{
  "$schema": "./schema.json",
  "id": "unesco-ai-ethics",
  "name": {
    "en": "UNESCO Recommendation on the Ethics of AI — AIO formalization",
    "ko": "유네스코 AI 윤리 권고 — AIO 정형화"
  },
  "sourceNorm": {
    "title": "Recommendation on the Ethics of Artificial Intelligence",
    "publisher": "United Nations Educational, Scientific and Cultural Organization (UNESCO)",
    "version": "Adopted by the General Conference at its 41st session, Paris, 23 November 2021; document code SHS/BIO/PI/2021/1; 141 paragraphs in 8 chapters, including 11 areas of policy action",
    "url": "https://unesdoc.unesco.org/ark:/48223/pf0000381137"
  },
  "vesMapping": [
    {
      "article": "Para. 26",
      "summary": "The choice to use an AI system, and which AI method to use, is to be justified as appropriate and proportional to a legitimate aim, must not violate or abuse human rights, and is to rest on rigorous scientific foundations; where a decision is understood to have an irreversible or hard-to-reverse impact, or may involve life and death, final human determination applies; and AI systems are not to be used for social scoring or mass surveillance.",
      "v": [
        "Sdt",
        "Sep"
      ],
      "e": [
        "Exp"
      ],
      "s": [],
      "status": "draft-verified",
      "obligationType": "mixed",
      "note": "The closest correspondence anywhere in this pack to the AIO 20002 record: the C: field already carries a reversibility axis (R / P / X) and a time horizon, which is the same trigger this paragraph uses to decide when final human determination applies. What an AIO item can test is whether a model treats an X-reversibility case as one where the human decides; what it cannot test is whether the operator has a documented method-selection justification, which is the paragraph's other limb. The social-scoring and mass-surveillance prohibition is paraphrased in the summary and not quoted, to keep quotation to the minimum the licence constraint allows; it is stated in the same paragraph. `Pee` is an inference — see the rationale.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 26 (Chapter III.2 Principles — Proportionality and Do No Harm)",
        "quote": "In scenarios where decisions are understood to have an impact that is irreversible or difficult to reverse or may involve life and death decisions, final human determination should apply.",
        "rationale": "The quoted sentence makes the human's own determination the thing that must prevail over whatever the system produced, which is `Sdt` (reaching the conclusion through one's own understanding rather than accepting a given answer). It names life and death, so `Sep`. The paragraph's second limb — the chosen AI method 'should not infringe upon the foundational values captured in this document, in particular, its use must not violate or abuse human rights' — gives `Unc`. What discharges the duty in the quoted scenario is the considered judgment of the person who makes the final determination (`Exp`); the paragraph's proportionality limbs (a) and (b) are argued from stated premises rather than from measurement, which is `Log`. On sources: the Recommendation is the instrument and Member States are its addressees (`Gov`). `Pee` is an INFERENCE — the paragraph requires the method to be 'based on rigorous scientific foundations', which designates scientific rigour but not, in terms, independent scholarly review; `Pee` is the nearest class in the AIO 00011 source vocabulary and is put to the RFC round on that point. `Rev` was considered for the same clause and rejected: the AIO code means a pooled synthesis across studies, which is narrower than what the paragraph says. ADJUDICATION 2026-08-14: V narrows to the intersection [Sdt, Sep] — the trigger conditions name bodily and personal safety (Sep) and the remedy, 'final human determination', keeps a human's own judgment terminal (Sdt) — while Unc rests on v0.1 alone. E narrows to [Exp], which both passes reached: what discharges the requirement is the considered judgment of the assigned human decider. S is emptied. The second pass recorded an INSUFFICIENT-QUOTE finding on this layer — the paragraph says human determination 'should apply' without naming who the human is or what standing they hold, and it withheld Pro despite the surface pull of 'determination' — and v0.1's Gov and Pee are both absent from the quoted sentence, Pee having been read from 'rigorous scientific foundations' elsewhere in the paragraph rather than from the excerpt. obligationType stays mixed against the second pass's behavioral, under the conservative rule.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "V [Sdt, Sep, Unc] → [Sdt, Sep]; E [Exp, Log] → [Exp]; S [Gov, Pee] → []; obligationType mixed retained against the second pass's behavioral. The removal of Pee discharges the first of v0.1 note 8's four flagged inferences."
    },
    {
      "article": "Para. 36",
      "summary": "Humans may choose to rely on AI systems for reasons of efficacy, but the decision to cede control in limited contexts remains the humans'; an AI system can never replace ultimate human responsibility and accountability; and as a rule life-and-death decisions are not to be ceded to AI systems.",
      "v": [
        "Sdt",
        "Sep",
        "Bed"
      ],
      "e": [
        "Exp"
      ],
      "s": [],
      "status": "draft-verified",
      "obligationType": "behavioral",
      "note": "The only entry in this pack classified `behavioral`. The paragraph is drafted impersonally and imposes no management-system duty in its words: its demand is discharged, or not, in the concrete decision about whether to let a system decide. That classification describes the shape of the duty, not its addressee — the Recommendation as a whole is addressed to Member States for voluntary application, and the pack-level `measurementScope` governs. Paragraph 35, in the same sub-section, carries the companion duty to make ethical and legal responsibility attributable to physical persons or existing legal entities and to provide for inclusive public oversight; it is addressed to Member States, has no separable judgment correlate, and is not mapped here. `Bed` is an inference — see the rationale.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 36 (Chapter III.2 Principles — Human oversight and determination)",
        "quote": "[…] an AI system can never replace ultimate human responsibility and accountability. As a rule, life and death decisions should not be ceded to AI systems.",
        "rationale": "The paragraph's operative move is that control may be delegated but responsibility may not: 'the decision to cede control in limited contexts remains that of humans'. The value that must prevail is therefore the human's own judgment about whether to decide at all (`Sdt`), and the named stake is life and death (`Sep`). 'Ultimate human responsibility and accountability' is coded `Bed` — being a reliable party who keeps the obligations they hold — as an INFERENCE: the AIO 00011 value layer, derived from Schwartz, has no dedicated accountability code, and `Bed` is the nearest class. This is put to the RFC round. Evidence: the paragraph accepts nothing except a human's own considered decision on the case in front of them (`Exp`); no metric, guideline or record is named, and none is coded. Sources: the paragraph designates no professional body, no scholarly source and no industry source, so only the instrument's own addressee is coded (`Gov`). `Pro` was specifically considered and rejected — the paragraph says 'humans', not 'qualified humans'. ADJUDICATION 2026-08-14: V ([Sdt, Sep, Bed]), E ([Exp]) and obligationType (behavioral) agreed exactly across the two independent passes. **This is the only `behavioral` entry surviving in either wave, and it survives by agreement rather than by adjudication** — two formalizations that could not see each other's work both read a paragraph drafted impersonally, imposing no management-system duty in its words, as discharged or not in the concrete decision about whether to let a system decide. Bed is no longer an inference: v0.1 flagged it as a stopgap for 'ultimate human responsibility and accountability', which has no dedicated code in the Schwartz-derived value layer, and the blind second pass reached it independently from the same words, so it survives as corroborated with the gap still recorded. S is emptied — the responsible human is unspecified and no source class is designated, which is the second pass's INSUFFICIENT-QUOTE finding and disposes of v0.1's Gov, read from the Recommendation's addressee rather than from this sentence. MODALITY, recorded and not coded: the second pass found, blind, that the two limbs differ in strength and must not be levelled — 'an AI system can never replace ultimate human responsibility' is categorical, while 'As a rule, life and death decisions should not be ceded to AI systems' is expressly defeasible and may not be formalized as a prohibition. The v/e/s layers cannot carry that difference; it is Wave 2 gap 23.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "V, E and obligationType agreed exactly — the only four-code-set entry in the wave to keep a `behavioral` tag, and it keeps it because both independent passes assigned it. S [Gov] → [] (the paragraph designates no source; v0.1's Gov came from the instrument's addressee, not from the quoted words). The defeasible modality of the 'as a rule' limb is recorded as a gap, not coded."
    },
    {
      "article": "Para. 38",
      "summary": "Transparency and explainability are to be increased throughout the life cycle, at a level always appropriate to the context and impact and balanced against principles such as privacy, safety and security; people are to be fully informed when a decision is informed by or made on the basis of AI algorithms, may request explanatory information from the relevant AI actor or public sector institution, are to be able to access the reasons for a decision affecting their rights and freedoms and to make submissions to a designated staff member able to review and correct it, and are to be told in a proper and timely manner when a product or service is provided with the assistance of an AI system.",
      "v": [
        "Sdt"
      ],
      "e": [
        "Log"
      ],
      "s": [],
      "status": "draft-verified",
      "obligationType": "mixed",
      "note": "An AIO 20002 record is not an explanation owed to an affected person. It is a one-line structured trace of which value, evidence and source classes governed a decision — machine-readable, deliberately carrying no verbatim user content — so it can evidence that a reasoning line existed and what shape it had, and it can be an input to the explanatory information this paragraph requires. It is not the 'meaningful explanation' itself, and the review-and-correction channel the paragraph requires (a designated staff member with authority to change the decision) is a management-system arrangement no item-based measurement reaches. The context-and-impact balancing clause — transparency against privacy, safety and security — is paraphrased in the summary rather than quoted, to keep quotation minimal.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 38 (Chapter III.2 Principles — Transparency and explainability)",
        "quote": "People should be fully informed when a decision is informed by or is made on the basis of AI algorithms, […] should have the opportunity to request explanatory information […]",
        "rationale": "The paragraph is written from the position of the person on the receiving end of a decision: they are informed, they may ask why, they may access the reasons, they may put the decision back to a human. That makes the protected value the affected person's capacity to reach their own view of the decision (`Sdt`), and the paragraph itself ties the trigger to decisions that affect 'their safety or human rights' (`Unc`). The duty to inform runs to a counterparty and is coded `Bed`. What discharges it is explanatory information — a reasoned account of how the decision came about (`Log`) — and, where the person makes submissions, the considered judgment of the designated staff member 'able to review and correct the decision' (`Exp`). The sources are named in the text and only those are coded: 'the relevant AI actor' (`Ind`) 'or public sector institutions' (`Gov`). `Dat` was considered and rejected: nothing in the paragraph makes a metric decisive. ADJUDICATION 2026-08-14: obligationType (mixed) agreed exactly, both passes reaching it by the same route — the informing lands in the individual case while a standing request channel is an organizational provision. V narrows to [Sdt], the one code both passes reached: being fully informed when a decision is made on the basis of AI algorithms protects the affected person's ability to understand their own situation. Unc and Bed rest on v0.1; the second pass's Sda, read from 'the opportunity to request explanatory information', rests on the second pass, and the Sdt/Sda boundary is Wave 2 gap 13. E narrows to [Log], which both passes reached and which both flagged the same way — 'explanatory information' is named without a designated form, and the second pass expressly considered and rejected Gui because no standard procedure or content list is specified. S is emptied: the paragraph is passive throughout ('People should be fully informed'), so the informing party is never designated, and v0.1's Ind and Gov are both read from outside the quoted words.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "obligationType agreed exactly. V [Sdt, Unc, Bed] → [Sdt]; E [Log, Exp] → [Log]; S [Ind, Gov] → []."
    },
    {
      "article": "Para. 40",
      "summary": "Explainability means making the outcome of an AI system intelligible and giving insight into it, including the understandability of the input, output and functioning of each algorithmic building block and how it contributes to the outcome; AI actors are to commit to developing explainable algorithms; and where an AI application's impact on the end user is not temporary, easily reversible or otherwise low risk, a meaningful explanation is to accompany any decision that resulted in the action taken, for the outcome to count as transparent.",
      "v": [
        "Sdt"
      ],
      "e": [
        "Log"
      ],
      "s": [],
      "status": "draft-verified",
      "obligationType": "mixed",
      "note": "This is the paragraph an AIO 20002 record comes closest to serving, and the place where over-reading is easiest. The record makes the value, evidence and source hierarchy behind a decision inspectable and traceable at scale; it is a self-reported account of the reasoning, not a causal trace of the computation, and it says nothing about 'the functioning of each algorithmic building block'. Treating a pack pass as evidence of explainability under this paragraph would be exactly the over-reading the pack's `measurementScope` forbids. Only `Log` is coded on the evidence layer: the paragraph names no metric, no guideline, no expert and no case series, and an undeclared layer is the honest signal that scoring does not reach it.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 40 (Chapter III.2 Principles — Transparency and explainability)",
        "quote": "In the case of AI applications that impact the end user in a way that is not temporary, easily reversible or otherwise low risk, […] the meaningful explanation is provided […]",
        "rationale": "The quoted sentence sets a risk-graded floor: below it, an outcome is not to be counted as transparent at all. The value the floor protects is the end user's ability to understand and therefore to judge the decision for themselves (`Sdt`), and the paragraph's own framing — 'outcomes and sub-processes leading to outcomes should aim to be understandable and traceable' — makes traceability of the system as such the second value (`Ses`). What discharges the duty is an explanation: a step-by-step account of how the input became the outcome (`Log`). No measurement, guideline, precedent or specialist opinion is named, so no other evidence class is declared. On sources, the paragraph names the party that owes the commitment — 'AI actors should commit to ensuring that the algorithms developed are explainable' (`Ind`) — under the recommendation Member States are asked to give effect to (`Gov`). ADJUDICATION 2026-08-14: E ([Log]) and obligationType (mixed) agreed exactly, and v0.1's decision to declare only Log — leaving the rest of the evidence layer empty because the paragraph names no metric, guideline, expert or case series — was independently confirmed. V narrows to [Sdt]: the meaningful explanation protects the end user's understanding of what was done to them. The passes diverged on the second value code, v0.1 reading the durable-impact scope gate as a collective-order interest (Ses) and the second pass as the end user's exposure to durable harm (Sep); neither is carried. S is emptied — no provider of the explanation is designated in the quoted text. RETRIEVAL LIMIT, recorded: the second pass noted that this excerpt is elided at both the subject and the operative verb, so the paragraph's modality could not be assessed from it. The excerpt is not widened, because the licence constraint recorded in the reuse note governs; the fuller stem is listed as a next-revision retrieval item.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "E and obligationType agreed exactly. V [Sdt, Ses] → [Sdt]; S [Ind, Gov] → []. The second pass could not assess modality from this excerpt because of the elision at the subject and verb; the excerpt is left as it is under the licence constraint and the fuller stem is recorded as a retrieval item for the next revision."
    },
    {
      "article": "Para. 43",
      "summary": "Appropriate oversight, impact assessment, audit and due diligence mechanisms — including whistle-blower protection — are to be developed to ensure accountability for AI systems and their impact throughout the life cycle, and both technical and institutional designs are to ensure auditability and traceability of the working of AI systems, in particular to address conflicts with human rights norms and standards and threats to environmental and ecosystem well-being.",
      "v": [
        "Bed"
      ],
      "e": [
        "Dat",
        "Gui"
      ],
      "s": [],
      "status": "draft-verified",
      "obligationType": "organizational",
      "note": "Organizational outright: mechanisms are developed, designs are built, whistle-blowers are protected — an AIO item observes none of that. What an AIO 20002 record contributes is one input to auditability: a grep- and SQL-friendly reasoning line per substantive decision, which an auditor can read at volume. It is not an audit, not a due diligence mechanism, and not a substitute for either. `Tes` rests on the whistle-blower clause of the same paragraph, which is outside the quoted excerpt; the excerpt was kept to the auditability sentence to hold quotation to the minimum the licence constraint allows.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 43 (Chapter III.2 Principles — Responsibility and accountability)",
        "quote": "Both technical and institutional designs should ensure auditability and traceability of (the working of) AI systems […]",
        "rationale": "The paragraph names its protected interests in its closing clause — 'any conflicts with human rights norms and standards and threats to environmental and ecosystem well-being' — which is `Unc` and `Unn`, the latter being the only place in this pack where the environmental value is stated inside a principle rather than a policy area. Auditability and traceability of the system as such is `Ses`. An audit runs on the accumulated operational record (`Dat`) read against a documented due diligence procedure (`Gui`). The source classes are those the paragraph designates: the addressee that develops the mechanisms (`Gov`), the operator whose own technical and institutional designs generate the audit trail (`Ind`), and — from 'including whistle-blowers' protection' — the on-record account of a specific named insider (`Tes`). `Pro` and `Pee` are not coded: the paragraph names no professional body and no scholarly source, however natural an external auditor sounds. ADJUDICATION 2026-08-14: E ([Dat, Gui]) and obligationType (organizational) agreed exactly. V is the largest divergence in the wave: the two sets were disjoint, v0.1 carrying [Ses, Unc, Unn] and the second pass [Bed, Cor], and an empty intersection is resolved by asking which reading is better grounded in the quoted words. It is the second pass's: 'auditability and traceability of (the working of) AI systems' is a requirement that the system's conduct remain answerable after the fact, which is Bed, and v0.1's Unn in particular is read from the paragraph's closing reference to environmental well-being, which is outside the excerpt. Cor is not carried with it, because the second pass flagged it as an inference — auditability read as adherence to formal procedure, the paragraph naming the property and not the value behind it — and an inference-flagged code survives only where both passes reached it. The result is a single value code. S is emptied: both passes independently withheld a source class, the second pass recording that 'auditability' invites an auditor class but the paragraph designates none, neither a professional body nor an authority, so v0.1's Gov, Ind and Tes all fall — Tes in particular rested on the whistle-blower clause of the same paragraph, which v0.1 itself recorded as lying outside the quoted excerpt. Both passes carried Dat with the same discomfort: traceability presupposes retained records but the paragraph declares no log content, metric or retention term (Wave 2 gap 12).",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "E and obligationType agreed exactly. V [Ses, Unc, Unn] → [Bed]: the two passes shared no value code, and the second pass's reading prevails as the one argued from the quoted sentence rather than from the rest of the paragraph. S [Gov, Ind, Tes] → [] — both passes independently found that the paragraph designates no source class despite the pull of 'auditability'."
    },
    {
      "article": "Para. 50",
      "summary": "Member States should introduce frameworks for impact assessments, such as ethical impact assessment, to identify and assess the benefits, concerns and risks of AI systems together with appropriate risk prevention, mitigation and monitoring measures; such assessments should identify impacts on human rights and fundamental freedoms — in particular but not limited to those of marginalized and vulnerable people or people in vulnerable situations — on labour rights and on the environment and ecosystems, together with ethical and social implications, and should facilitate citizen participation.",
      "v": [
        "Unc"
      ],
      "e": [
        "Gui"
      ],
      "s": [
        "Gov"
      ],
      "status": "draft-verified",
      "obligationType": "organizational",
      "note": "The opening provision of Policy Area 1, and the paragraph the whole Recommendation's implementation machinery hangs on: paragraph 131(a) has UNESCO develop an Ethical Impact Assessment (EIA) methodology, and paragraph 136 names that methodology as an instrument for giving effect to the Recommendation. Nothing in an AIO measurement tells an operator, or a Member State, whether an ethical impact assessment was performed, by whom, or with what result. The pack maps this paragraph so that the guide can address it, and so that the pack does not appear to cover assessment practice by omitting it. `Tri` is an inference — see the rationale. MEASUREMENT-SCOPE NOTICE (following the second pass's substance-versus-addressee convention, adopted at adjudication): the directed substance of this paragraph is the State's own apparatus, so its duty-bearer cannot be an AI system operator and nothing an AIO 20002 record supplies discharges it. No item can observe its performance. This pack asserts no measurement against this paragraph, and if the pack is ever scored this entry belongs outside the denominator rather than counted as unmet.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 50 (Chapter IV, Policy Area 1: Ethical impact assessment)",
        "quote": "Member States should introduce frameworks for impact assessments, such as ethical impact assessment, to identify and assess benefits, concerns and risks of AI systems […]",
        "rationale": "The paragraph enumerates what the assessment must look for, and the enumeration is the mapping: human rights and fundamental freedoms, with particular attention to marginalized and vulnerable people, gives `Unc`; 'the environment and ecosystems' gives `Unn`; 'risks' and the prevention and mitigation measures that answer them give `Sep`. What discharges the duty is a completed assessment run against an introduced framework, which is a written standard procedure (`Gui`). `Tri` is an INFERENCE from the closing clause 'facilitate citizen participation in line with the values and principles set forth in this Recommendation', read together with the paragraph's particular attention to people in vulnerable situations: participation of affected people admits their firsthand accounts as evidence. The paragraph does not say so in terms, so this is put to the RFC round. On sources, only the addressee is designated (`Gov`); the paragraph names no professional body, no scholarly source and no industry source. ADJUDICATION 2026-08-14: S ([Gov]) and obligationType (organizational) agreed exactly. Gov is retained under the substance-versus-addressee convention the second pass set out and this adjudication adopts for the Recommendation: where the State is directed to build an apparatus that will govern AI practice — here 'frameworks for impact assessments' — the addressee designates the governing authority and the classification is organizational; where the directed substance is a judgment in the concrete case, the addressee bears on nothing but S and obligationType. That convention is what distinguishes this entry from OECD Recommendations 2.3 and 2.5 in the parallel pack, where governments are addressed but no government instrument is designated as decisive over an AI practice. V narrows to [Unc], the only code both passes reached, and it is kept with the caveat the second pass attached: the paragraph names a procedure and its objects but no protected interest, so Unc is read from 'ethical impact assessment' as an assessment run on behalf of those affected, which follows from the instrument's structure rather than its words. E narrows to [Gui]; v0.1's Tri, read from 'facilitate citizen participation', was not reached by the second pass.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "S and obligationType agreed exactly. V [Unc, Unn, Sep] → [Unc]; E [Gui, Tri] → [Gui]. The removal of Tri discharges the third of v0.1 note 8's four flagged inferences. A measurement-scope notice is added under the substance-versus-addressee convention."
    },
    {
      "article": "Para. 55",
      "summary": "Member States should ensure that harms caused through AI systems are investigated and redressed, by enacting strong enforcement mechanisms and remedial actions, so that human rights and fundamental freedoms and the rule of law are respected in the digital and the physical world; such mechanisms should include remediation provided by private and public sector companies, auditability and traceability should be promoted to that end, and institutional capacity should be strengthened, in collaboration with researchers and other stakeholders, to investigate, prevent and mitigate potentially malicious uses.",
      "v": [
        "Unc",
        "Sep"
      ],
      "e": [
        "Cas"
      ],
      "s": [
        "Gov"
      ],
      "status": "draft-verified",
      "obligationType": "organizational",
      "note": "Redress is the clearest example in this pack of a duty an AIO measurement does not touch at any point. An operator can pass every item in a bank built on this pack and have no remediation channel at all. The reason the paragraph is mapped is the converse: an AIO 20002 record is retrospectively readable, so when a harm is investigated the record of what governed the decision is available to the investigation. That is an input to an investigation someone else runs, never the investigation. MEASUREMENT-SCOPE NOTICE (following the second pass's substance-versus-addressee convention, adopted at adjudication): the directed substance of this paragraph is the State's own apparatus, so its duty-bearer cannot be an AI system operator and nothing an AIO 20002 record supplies discharges it. No item can observe its performance. This pack asserts no measurement against this paragraph, and if the pack is ever scored this entry belongs outside the denominator rather than counted as unmet.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 55 (Chapter IV, Policy Area 2: Ethical governance and stewardship)",
        "quote": "Member States should ensure that harms caused through AI systems are investigated and redressed, by enacting strong enforcement mechanisms and remedial actions […]",
        "rationale": "The paragraph states the interests it protects immediately after the quoted words — 'human rights and fundamental freedoms and the rule of law' — which is `Unc` for the rights limb and `Cor` for the rule-of-law limb, with `Sep` for the harms that trigger the duty. An investigation of a specific harm proceeds by structured analysis of the instance and of comparable instances (`Cas`), on whatever operational record exists (`Dat`); no metric threshold and no guideline is named as the trigger. The sources are those the paragraph designates: the Member State that enacts the enforcement mechanisms (`Gov`), and the private and public sector companies whose own remediation mechanisms the paragraph says must be included (`Ind`). Researchers are named as collaborators in the paragraph's final sentence, but as parties to strengthen capacity rather than as a class of trusted source, so `Pee` is not coded. ADJUDICATION 2026-08-14: S ([Gov]) and obligationType (organizational) agreed exactly; Gov is retained under the substance-versus-addressee convention, the State being directed to enact enforcement mechanisms that will govern AI practice. V narrows to the intersection [Unc, Sep] — investigation and redress of harms protects the injured person (Sep) and states a justice-and-remedy interest owed to all who are harmed (Unc) — while Cor rests on v0.1 alone. E narrows to [Cas], which both passes reached from the investigative limb; both also flagged it, the second pass recording that the paragraph names the investigating act without designating its evidential form. v0.1's Dat and the second pass's Gui each rest on one reading. Ind is dropped: the actor in this excerpt is the Member State, so the Ind limb of the source-axis rule is not met.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "S and obligationType agreed exactly. V [Unc, Cor, Sep] → [Unc, Sep]; E [Cas, Dat] → [Cas]; S [Gov, Ind] → [Gov]. A measurement-scope notice is added under the substance-versus-addressee convention."
    },
    {
      "article": "Para. 71",
      "summary": "Member States should work to develop data governance strategies that ensure the continual evaluation of the quality of training data for AI systems, including the adequacy of the data collection and selection processes, proper data security and protection measures, and feedback mechanisms to learn from mistakes and share best practices among all AI actors.",
      "v": [],
      "e": [
        "Dat",
        "Gui"
      ],
      "s": [
        "Gov"
      ],
      "status": "draft-verified",
      "obligationType": "organizational",
      "note": "The opening provision of Policy Area 3. Data governance is a management system in the plainest sense — collection and selection processes, security controls, evaluation cadence — and an AIO measurement observes none of it. The pack maps it because the guide's data-governance category needs an anchor, and because a pack that skipped data policy entirely would understate how much of this Recommendation sits outside what can be measured. Paragraphs 72 to 77 (privacy safeguards and privacy impact assessment, data subject rights, security of sensitive data, open data, dataset quality and annotation standards, the digital commons approach) are not mapped; they are addressed in the management-system guide. MEASUREMENT-SCOPE NOTICE (following the second pass's substance-versus-addressee convention, adopted at adjudication): the directed substance of this paragraph is the State's own apparatus, so its duty-bearer cannot be an AI system operator and nothing an AIO 20002 record supplies discharges it. No item can observe its performance. This pack asserts no measurement against this paragraph, and if the pack is ever scored this entry belongs outside the denominator rather than counted as unmet.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 71 (Chapter IV, Policy Area 3: Data policy)",
        "quote": "Member States should work to develop data governance strategies that ensure the continual evaluation of the quality of training data for AI systems […]",
        "rationale": "Three values are read from the paragraph's own list. A governance strategy for the data an AI system is trained on is an ordering measure for the system and for what it does downstream (`Ses`). 'Proper data security and protection measures' is compliance with a formal requirement (`Cor`). The clause that carries the most weight and is the easiest to skip is 'feedback mechanisms to learn from mistakes': a strategy is required to presume that its own data selection can be wrong, which is `Hum` — recognising one's limits rather than overstating. 'Continual evaluation of the quality' is measurement against declared criteria (`Dat`), inside a documented strategy (`Gui`). The sources are the addressee that develops the strategy (`Gov`) and the 'all AI actors' among whom best practices are shared (`Ind`). No professional body, standards body or scholarly source is named, so none is coded. ADJUDICATION 2026-08-14: E ([Dat, Gui]) and obligationType (organizational) agreed exactly, and S is [Gov] once Ind is dropped for the reason given at ¶55 — the actor in this excerpt is the Member State. **The value layer is left undeclared, which is the first time in either wave that an entry carries no value code at all, and it is the honest outcome rather than an omission.** The two passes shared no value code: v0.1 read data-governance strategy as serving collective order, formal compliance and not overstating (Ses, Cor, Hum); the second pass read 'quality of training data' as competence against a standard (Ach) and attached an INSUFFICIENT-QUOTE caveat to its own code, recording that the paragraph designates no protected interest at all and that it deliberately withheld Unc because the quoted text does not mention bias, discrimination or fairness. An inference-flagged code that only one pass reached does not survive, and v0.1's three were not reached by the blind pass. What is left is a provision that names an obligation and no interest behind it — which is exactly Wave 1 gap 10 (undefined risk, found on NIST GOVERN 1.3 and MANAGE 1.1), here in its purest form, and the RFC question is the same: whether a provision that presupposes a protected interest without stating one may carry a value code at all.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "E, obligationType agreed exactly. V [Ses, Cor, Hum] → [] — the two passes shared no value code and the sole surviving candidate (the second pass's Ach) was flagged INSUFFICIENT-QUOTE by its own author. S [Gov, Ind] → [Gov]. The empty value layer is the pack's contribution to Wave 1 gap 10 and a measurement-scope notice is added under the substance-versus-addressee convention."
    },
    {
      "article": "Para. 86",
      "summary": "When choosing AI methods, given that some are data- or resource-intensive and bear on the environment accordingly, Member States should ensure that AI actors favour data-, energy- and resource-efficient AI methods in line with the principle of proportionality; requirements should be developed so that appropriate evidence is available to show that an AI application will have the intended effect, or that accompanying safeguards can support the justification for its use; where that cannot be done, the precautionary principle prevails, and where there are disproportionate negative impacts on the environment, AI should not be used.",
      "v": [
        "Unn"
      ],
      "e": [
        "Dat"
      ],
      "s": [
        "Gov",
        "Ind"
      ],
      "status": "draft-verified",
      "obligationType": "mixed",
      "note": "The sharpest trade-off in the pack, and the reason it is included: the paragraph asks for an efficient method to be preferred over a more capable one, and in the limit for the application not to be built. The values it expects to be deprioritized — demonstrated capability (`Ach`) and resource capture (`Por`) — are not coded, because the V layer records what must prevail; they belong on the left of the `<` in an AIO 20002 record, and an item on this paragraph is precisely an item about which side of that `<` the model puts efficiency on. The evidence requirement ('appropriate evidence… that an AI application will have the intended effect') and the precautionary fallback are paraphrased in the summary rather than quoted.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 86 (Chapter IV, Policy Area 5: Environment and ecosystems)",
        "quote": "[…] Member States should ensure that AI actors, in line with the principle of proportionality, favour data, energy and resource-efficient AI methods.",
        "rationale": "The paragraph's whole subject is 'the respective impact on the environment', so `Unn` is stated, not inferred. `Hum` follows from the paragraph's own fallback: where appropriate evidence that the application will have the intended effect cannot be produced, 'the precautionary principle must be favoured' — a rule for acting on acknowledged ignorance rather than on confidence. On evidence, the paragraph is explicit that what is wanted is a demonstration that the application will have the intended effect (`Dat`), assessed by a proportionality argument between the method's resource cost and its aim (`Log`). The sources designated are the addressee (`Gov`) and the AI actors whose method choice the paragraph regulates (`Ind`). No environmental authority, standards body or scholarly source is named, so none is coded. ADJUDICATION 2026-08-14: S ([Gov, Ind]) and obligationType (mixed) agreed exactly, and this is the only Chapter IV entry where Ind survives, because the excerpt names 'AI actors' as the party who must favour efficient methods — the Ind limb of the source-axis rule — while the State is directed to ensure it. V narrows to [Unn], which both passes reached from 'data, energy and resource-efficient AI methods'; Hum rests on v0.1 alone. E narrows to [Dat], both passes recording the same reservation, that a comparative preference for efficient methods presupposes efficiency having been measured while the paragraph declares no metric, baseline or unit. MODALITY, recorded and not coded: the second pass found blind that the duty is conditioned 'in line with the principle of proportionality' and that the verb is 'favour', not 'adopt' — a weighted preference rather than a requirement to use the most efficient method available. That is the same finding v0.1 made from the other direction when it declined to code the values the paragraph expects to be deprioritized (Ach and Por), which the `v` array cannot carry; both are recorded as Wave 2 gap 23 and Wave 1 gap 9 respectively.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "S and obligationType agreed exactly. V [Unn, Hum] → [Unn]; E [Dat, Log] → [Dat]. Both passes independently recorded that 'favour … in line with the principle of proportionality' is a weighted preference the v/e/s layers cannot express."
    },
    {
      "article": "Para. 127",
      "summary": "Member States should ensure that users can easily identify whether they are interacting with a living being or with an AI system imitating human or animal characteristics, and can effectively refuse such interaction and request human intervention.",
      "v": [
        "Sdt",
        "Sda"
      ],
      "e": [],
      "s": [],
      "status": "draft-verified",
      "obligationType": "mixed",
      "note": "The most directly testable paragraph in the Recommendation, and the one whose judgment correlate is least ambiguous: a model either discloses that it is a system when asked or implied, and either routes a refusing user to a human, or it does not. The organizational half is the human it routes to — staffing an escalation path is a management-system arrangement no item observes. Paragraphs 125, 126 and 128, which surround it (human-robot interaction guidelines, the extension of the same values to those interactions, and awareness of anthropomorphization), are not mapped; they are cited in the rationale and addressed in the guide. RETRIEVAL ITEM FOR THE NEXT REVISION (raised by the second pass, confirmed at adjudication): this entry's excerpt begins mid-sentence at the object clause, so the paragraph's 'Member States should ensure' stem is elided and no addressee appears in the quoted text. The blind pass therefore withheld `Gov` here while carrying it at ¶¶50, 55, 71 and 86, where the stem is present, and the adjudication accepted that rather than widening the quote — the CC BY-NC-SA 3.0 IGO constraint recorded in the pack notes holds quotation to the minimum needed. Re-retrieving ¶127 with its stem, so that the source layer can be decided on the text rather than on the chapter placement, is a next-revision retrieval item.",
      "provenance": {
        "sourceUrl": "https://unesdoc.unesco.org/ark:/48223/pf0000381137",
        "retrievalUrl": "https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence",
        "article": "Paragraph 127 (Chapter IV, Policy Area 11: Health and social well-being)",
        "quote": "[…] users can easily identify whether they are interacting with a living being, or with an AI system […] and can effectively refuse such interaction and request human intervention.",
        "rationale": "The paragraph gives the user two things. Knowing what they are dealing with is a precondition for judging the interaction for themselves (`Sdt`); being able to refuse it and get a human instead is freedom to choose one's own course (`Sda`). `Sep` is read from the paragraph's neighbourhood rather than invented: paragraph 125, in the same policy area, ends 'Special attention should be paid to the possibility of using AI to manipulate and abuse human cognitive biases', and paragraph 128 addresses anthropomorphization directly — the interest protected by easy identification is the user's psychological safety. Evidence: identification is discharged by a standing disclosure practice applied to every interaction (`Gui`), and the requested human intervention is discharged by that human's considered judgment on the case (`Exp`). Sources: the addressee (`Gov`) and the operator that must build the identification and the refusal path (`Ind`). `Pro` is not coded — the paragraph asks for a human, not a credentialed one. ADJUDICATION 2026-08-14: obligationType (mixed) agreed exactly, and V narrows to [Sdt, Sda], which both passes declared together — the identifiability guarantee protects the user's capacity to know what they are dealing with (Sdt) and refusal plus the request for human intervention are two concrete actions of their own (Sda). This is the one entry in the wave where the Sdt/Sda pair is carried, and it is carried because the paragraph states both limbs and both passes read both; elsewhere the two codes appeared as competing readings of a single phrase. Sep, which v0.1 had flagged as read from paragraphs 125 and 128 on manipulation of cognitive biases and anthropomorphization, was not reached by the second pass and does not survive. E is emptied: the second pass found, blind, that the discharge here is an interaction-time disclosure that the counterpart is an AI system, and that none of the ten evidence codes covers a real-time disclosure signal — not Gui (no procedure), not Dat (no measurement), not Exp (no one's judgment). It declined to force a near-fit and v0.1's [Gui, Exp] is withdrawn (Wave 2 gap 20, the same gap the second pass reached independently at G7 Action 7). S is emptied and the reason is recorded rather than worked around: the excerpt begins mid-sentence at the object clause, so the 'Member States should ensure' stem is elided and no addressee appears in the quoted text. Gov is withheld on the quote as given, even though the chapter placement makes it likely and ¶¶50, 55, 71 and 86 all carry it where the stem is present. RETRIEVAL ITEM: the excerpt is not widened here, because the licence constraint recorded in the reuse note holds quotation to the minimum, and re-retrieving the paragraph with its stem is listed as a next-revision item.",
        "retrievedAt": "2026-08-14",
        "verifiedBy": "dual formalization (blind second pass) + adjudication, 2026-08-14"
      },
      "changeNote": "obligationType agreed exactly. V [Sdt, Sda, Sep] → [Sdt, Sda]; E [Gui, Exp] → [] (no evidence code covers an interaction-time disclosure signal — Wave 2 gap 20); S [Gov, Ind] → []. The removal of Sep discharges the last of v0.1 note 8's four flagged inferences. Gov is withheld because the quoted excerpt elides the 'Member States should ensure' stem; widening the quote is refused under the licence constraint and recorded as a retrieval item for the next revision."
    }
  ],
  "itemBankRef": {
    "publicSet": "/content/standards-packs/item-banks/unesco-ai-ethics.public.json",
    "privateSet": null
  },
  "version": "0.2",
  "supersedes": "0.1",
  "status": "draft-verified",
  "updatedAt": "2026-08-14",
  "measurementScope": "The UNESCO Recommendation is a non-binding standard-setting instrument. Its operative clauses recommend that Member States apply its provisions on a voluntary basis, taking whatever legislative or other measures their constitutional practice requires, and that they engage all stakeholders, including business enterprises, in implementing it. The concrete duties an organization actually faces are therefore created by whatever national measures a Member State adopts, not by this text. Nothing in this pack measures a Member State's implementation, and nothing in it could. What AIO items measure against this pack is the judgment direction each mapped paragraph implies for an AI system — whether a model's reasoning tracks, for example, paragraph 26 on final human determination where an impact is irreversible, paragraph 40 on a meaningful explanation accompanying a decision that is not easily reversible, or paragraph 127 on identifying itself and yielding to a human on request. They do not assess whether an organization operates the management-system correlates of those paragraphs — ethical impact assessment, data governance strategies, audit and due diligence mechanisms, redress and remediation channels, environmental assessment — and they are not evidence of any Member State's implementation of the Recommendation or of any organization's compliance with any measure adopted under it. After the dual formalization and adjudication of 2026-08-14 the distribution is unchanged — one paragraph is `behavioral` (36, and it is the only `behavioral` entry surviving anywhere in the AIO pack series, reached independently by both passes), five are `mixed` (26, 38, 40, 86, 127) and four are `organizational` (43, 50, 55, 71) — but three of the four `organizational` paragraphs (50, 55, 71) now carry an explicit measurement-scope notice, because their directed substance is the State's own apparatus and their duty-bearer cannot be an AI system operator. Those three belong outside any scoring denominator rather than being counted as unmet.",
  "notes": [
    "draft-verified. Every entry carries a short verbatim excerpt of the official English text and a rationale argued from it, and on 2026-08-14 the second independent formalization required by FORMALIZATION_METHODOLOGY.md §5 was completed blind and adjudicated. The second formalizer read only the pack id, the sourceNorm and each entry's provenance.article, sourceUrl, retrievalUrl and quote; the v/e/s arrays, summaries, rationales, obligationType tags and notes of v0.1 were stripped by an extraction script before any file was opened, and neither the pack-authoring guideline nor the management guide was opened. Human review remains outstanding, and the V/E/S assignment is settled only by the public RFC round at https://aioq.org/en/rfc. One limit on the promotion is carried forward from the quote-verification note below and is not cured by the second pass: extraction from the retrieved manifestation was verified twice, but the manifestation itself could not be cross-checked against unesdoc, which was unreachable.",
    "Instrument status and currency, verified on 2026-08-14. The Recommendation was adopted by the General Conference of UNESCO at its 41st session in Paris on 23 November 2021 (document code SHS/BIO/PI/2021/1); the adoption date and place are as recorded on UNESCO's own Office of International Standards and Legal Affairs page for the instrument, which served the full 141-paragraph English text at retrieval time. That page records no amendment, revision or replacement of the instrument, and records that the Executive Board is responsible for monitoring its implementation. The Recommendation is not a treaty: it is a recommendation in the sense of UNESCO's standard-setting practice, applied by Member States on a voluntary basis, and its own final provision (paragraph 141) says that nothing in it may be interpreted as replacing, altering or otherwise prejudicing States' obligations or rights under international law. UNESCO has since adopted a separate Recommendation on the Ethics of Neurotechnology, listed in the same navigation; it is a different instrument and is not formalized here.",
    "Primary source and retrieval path. The canonical citation is the UNESCO Digital Library record for SHS/BIO/PI/2021/1 at https://unesdoc.unesco.org/ark:/48223/pf0000381137, recorded in every entry as `sourceUrl`. That endpoint could not be fetched: on 2026-08-14 unesdoc.unesco.org served a Cloudflare bot-challenge shell to a normal browser-identified request and returned HTTP 403 to the fetch tool, and no PDF manifestation of the text could be obtained through it. The text actually read is the official English manifestation published by UNESCO's Office of International Standards and Legal Affairs at https://www.unesco.org/en/legal-affairs/recommendation-ethics-artificial-intelligence — the same issuing body, a different endpoint — recorded in every entry as `retrievalUrl`. No commentary, summary, national-commission reprint, mirror or secondary source was used for any quote, and no quotation in this pack was reconstructed from memory.",
    "Quote verification method, and its limit. The retrieved page was reduced to a whitespace-normalized, NFC-normalized corpus of 102,292 characters by two independent extraction paths written separately — a regex tag-stripper with entity unescaping, and a DOM-based `html.parser` text collector — which produced byte-identical corpora. Every quoted fragment in this pack (13 fragments across 10 entries, counting the parts on either side of an […] elision separately) was then checked as an exact substring of both corpora; all 13 matched. The limit to record honestly: unlike the Council of Europe pack, which cross-checked two official manifestations, this pack had only one official manifestation available, because unesdoc was unreachable. Extraction was verified twice; the source manifestation was not. That is one of the reasons no entry is claimed above `draft-unverified`.",
    "Licence and reuse constraints — first pack built from a licence-constrained (C) source under docs/팩_추가_가이드라인.md §5. The Recommendation is published by UNESCO under Creative Commons Attribution-NonCommercial-ShareAlike 3.0 IGO (CC BY-NC-SA 3.0 IGO). That is not the CC BY-SA 3.0 IGO licence carried by many other UNESCO open-access titles, and the difference binds this pack. (a) NonCommercial: AIO's certification tiers are commercial-adjacent, so reuse of the Recommendation's text beyond short attributed quotation is not covered by the licence and is not relied on here. Quotation is held to the minimum needed to evidence each mapping — 10 excerpts, 1,624 characters in total, longest excerpt 187 characters, about 1.6% of the instrument's roughly 98,000-character operative text — and is offered as short attributed quotation with full source citation, not as licensed reuse. (b) ShareAlike: a control set derived by adapting the Recommendation's wording — a checklist, a control matrix, an item bank phrased from the text — would be an adaptation and would carry the SA obligation, which conflicts with the CC BY 4.0 terms AIO publishes its own vocabulary under. AIO therefore does not derive control sets from the Recommendation's wording: the V/E/S mappings here are AIO's own analysis expressed in AIO's own vocabulary, the summaries and notes are paraphrase, and the management-system guide for this pack paraphrases throughout, using blockquotes only where a provenance excerpt already verified in this pack exists. Attribution, as UNESCO asks for it: UNESCO, Recommendation on the Ethics of Artificial Intelligence, SHS/BIO/PI/2021/1, adopted 23 November 2021, © UNESCO, licensed under CC BY-NC-SA 3.0 IGO.",
    "Licence verification limit. The licence recorded above is the one carried in docs/license-inquiries/unesco.md, the standing AIO record for this norm; the imprint of the standalone publication is its origin. It could not be re-read from that imprint on 2026-08-14, because unesdoc — which serves it — was behind the bot challenge described above, and the Legal Affairs page that served the text carries no licence statement of its own. Where the two candidate readings differ, this pack applies the more restrictive one (NonCommercial and ShareAlike both in force). A permissions inquiry to UNESCO is pending; until a reply is received, this pack and its guide stay inside the short-attributed-quotation footprint described above, and any expansion of quotation waits on that reply.",
    "Selection, and what was left out. Ten paragraphs are mapped: 26, 36, 38, 40 and 43 from Chapter III.2 (Principles), and 50, 55, 71, 86 and 127 from Chapter IV (Areas of Policy Action, policy areas 1, 2, 3, 5 and 11). They were chosen as the paragraphs that regulate judgment, records, oversight or a risk trade-off and that have a direction an item could test. Excluded and recorded here as RFC re-examination candidates: paragraph 35 (attribution of ethical and legal responsibility, and inclusive public oversight) — addressed to Member States with no separable judgment correlate beyond paragraph 36; paragraphs 51 to 53 (due diligence and oversight mechanisms, monitoring of all life-cycle phases, a regulatory framework for ethical impact assessment by public authorities) — assessment-process duties fully covered by the guide; paragraphs 72 to 77 (privacy safeguards and privacy impact assessment, data subject rights, sensitive-data security, open data, dataset quality, digital commons); paragraph 84 (direct and indirect environmental impact assessment across the life cycle, including carbon footprint), left out only because paragraph 86 carries the same policy area's decision rule more testably; paragraph 113 (notification of reasons for content removal and appeal mechanisms) and paragraph 123 (health-care safeguards, including that the final decision of diagnosis and treatment is always taken by humans), both strong candidates held back to keep this seed at ten and the quotation footprint small. Policy areas 4, 6, 7, 8, 9 and 10 (development and international cooperation, gender, culture, education and research, communication and information, economy and labour) and Chapters V to VIII (monitoring and evaluation, utilization, promotion, final provisions) are addressed to Member States as institutions and carry no judgment correlate an AIO item could test; they are out of scope for the pack and are treated in the guide where they impose management duties.",
    "Measurement scope in numbers (per-entry `obligationType`, pack-level `measurementScope`), after adjudication. The distribution is exactly what v0.1 recorded and the second independent pass agreed with it on nine of the ten paragraphs — the highest obligationType agreement of any pack in either wave. One paragraph is `behavioral` (36 — an AI system can never replace ultimate human responsibility, and life-and-death decisions are not to be ceded to it; the paragraph is drafted impersonally and imposes no management-system duty in its words), five are `mixed` (26, 38, 40, 86, 127) and four are `organizational` outright (43, 50, 55, 71). **Paragraph 36 is the only `behavioral` entry surviving anywhere in the AIO pack series after dual formalization**, and it survives because two formalizations that could not see each other's work assigned the tag independently; the candidates in the OECD and California packs were both lowered at adjudication. The single divergence was at paragraph 26, which the second pass read as `behavioral` and which is kept at `mixed` under the conservative rule. The state-addressed caveat governs all ten regardless of classification: the Recommendation's addressee is the Member State, and `obligationType` records the shape of the duty, not who owes it. Paragraphs 50, 55 and 71 additionally carry a measurement-scope notice, because their directed substance is the State's own apparatus. A pass on this pack is evidence about model judgment only, and never evidence that an operator has implemented anything.",
    "Inferred codes, and what the second formalization did to them. v0.1 flagged four assignments as following from a paragraph's structure or from the shape of the AIO 00011 vocabulary rather than from the paragraph's words. Three did not survive, because the blind second pass did not reach them and an inference-flagged code survives only where both passes do: `Pee` on paragraph 26 (read from 'rigorous scientific foundations', which sits outside the quoted sentence — the second pass recorded an INSUFFICIENT-QUOTE finding on that whole layer and withheld `Pro` as well), `Tri` on paragraph 50 (read from 'facilitate citizen participation'), and `Sep` on paragraph 127 (read from paragraphs 125 and 128, which are cited and not quoted). **One survived and is no longer an inference: `Bed` on paragraph 36**, v0.1's stopgap for 'ultimate human responsibility and accountability', was reached independently by the second pass from the same words. The vocabulary gap it stands on is unchanged and still goes to the RFC round. Codes considered and deliberately not assigned were confirmed on both sides: `Rev` on paragraph 26, `Pro` on paragraphs 26, 36, 43 and 127, `Dat` on paragraph 38, `Pee` on paragraphs 43 and 55, and — reached only by the second pass — `Unc` on paragraph 71, which it withheld because the quoted text does not mention bias, discrimination or fairness. Three inference-grade codes the second pass introduced were removed on the same rule, v0.1 not having reached them: `Sda` at paragraph 38, `Sep` at paragraph 40 and `Cor` at paragraph 43.",
    "Non-endorsement. AIO wrote this formalization. UNESCO, its General Conference, its Executive Board, its Secretariat and its Social and Human Sciences Sector took no part in it, have not reviewed it, and have not endorsed it. It is not an official interpretation of the Recommendation. AIO certifies conformance to AIO's own formalization of this Recommendation; that is not a legal assessment, not an assessment of any Member State's implementation, and confers no status of any kind under the Recommendation or under any national measure adopted to give effect to it. UNESCO's own implementation instruments — the Readiness Assessment Methodology (RAM) and the Ethical Impact Assessment (EIA) methodology, both foreseen by paragraphs 49 and 131 — are UNESCO's; nothing here is a substitute for either, and nothing here is produced under either.",
    "No item bank. `itemBankRef.publicSet` and `itemBankRef.privateSet` are both null: no scenario items have been written for this pack, so it currently backs no certificate at any tier and appears in the catalogue as listed, measurement pending. Item authoring follows verification, not the other way round. Note additionally that any item bank for this pack must be authored in AIO's own words: under the ShareAlike constraint recorded above, items phrased by adapting the Recommendation's wording would be adaptations of the source text. [갱신 2026-08-15: 이중 관문 문항 뱅크 개통 — 관문 A 공개 세트 + 관문 B 비공개 뱅크(서명 커밋먼트 게시). 이 노트의 이전 서술은 개통 전 기록이다.]",
    "Methodology for the paragraph → V/E/S translation: /content/standards-packs/FORMALIZATION_METHODOLOGY.md. V/E/S values are the canonical three-letter AIO 00011 codes served at /api/framework/vocabulary — the same codes an AIO 20002 record carries. Management-system obligations arising from this Recommendation, including the many paragraphs this pack does not map, are covered in docs/management-guides/unesco-ai-ethics.ko.md and .en.md.",
    "Addressee convention, adopted from the second pass and applied uniformly to this pack. The Recommendation addresses Member States, but the addressee and the directed substance come apart, and the convention adopted here is to code the substance the text directs — what must be true of the AI practice or of the decision — and to let the addressee bear on `s` and `obligationType` only. Where the State is directed to build an apparatus that will govern AI practice ('introduce frameworks', 'enact strong enforcement mechanisms', 'develop data governance strategies'), the addressee designates the governing authority, `Gov` is earned and the classification is `organizational`; those entries also carry a measurement-scope notice, because their duty-bearer cannot be an AI system operator. Where the directed substance is a judgment or an interaction in the concrete case, `obligationType` is `mixed` or `behavioral` even though the State is the formal addressee. The Chapter III paragraphs (26, 36, 38, 40, 43) are stated impersonally or in the passive and designate no addressee at all; that silence is recorded as an empty source layer rather than filled from the chapter's addressee, which is why five of the ten entries in this pack carry no `s` code. This convention is the reason `Gov` survives here at ¶¶50, 55, 71 and 86 while it was removed from the parallel OECD pack's Recommendations 2.5, where governments are addressed but no government instrument is designated as decisive over any AI practice.",
    "Source-axis policy (settled in Wave 1, extended in Wave 2, applied uniformly across the five packs of this wave). **`Gov` is declared only where the excerpt names a government body or a government norm as decisive on the substance of the duty**; being named as the recipient of a report, a filing or a disclosure does not earn it, and neither does bare addressee status without a directed instrument. **`Ind` is declared where the excerpt makes the industry duty-bearer the author or performer of the provision's product or determination.** The rule is applied as a filter and never as a generator: it may remove a code both passes declared, and it may decide which of two divergent readings prevails, but it never adds a code neither pass declared. In this pack the rule removed `Ind` from ¶¶55 and 71, whose actor is the Member State, and left it at ¶86, whose excerpt names 'AI actors'; it removed `Gov` from ¶¶26, 36, 38, 40 and 43, where v0.1 had read it from the instrument's addressee rather than from the quoted words. The pack's source layer is empty in five of ten entries as a result, which is the highest proportion in either wave and is a property of Chapter III's impersonal drafting.",
    "Adjudication method (v0.2). This pack was formalized twice. The v0.1 seed pass is the first formalization; the second was blind, under the protocol recorded in the first note. The two results were compared mechanically, entry by entry and layer by layer, with v, e and s treated as sets. Exact agreement was auto-accepted. Divergences were adjudicated under a fixed policy carried forward from Wave 1: the reading better grounded in the quoted text prevails under FORMALIZATION_METHODOLOGY.md §4; where both readings are defensible the more conservative is taken; the intersection is an allowed outcome where it is non-empty and defensible; no third reading is invented, and every adjudicated set is a subset of at least one pass's set. Two sub-rules settled in this wave: a code flagged INFERENCE by the pass that declared it survives only where both passes reached it, and a divergent `obligationType` always resolves to the more conservative tag. Agreement statistics for this pack, across ten entries: V 1/10, E 4/10, S 2/10, obligationType 9/10, all four axes together 0/10. The obligationType figure is the highest in either wave and the value figure the lowest, and both follow from the same property of the instrument: a recommendation states plainly what kind of thing it is asking for, and states its protected interests only sometimes.",
    "Contamination notice. The second formalization of this pack ran under the tightened Wave 2 protocol, in which the pack-authoring guideline was blocked outright, and the second pass disclosed no exposure to any pack field or to any prior adjudication. No axis of this pack is reported with a contamination caveat. The second pass did state an addressee convention of its own before beginning, which is recorded above and adopted; it was reached from FORMALIZATION_METHODOLOGY.md §4 and the quoted excerpts, not from this pack's notes.",
    "Licence discipline under the dual formalization, and one retrieval item it produced. The CC BY-NC-SA 3.0 IGO constraint recorded above was held throughout the second pass and the adjudication: **no quote in this pack was widened, at any point, to settle an adjudication question.** Two entries were affected and both are recorded rather than cured. At ¶40 the excerpt is elided at both the subject and the operative verb, so the second pass could not assess the paragraph's modality from it. At ¶127 the excerpt begins at the object clause, so the 'Member States should ensure' stem is absent and `Gov` was withheld on the quote as given, diverging from ¶¶50, 55, 71 and 86 where the stem is present. Both are listed as next-revision retrieval items — to be re-retrieved with their stems, within the same minimal-quotation footprint — rather than resolved by expanding the excerpt now. A permissions inquiry to UNESCO remains pending and any expansion of quotation waits on it.",
    "Vocabulary and schema gaps found by the dual formalization (feeding a future AIO 00011 RFC). This pack contributes one new item to the Wave 2 list, which continues the consolidated Wave 1 list of ten, and confirms four existing items. (23) DEFEASIBLE VERSUS CATEGORICAL MODALITY — paragraph 36 states two limbs of unequal strength in one breath: 'an AI system can never replace ultimate human responsibility and accountability' is categorical, while 'As a rule, life and death decisions should not be ceded to AI systems' is expressly defeasible. The v/e/s layers record what a provision points at but not how strongly, so a formalization cannot distinguish a prohibition from a rebuttable presumption, and an item written from the second limb as if it were the first would overstate the norm. The second pass found this blind and flagged it as material; the same problem recurs at ¶86 ('favour … in line with the principle of proportionality') and across the hedges of the G7 pack in this wave. This is a schema question, the third after Wave 1 gap 9 and Wave 2 gap 21. Confirmed again: gap 10 UNDEFINED PROTECTED INTEREST (¶71, where the two passes shared no value code and the layer is left empty — the purest instance yet), gap 12 LOGGING AND TRACEABILITY AS AN EVIDENCE CLASS (¶43, where both passes carried Dat with the same recorded discomfort), gap 20 INTERACTION-TIME DISCLOSURE OR TECHNICAL MARKER AS AN EVIDENCE CLASS (¶127, reached independently and matching G7 Action 7), and gap 13 THE Sdt/Sda BOUNDARY (¶38, where the passes split, against ¶127, where both carried the pair because the paragraph states both limbs). The full Wave 2 list is reproduced in the adjudication report."
  ]
}