Can the model preserve what the provision actually says?
Five real provisions, structured before relying on automated extraction. Inspect the actor, force, scope and exceptions—and the points where the current schema needs a decision.
AI-assisted annotation 2026-10-11 Human review pending IR 0.1.0 · no parser run
Read primary passages and surrounding qualifiers; structure candidates without calling the extraction pipeline. These are proposed readings, not expert gold labels.
Five purposively selected provisions, not representative coverage or a parser evaluation. EU examples read the original 2024 text, not current consolidated law or applicability to a person. Source hashes bind the curated excerpts, not complete issuer files. No requirements, interpretations or controls are approved.
Three legal bundles produce four unaccepted candidate requirements. Two voluntary readings are held outside the IR. Passing structural checks is not expert validation.
Case 01 · Candidate only
A retention duty with an exception
The duty fits the IR, but free-text conditions and exceptions do not themselves enforce its boundary. Applicability and the governing retention law remain unresolved.
Regulation (EU) 2024/1689 — original text
Article 26(6), first subparagraph; second subparagraph retained as context
Original Official Journal text, 2024-07-12 · checked 2026-10-11
Inspect source excerpts and context — eu-log-retention
Article 3 actor definitions, Article 6 and Annex III classification, Articles 2 and 113 scope/application, Article 12 logging and applicable data-protection/financial-services law need separate review. This excerpt alone does not establish applicability.
art-26/para-6/subpara-1
Deployers of high-risk AI systems shall keep the logs automatically generated by that high-risk AI system to the extent such logs are under their control, for a period appropriate to the intended purpose of the high-risk AI system, of at least six months, unless provided otherwise in applicable Union or national law, in particular in Union law on the protection of personal data.
art-26/para-6/subpara-2
Deployers that are financial institutions subject to requirements regarding their internal governance, arrangements or processes under Union financial services law shall maintain the logs as part of the documentation kept pursuant to the relevant Union financial service law.
retain controlled logs
Actor
Deployers of high-risk AI systems
Speech act
obligation
Action
Keep automatically generated logs
Object
Logs automatically generated by the high-risk AI system
Counterparty
Not separately annotated
Timing
A period appropriate to the intended purpose, of at least six months, subject to the legal exception.
Conditions
The actor is a deployer of a high-risk AI system within the Regulation's applicable scope.
Only to the extent the logs are under the deployer's control.
Exceptions
Applicable Union or national law provides otherwise, particularly personal-data protection law.
Formalization boundary: blocked · Only the first subparagraph is annotated as a duty. The financial-institution documentation duty is retained but not extracted; no completeness claim. Do not replace 'at least six months' with a universal deletion deadline.
Unresolved question · blocks compilation
Which retention law, intended purpose and control boundary govern the deployment, including any financial-services documentation duties?
Who must resolve it: Qualified legal reviewer with the deployer's records and data-protection owners
The external law and organizational facts required for operational retention are not supplied. This is an unresolved dependency, not a claim that the statute is inherently ambiguous.
No other applicable law changes the baseline: Choose and justify a purpose-appropriate period not shorter than six months; do not assume six months is always sufficient.
Other applicable law supplies a different rule: Record that authority and review the combined duties before configuring retention.
Inspect proposed controls and boundary scenarios — eu-log-retention
Editorial design proposals, not adopted controls or executable entitlement tests. No policy decision has been approved.
Possible implementation · retain-controlled-logs
Maintain a scoped log inventory and a versioned retention schedule recording the controlling authority, purpose, period and access arrangements. Do not configure a universal deletion timer from this excerpt.
Owner: Deployer's records owner, with legal and data-protection review
Evidence to examine
Log inventory and evidence of control
Approved retention rationale and legal dependency record
Retention configuration and sampled preservation/deletion events
What this cannot establish: Configuration evidence cannot establish legal applicability, completeness of the log inventory or the effectiveness of review.
Boundary scenarios for a reviewer
A log is outside the deployer's control
Do not infer this retention duty covers that log; assess other duties separately.
A controlled log is five months old and no legal exception changes the baseline
The proposed minimum-period interpretation does not permit deletion yet.
A controlled log reaches six months
No automatic deletion authorization: purpose-appropriateness and other applicable law still need review.
A financial institution is the deployer
Review the second subparagraph and relevant financial-services documentation rules; the first annotated duty is not exhaustive.
Inspect annotation JSON — eu-log-retention
Case 02 · Candidate only
A complaint right and an authority's duty
Two directed statements are needed: a complainant's right and an authority's handling duty. A complaint does not entail compensation, a successful challenge or a response deadline.
Regulation (EU) 2024/1689 — original text
Article 85, both sentences
Original Official Journal text, 2024-07-12 · checked 2026-10-11
Inspect source excerpts and context — eu-complaint-route
Regulation (EU) 2019/1020 and the authority's dedicated procedures are external dependencies. Scope and application must be assessed separately. 'Any natural or legal person' is not confined to people personally affected by a decision.
art-85/sentence-1
Without prejudice to other administrative or judicial remedies, any natural or legal person having grounds to consider that there has been an infringement of the provisions of this Regulation may submit complaints to the relevant market surveillance authority.
art-85/sentence-2
In accordance with Regulation (EU) 2019/1020, such complaints shall be taken into account for the purpose of conducting market surveillance activities, and shall be handled in line with the dedicated procedures established therefor by the market surveillance authorities.
submit complaint
Actor
Any natural or legal person
Speech act
right
Action
Submit a complaint to the relevant market surveillance authority
Object
An alleged infringement of the Regulation
Counterparty
Relevant market surveillance authority
Timing
No deadline or period inferred
Conditions
The person has grounds to consider that there has been an infringement of the Regulation.
Exceptions
None stated in this reading; not proof that none exist elsewhere.
Formalization boundary: procedurally-verifiable · The heading explicitly names a right; the sentence says 'may'. This provisional annotation uses right rather than permission and needs human review. Other administrative and judicial remedies remain unaffected. No personal-impact condition is added.
Unresolved question · blocks compilation
Which authority and submission procedure are relevant in the jurisdiction and matter concerned?
Who must resolve it: Competent authority and qualified jurisdictional reviewer
The competent authority and intake procedure cannot be identified from this sentence alone.
The designated authority and intake channel are identified: Expose the applicable channel, recording source and procedure version.
Competence or routing remains unresolved: Escalate for routing review; do not invent an authority or reject the complaint as ineligible.
handle complaint
Actor
Market surveillance authorities
Speech act
obligation
Action
Take complaints into account for market surveillance and handle them under the dedicated procedures
Object
Complaints submitted under Article 85
Counterparty
Complainant
Timing
No deadline or period inferred
Conditions
In accordance with Regulation (EU) 2019/1020.
Follow the dedicated procedures established by the market surveillance authorities.
Exceptions
None stated in this reading; not proof that none exist elsewhere.
Formalization boundary: blocked · An authority duty is separate from the complainant's right. No response deadline or compensation obligation is present in this excerpt.
Unresolved question · blocks compilation
Which version of the dedicated procedure governs handling, and what records demonstrate it was followed?
Who must resolve it: Market surveillance authority's procedure owner
The procedure, routing and evidential steps are not included in the source bundle.
A governing procedure is available: Map its steps and evidence without inventing a deadline from Article 85.
The governing procedure is not available: Keep the handling control blocked pending the external procedure.
Inspect proposed controls and boundary scenarios — eu-complaint-route
Editorial design proposals, not adopted controls or executable entitlement tests. No policy decision has been approved.
Possible implementation · handle-complaint
A complaint intake and handling procedure that records routing, consideration and disposition against the governing procedure. Accessibility and assisted submission should be assessed as design proposals, not silently attributed to this sentence.
Owner: Relevant market surveillance authority
Evidence to examine
Procedure version and competence/routing record
Complaint receipt and consideration record
Disposition or escalation record
What this cannot establish: A receipt proves intake, not meaningful consideration, effective remedy or compensation.
Boundary scenarios for a reviewer
A legal person alleges an infringement but is not personally subject to an AI decision
Do not add Article 86's personal-impact condition to Article 85's complaint route.
A complaint is received and logged
A receipt alone does not demonstrate the handling duty was fulfilled or that compensation follows.
No response deadline is supplied in this excerpt
Do not manufacture a 30-day deadline; obtain the dedicated procedure and other applicable law.
Inspect annotation JSON — eu-complaint-route
Case 03 · Candidate only
An explanation right with a narrow scope
A right fits, but the IR has no typed relationship between right-holder and duty-bearer and no executable representation of cumulative scope or legal overrides. Preserve them in fields and review notes rather than flattening them.
Regulation (EU) 2024/1689 — original text
Article 86(1)–(3), read together
Original Official Journal text, 2024-07-12 · checked 2026-10-11
Inspect source excerpts and context — eu-explanation-right
Article 6, Annex III, Union/national law and overlapping Union rights must be examined. Missing a condition here does not establish that no explanation right exists under other law.
art-86/para-1
Any affected person subject to a decision which is taken by the deployer on the basis of the output from a high-risk AI system listed in Annex III, with the exception of systems listed under point 2 thereof, and which produces legal effects or similarly significantly affects that person in a way that they consider to have an adverse impact on their health, safety or fundamental rights shall have the right to obtain from the deployer clear and meaningful explanations of the role of the AI system in the decision-making procedure and the main elements of the decision taken.
art-86/para-2
Paragraph 1 shall not apply to the use of AI systems for which exceptions from, or restrictions to, the obligation under that paragraph follow from Union or national law in compliance with Union law.
art-86/para-3
This Article shall apply only to the extent that the right referred to in paragraph 1 is not otherwise provided for under Union law.
obtain explanation
Actor
Any affected person subject to the decision
Speech act
right
Action
Obtain clear and meaningful explanations from the deployer
Object
The AI system's role in the decision-making procedure and the main elements of the decision taken
Counterparty
Deployer taking the decision
Timing
No deadline or period inferred
Conditions
The deployer takes a decision on the basis of output from a high-risk AI system listed in Annex III.
The decision produces legal effects or similarly significantly affects the person.
The person considers the impact adverse to their health, safety or fundamental rights.
The Article applies only to the extent the right is not otherwise provided under Union law.
Exceptions
Systems listed under Annex III point 2 are excluded.
Exceptions or restrictions following from Union or national law in compliance with Union law.
Formalization boundary: judgment-required · The right-holder is not the operator of the proposed control. Keep the deployer counterpart explicit in this annotation sidecar. Neither a remedy outcome nor a deadline is inferred. All three paragraphs support the reading.
Unresolved question · blocks compilation
After applicability and legal overrides are resolved, what evidence and review establish that this explanation is clear and meaningful?
Who must resolve it: Qualified legal reviewer and deployer's explanation owner, informed by affected-person testing
Adequacy is not settled by an explanation being delivered or its fields passing a machine check. The applicable legal standard and affected person's context need review.
A template delivers the named explanation elements: Check delivery and coverage but do not treat completeness alone as proof of meaningfulness.
Contextual human review also examines intelligibility: Add review and comprehension evidence while leaving the legal adequacy judgment with the authorized reviewer.
Inspect proposed controls and boundary scenarios — eu-explanation-right
Editorial design proposals, not adopted controls or executable entitlement tests. No policy decision has been approved.
Possible implementation · obtain-explanation
An applicability-aware request workflow preserving the decision record, AI contribution and principal decision elements, with human review of intelligibility. Do not automatically deny other explanation rights when this Article's conditions are unmet.
Owner: Deployer's explanation owner with qualified legal review
Evidence to examine
Applicability and legal-override review record
Versioned explanation and delivery record
Human adequacy review and affected-person feedback
What this cannot establish: Delivery and schema completeness cannot prove that an explanation is legally sufficient, meaningful or practically usable.
Boundary scenarios for a reviewer
An Annex III point 2 system is used
Do not attribute the explanation right to Article 86; assess any parallel rights separately.
An ordinary AI suggestion has no legal or similarly significant effect
Do not infer Article 86 entitlement from AI involvement alone.
Paragraph 1 appears satisfied but an applicable legal restriction is unresolved
Keep applicability unresolved rather than granting or denying a right automatically.
All explanation fields are populated, but the response is incomprehensible to its recipient
Do not label the explanation meaningful on field completeness alone.
Inspect annotation JSON — eu-explanation-right
Case 04 · Translation held
A voluntary framework's recommended outcome
The source describes a desired organizational outcome, not a statutory duty. The four IR modalities cannot represent this without changing its force. Hold translation rather than relabeling it obligation.
NIST AI Risk Management Framework (AI RMF 1.0)
GOVERN 2.1, printed page 23 / PDF page 28; voluntary framing in executive summary
Inspect source excerpts and context — nist-govern-roles
The RMF is voluntary. Adoption in a contract, policy or law could create a separate obligation, but no such adoption instrument is supplied here. GOVERN 2.2–2.3 on training and leadership remain outside this single-outcome annotation.
govern-2-1
Roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented and are clear to individuals and teams throughout the organization.
document roles
Actor
Organization using the framework
Speech act
framework-outcome
Action
Document roles, responsibilities and communication and make them clear across the organization
Object
Mapping, measuring and managing AI risks
Counterparty
Not separately annotated
Timing
No deadline or period inferred
Conditions
Voluntary framework use; no binding adoption instrument is supplied.
Exceptions
None stated in this reading; not proof that none exist elsewhere.
Formalization boundary: procedurally-verifiable · The organization is an inferred subject, not an expressly named legal duty-bearer. Documentary presence and staff clarity are different checks. This reading stays outside canonical Requirement objects until the speech-act/force issue is resolved.
Inspect proposed controls and boundary scenarios — nist-govern-roles
Editorial design proposals, not adopted controls or executable entitlement tests. No policy decision has been approved.
Possible implementation · document-roles
Versioned role and escalation map with checks that relevant people understand it. This is an adoption-dependent design idea, not a legal mandate extracted from the RMF.
Owner: Organization's AI risk lead, if it chooses to adopt the outcome
Evidence to examine
Role and communication map
Ownership and escalation review
Sampled staff understanding checks
What this cannot establish: A complete document does not establish practical authority, independence or staff understanding.
Boundary scenarios for a reviewer
No contract, law or internal policy adopts the framework
Do not convert the outcome into a binding obligation.
A role chart exists but staff do not know whom to contact
Documentation alone does not establish the recommended clarity outcome.
Inspect annotation JSON — nist-govern-roles
Case 05 · Translation held
A context-sensitive recommendation
'Should' is recommendatory, not permission or obligation. The IR lacks a recommendation modality and a typed force/adoption relationship. Preserve the recommendation as a sidecar reading while testing a schema extension.
OECD AI Principles — Accountability, Principle 1.5
Principle 1.5, traceability paragraph (short opening excerpt); full paragraph read for context
Inspect source excerpts and context — oecd-traceability
The full paragraph connects traceability across data, processes and lifecycle decisions to examining outputs and answering inquiries. It qualifies the practice by context and the state of the art. The short quote is an anchor, not the complete recommendation.
principle-1-5/traceability-excerpt
AI actors should ensure traceability, including in relation to datasets, processes and decisions
ensure traceability
Actor
AI actors
Speech act
recommendation
Action
Ensure traceability enabling output analysis and responses to inquiry
Object
Datasets, processes and decisions across the AI system lifecycle
Counterparty
Not separately annotated
Timing
No deadline or period inferred
Conditions
Appropriate to context and consistent with the state of the art.
Exceptions
None stated in this reading; not proof that none exist elsewhere.
Formalization boundary: judgment-required · The quote verifies the anchor only; full-paragraph context supports the qualifiers in this proposed reading. No statutory duty, fixed record format or retention period is inferred.
Inspect proposed controls and boundary scenarios — oecd-traceability
Editorial design proposals, not adopted controls or executable entitlement tests. No policy decision has been approved.
Possible implementation · ensure-traceability
A context-specific provenance and inquiry process that connects relevant lifecycle records to an authorized responder. Record the adoption rationale separately from the OECD recommendation.
Owner: Responsible AI actors, under an explicit adoption decision
Evidence to examine
Context and adoption rationale
Traceability records
Sample inquiry reconstruction and response
What this cannot establish: Record presence does not prove state-of-the-art adequacy, accountability or a legal duty to disclose all records.
Boundary scenarios for a reviewer
Traceability is represented as 'permission' because the source is not binding
Reject the force change: a recommendation is neither permission nor a mandatory duty.
A universal retention period is added
Reject the unsupported addition; preserve the context-sensitive recommendation.
Inspect annotation JSON — oecd-traceability
What must happen before trusting extraction?
Have a qualified reviewer assess these proposed readings, resolve how recommendations and adoption are represented, and measure scope preservation separately from quote matching. The parser already exists; this corpus tests the contract it should preserve.
Local check: npm run review:manual. Inspect canonical candidate exports with npm -w packages/spec run manual:check -- --json. See the repository review report for the open gaps and decision sequence.