm.

Configurations

atlas is the default denormalized view, one row per implementation. The other configurations expose the normalized objects and bridge tables. Parquet keeps multi-value ID fields as typed lists. CSV companions remain in the repository for transparent diffs.

Responsible use

Use the Atlas to compare intervention points, trace claims to sources, identify coverage gaps, and design bounded evaluations. Do not use a decision tier as an effectiveness estimate. Do not infer vote effects from reach, virality, or case linkage. The dataset contains no targeting profiles, current-campaign playbooks, safeguard-bypass prompts, personally identifiable data, or raw harmful model outputs.

Licensing and citation

CC BY 4.0 covers project-authored taxonomy, annotations, relations, and metadata. It does not relicense linked statutes, papers, reports, or other third-party material. No DOI is minted for this beta. Cite the repository and version using CITATION.cff until a stable, independently reviewed release.

Canonical repository: https://github.com/apolmig/agencytransfer/tree/main/policy-atlas

Immutable dataset version: https://huggingface.co/datasets/apol/agency-transfer-policy-atlas/tree/v0.1.0-beta.2

Downloads last month: 70

Collection including apol/agency-transfer-policy-atlas

Agency-Transfer

Collection

agency transfer in relation to election security, harmful manipulation, persuasiveness and AI epistemic risks • 4 items • Updated 4 days ago