{"schema":"https://ver.cy/schemas/card/1.0.0","id":"vr.wm-ai-011","code":"wm-ai-011-evaluation-harness-and-scorer","url":"https://ver.cy/models/wm-ai-011-evaluation-harness-and-scorer/","name":"Evaluation Harness and Scorer","alternateNames":[],"kind":"world-model","status":"published","version":"0.1.0","language":"en","classifiers":{"family":"World Models","category":"Information and virtual systems","entryKind":"entity","plane":"","domain":["INF.AI.HRN"],"industry":["Cross-industry"],"navPath":"NAV.INF.AI.HRN","tags":["evaluation","harness","scorer","inf.ai.hrn"],"facets":{}},"whatItIs":"One identifiable procedure definition across immutable revisions. It owns input and output contracts, scoring recipes, judge qualification requirements and revision comparability. It does not own an evaluation run or authorize operational decisions.","purpose":"Describe a versioned scoring procedure and its harness contract, including optional model judges, with reproducible bindings and explicit validity limits.","scope":{"in":["Procedure identity, semantic revision and accountable custodian","Task applicability, adapters, metrics, rubrics and aggregation recipes","Optional model judge bindings, calibration and trust policy","Runtime request and response contracts, replay requirements and qualification evidence","Definition change, comparability and governed retirement"],"out":["WM-AI-003 evaluation run, finding and approval lifecycle","Dataset curation, training, benchmark ownership and model artifact lifecycle","Actual scheduling, execution, resource provisioning and external tool effects","Release approval, certification, regulatory compliance conclusions and universal rankings","Operational instructions for dangerous evaluations"],"boundaries":[{"neighbor":"WM-AI-003 AI Model Evaluation","distinction":"Registry parent only; no incident relation-ledger edges exist at freeze. A proposed REFERENCE binding passes the procedure revision to the evaluation record. Runs, observations, findings and approvals remain external."},{"neighbor":"Dataset and model master records","distinction":"The harness pins immutable material and subject/judge model references. It does not duplicate their ownership or lifecycle. Exact neighbor identifiers require an adopting profile."},{"neighbor":"Execution runtime and audit record","distinction":"This entity describes required behavior and evidence shape. Runtime execution and authoritative event storage are delegated; local inspection is not execution."},{"neighbor":"External open specifications","distinction":"PROV-O and EARL are conceptual mappings. RFC 8785 is an optional serialization binding, not the format-neutral meaning of a score."}]},"distinguishingFeatures":["Unlike WM-AI-003 AI Model Evaluation: Registry parent only; no incident relation-ledger edges exist at freeze. A proposed REFERENCE binding passes the procedure revision to the evaluation record. Runs, observations, findings and approvals remain external.","Unlike Dataset and model master records: The harness pins immutable material and subject/judge model references. It does not duplicate their ownership or lifecycle. Exact neighbor identifiers require an adopting profile.","Unlike Execution runtime and audit record: This entity describes required behavior and evidence shape. Runtime execution and authoritative event storage are delegated; local inspection is not execution.","Unlike External open specifications: PROV-O and EARL are conceptual mappings. RFC 8785 is an optional serialization binding, not the format-neutral meaning of a score."],"structure":{"bundles":[{"id":"bundle-definition","name":"Definition and accountability","description":"A stable procedure boundary and its intended construct.","layers":[{"id":"layer-identity","name":"Identity and responsibility","description":"Context for identity and responsibility within a versioned scoring procedure.","findings":[{"id":"procedure-identity","name":"Procedure identity and revision","description":"A harness definition has a persistent identity and immutable semantic revisions. Its configuration is distinct from a particular invocation and from an evaluated model. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which master identity and revision identify this scoring procedure?","id":"procedure-identity-q1","kind":"identity"},{"text":"Which scoring mode and supported task family classify this revision?","id":"procedure-identity-q2","kind":"classification"},{"text":"Which predecessor and dependency changes produced this revision?","id":"procedure-identity-q3","kind":"provenance"}]},{"id":"ownership-boundary","name":"Accountability and evaluation binding","description":"The procedure custodian governs definitions; the evaluation owner governs runs, findings and approvals. A registry parent is contextual and does not establish executable inheritance. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which accountable role maintains the procedure and which role owns evaluation decisions?","id":"ownership-boundary-q1","kind":"ownership"},{"text":"Which evaluation record may reference this revision without transferring its lifecycle here?","id":"ownership-boundary-q2","kind":"relationship"},{"text":"What evidence permits a custodian to issue a new procedure revision?","id":"ownership-boundary-q3","kind":"authority"}]}]},{"id":"layer-applicability","name":"Measurement applicability","description":"Context for measurement applicability within a versioned scoring procedure.","findings":[{"id":"intended-construct","name":"Intended construct and applicability","description":"The author declares what the procedure intends to measure and where it is unsupported. A computed value alone cannot establish suitability for an evaluation decision. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"What construct is measured and what observable evidence is used as its proxy?","id":"intended-construct-q1","kind":"definition"},{"text":"Which task, language, modality and deployment contexts are inside the claimed scope?","id":"intended-construct-q2","kind":"constraint"},{"text":"Which unmet conditions require rejection or an explicitly unsupported result?","id":"intended-construct-q3","kind":"exception"}]}]}]},{"id":"bundle-inputs","name":"Inputs and preparation","description":"Bindings and transformations that precede scoring.","layers":[{"id":"layer-material","name":"Input roles and material bindings","description":"Context for input roles and material bindings within a versioned scoring procedure.","findings":[{"id":"input-contract","name":"Input roles and admissibility","description":"The adapter declares subject output, reference material, item identity and task context separately. Missing labels, reordered items and unavailable material cannot silently become valid inputs. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which input roles and schema versions are required for this scoring mode?","id":"input-contract-q1","kind":"composition"},{"text":"How are item identity, output alignment and missing required inputs checked?","id":"input-contract-q2","kind":"validation"},{"text":"Which adapter transformations are permitted and which information losses must be reported?","id":"input-contract-q3","kind":"interoperability"}]},{"id":"test-material-binding","name":"Test material and separation rules","description":"The definition references externally mastered test material and declares split eligibility, exposure restrictions and contamination checks. It neither curates the dataset nor certifies absence of contamination. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which immutable dataset or reference-corpus revision and selection rule does the procedure accept?","id":"test-material-binding-q1","kind":"provenance"},{"text":"What training, tuning or judge exposure would undermine the proposed measurement?","id":"test-material-binding-q2","kind":"quality"},{"text":"Which actor may view held-out labels or reference answers during each scoring phase?","id":"test-material-binding-q3","kind":"access"}]}]},{"id":"layer-preparation","name":"Preparation constraints","description":"Context for preparation constraints within a versioned scoring procedure.","findings":[{"id":"preparation-policy","name":"Preprocessing and resource policy","description":"A proposed preparation contract pins parsing, normalization, truncation and resource ceilings before scoring. These settings can affect meaning and belong in comparison criteria. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which ordered normalization and extraction steps preserve or change score-relevant meaning?","id":"preparation-policy-q1","kind":"process"},{"text":"Which input limits and truncation behavior apply to every supported modality?","id":"preparation-policy-q2","kind":"requirement"},{"text":"Which isolation and resource restrictions must a runtime attest before using this definition?","id":"preparation-policy-q3","kind":"security"}]}]}]},{"id":"bundle-scoring","name":"Scoring semantics","description":"Formula, rubric and aggregation definitions.","layers":[{"id":"layer-criteria","name":"Criteria and rubrics","description":"Context for criteria and rubrics within a versioned scoring procedure.","findings":[{"id":"metric-contract","name":"Metric formula and score semantics","description":"Each selected metric needs a precise value domain, orientation and interpretation. Metric definitions can be externally referenced; this procedure owns their binding, parameterization and composition. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which formula, value type, unit and direction define each returned score?","id":"metric-contract-q1","kind":"measurement"},{"text":"How are ties, boundary values and undefined arithmetic handled?","id":"metric-contract-q2","kind":"constraint"},{"text":"Which thresholds yield a local criterion outcome and which release decisions remain external?","id":"metric-contract-q3","kind":"decision"}]},{"id":"rubric-contract","name":"Rubric and reference judgement","description":"A judge rubric describes criteria, scale anchors and admissible evidence. Exemplars and reference judgements are versioned inputs, not unquestionable ground truth. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which criterion wording and scale anchors constrain the judgement?","id":"rubric-contract-q1","kind":"definition"},{"text":"Which reference examples and counterexamples establish the intended interpretation?","id":"rubric-contract-q2","kind":"evidence"},{"text":"When should the rubric require abstention or referral instead of a numeric score?","id":"rubric-contract-q3","kind":"exception"}]}]},{"id":"layer-aggregation","name":"Aggregation policy","description":"Context for aggregation policy within a versioned scoring procedure.","findings":[{"id":"aggregate-contract","name":"Aggregation and denominator policy","description":"The proposed aggregation recipe declares weighting, eligible populations and missingness before results are inspected. Summary values preserve denominator and component context. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"How are item values, weights and group summaries combined?","id":"aggregate-contract-q1","kind":"measurement"},{"text":"Which attempted, eligible, valid and excluded counts accompany each summary?","id":"aggregate-contract-q2","kind":"quality"},{"text":"How are repeated items, correlated samples and undefined group summaries treated?","id":"aggregate-contract-q3","kind":"validation"}]}]}]},{"id":"bundle-judges","name":"Model judge qualification","description":"Optional model judges and their limitations.","layers":[{"id":"layer-configuration","name":"Configuration and calibration","description":"Context for configuration and calibration within a versioned scoring procedure.","findings":[{"id":"judge-binding","name":"Model judge configuration","description":"A model judge is an optional scoring dependency with pinned model reference, prompt, rubric and output parser. Unknown remote revision or stochastic settings restrict replay claims. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which judge model, prompt, rubric and parser revisions form the scoring dependency?","id":"judge-binding-q1","kind":"composition"},{"text":"Which decoding settings, seeds and availability assumptions apply to the judge?","id":"judge-binding-q2","kind":"state"},{"text":"How is a resolved judge revision captured when an endpoint alias can change?","id":"judge-binding-q3","kind":"provenance"}]},{"id":"judge-calibration","name":"Judge calibration and bias checks","description":"A proposed judge qualification profile records reference-review agreement and sensitivity tests before adoption. It makes no universal accuracy claim and sets no default acceptance percentage. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which independent reference judgements and acceptance criteria qualify this judge for this task?","id":"judge-calibration-q1","kind":"validation"},{"text":"Which order, verbosity, language and self-preference sensitivity checks are required?","id":"judge-calibration-q2","kind":"quality"},{"text":"Which rubric, judge or context changes invalidate prior qualification?","id":"judge-calibration-q3","kind":"lifecycle"}]}]},{"id":"layer-trust","name":"Trust and review controls","description":"Context for trust and review controls within a versioned scoring procedure.","findings":[{"id":"judge-trust","name":"Judge trust boundary and review route","description":"Scored content is untrusted data. The harness declares least-authority requirements, evidence handling and escalation policy; dangerous evaluations remain at policy level. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"How must runtime policy isolate evaluated content from judge instructions and privileged actions?","id":"judge-trust-q1","kind":"security"},{"text":"Which sensitive input or rationale content may be withheld, minimized or redacted?","id":"judge-trust-q2","kind":"privacy"},{"text":"Who may review disputed judge outputs without rewriting the original verdict?","id":"judge-trust-q3","kind":"authority"}]}]}]},{"id":"bundle-invocation","name":"Invocation and evidence contract","description":"Required runtime behavior and evidence shapes without run ownership.","layers":[{"id":"layer-runtime","name":"Invocation and response","description":"Context for invocation and response within a versioned scoring procedure.","findings":[{"id":"execution-contract","name":"Invocation recipe and failure behavior","description":"The procedure defines the request contract and expected failure semantics. Execution, scheduling, retries actually taken and run state are owned by the evaluation runtime. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which steps and dependency checks must precede a scoring invocation?","id":"execution-contract-q1","kind":"process"},{"text":"How must timeouts, malformed responses, refusals and cancellation be represented?","id":"execution-contract-q2","kind":"exception"},{"text":"Which timeout, retry and cache rules preserve the declared observation unit?","id":"execution-contract-q3","kind":"temporal"}]},{"id":"result-envelope","name":"Result envelope and provenance linkage","description":"A response schema carries item and procedure references, outcome state, optional score and evidence pointers. The actual result is mastered with its evaluation run, not as the harness definition. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which invocation, item and procedure revision identifiers must each response retain?","id":"result-envelope-q1","kind":"identity"},{"text":"How are valid score, failure, indeterminate, inapplicable and untested outcomes distinguished?","id":"result-envelope-q2","kind":"state"},{"text":"Which event and observation times accompany provenance without serving as identity?","id":"result-envelope-q3","kind":"temporal"}]}]},{"id":"layer-reproducibility","name":"Replay requirements","description":"Context for replay requirements within a versioned scoring procedure.","findings":[{"id":"replay-contract","name":"Replay and uncertainty contract","description":"A proposed reproducibility profile distinguishes exact replay from distributional agreement. Environment dependencies and uncertainty methods are declared rather than inferred from a seed. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which code, dependencies, material, environment and configuration pins are needed for replay?","id":"replay-contract-q1","kind":"requirement"},{"text":"Which variability sources and uncertainty method apply to this score and sampling unit?","id":"replay-contract-q2","kind":"measurement"},{"text":"What evidence and tolerances distinguish successful replay from a merely similar result?","id":"replay-contract-q3","kind":"validation"}]}]}]},{"id":"bundle-assurance","name":"Assurance and change","description":"Qualification, comparison and retirement of the definition.","layers":[{"id":"layer-qualification","name":"Verification and comparison","description":"Context for verification and comparison within a versioned scoring procedure.","findings":[{"id":"conformance-plan","name":"Procedure verification and negative fixtures","description":"A qualification plan tests the scorer definition and bindings before use. Local research schema validation is not evidence that any deployed scorer implements them. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which known-answer fixtures verify formulas, parsers and aggregation behavior?","id":"conformance-plan-q1","kind":"validation"},{"text":"Which malformed, empty, duplicated and adversarial input classes must be rejected or qualified?","id":"conformance-plan-q2","kind":"quality"},{"text":"Which implementation revision and reviewer evidence support a conformance assertion?","id":"conformance-plan-q3","kind":"evidence"}]},{"id":"comparison-policy","name":"Change impact and comparability","description":"Procedure revisions are compared by declared semantic differences, not by shared display names. The evaluation owner decides whether historical results need a fresh run. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which changes create a new semantic revision rather than a metadata correction?","id":"comparison-policy-q1","kind":"lifecycle"},{"text":"Which bindings must match before scores from two revisions can be compared?","id":"comparison-policy-q2","kind":"relationship"},{"text":"Who determines whether a changed scorer requires re-evaluation of earlier results?","id":"comparison-policy-q3","kind":"decision"}]}]},{"id":"layer-retirement","name":"Definition retirement","description":"Context for definition retirement within a versioned scoring procedure.","findings":[{"id":"retirement-policy","name":"Deprecation and governed disposal","description":"The procedure can be withdrawn while historical references remain resolvable under the adopting retention policy. Confidential payload disposal is executed by its custodian; preservation is not perpetual. Proposed local profile design; cited sources ground the concern, not every implementation detail.","questions":[{"text":"Which conditions deprecate or withdraw this revision and identify a successor?","id":"retirement-policy-q1","kind":"lifecycle"},{"text":"Which local manifests, private rubrics and evidence copies may be retained or deleted?","id":"retirement-policy-q2","kind":"retention"},{"text":"What minimal tombstone remains after authorized disposal without exposing protected content?","id":"retirement-policy-q3","kind":"access"}]}]}]}]},"agentConduct":{"may":["Validate a procedure definition: Proposed, unimplemented local operation. Check references, applicable modes, formula domains and required failure policies.","Prepare a scoring request: Proposed, unimplemented local operation. Resolve permitted item, material and judge bindings into a request manifest.","Inspect a supplied score envelope: Proposed, unimplemented local operation. Check structure, status, provenance links and denominator disclosures against the chosen revision.","Record qualification evidence: Proposed, unimplemented local operation. Attach reviewed fixture and judge-calibration evidence to the exact procedure revision.","Assess revision comparability: Proposed, unimplemented local operation. Produce a semantic difference and comparability statement for a pair of definitions.","Record procedure retirement: Proposed, unimplemented local operation. Record authorized deprecation, successor and disposition references."],"mustNot":["Retain only permitted minimal tombstones and provenance, redact protected payloads when required, and never delete externally mastered datasets, evaluation runs or results through this model.","Deny ungranted access; public definition metadata does not grant access to labels, private prompts or run payloads.","Time-limited reviewer access requires purpose, authority and an audit reference; it never grants execution authority."],"requiresHuman":[]},"ethics":{"considerations":["Record disclosure, redaction and disposition decisions with minimal personal data.","Adopters determine privacy, rights, retention and review requirements.","A judge response that matches its schema can still be false or biased."],"affectedParties":[]},"owners":{"steward":"Name an accountable procedure custodian, namespace authority and external evaluation owner using role references, never a vendor owner.","roles":[{"name":"Procedure custodian","responsibilities":["Maintains definitions and proposes revisions within delegated authority."]},{"name":"Measurement reviewer","responsibilities":["Assesses construct validity, calibration and comparability evidence."]},{"name":"Evaluation owner","responsibilities":["Owns the external evaluation lifecycle and decisions based on results."]},{"name":"Runtime operator","responsibilities":["Implements invocation policy and records actual runtime events externally."]},{"name":"Data custodian","responsibilities":["Controls protected material, retention and authorized disposal."]}],"masterSystems":[]},"relations":[{"target":"WM-AI-003","type":"references","note":"Proposed binding to the evaluation owner. Registry parent is reconciled as context, not automatic inheritance; evaluation runs, results, findings and approvals remain there."},{"target":"W3C PROV-O","type":"aligned","note":"Map procedure revisions to entities/plans and external invocation evidence to activities and attribution. No implemented mapping asserted."},{"target":"W3C EARL 1.0 Schema","type":"aligned","note":"Candidate mapping of criterion, assertor, subject and outcome in the response envelope. Note status and lossy mappings must remain visible."},{"target":"RFC 8785","type":"aligned","note":"Optional JSON canonicalization binding for integrity checks, with no semantic or measurement conformance claim."},{"target":"WM-AI-003 AI Model Evaluation","type":"neighbor","note":"Registry parent only; no incident relation-ledger edges exist at freeze. A proposed REFERENCE binding passes the procedure revision to the evaluation record. Runs, observations, findings and approvals remain external."},{"target":"Dataset and model master records","type":"neighbor","note":"The harness pins immutable material and subject/judge model references. It does not duplicate their ownership or lifecycle. Exact neighbor identifiers require an adopting profile."},{"target":"Execution runtime and audit record","type":"neighbor","note":"This entity describes required behavior and evidence shape. Runtime execution and authoritative event storage are delegated; local inspection is not execution."},{"target":"External open specifications","type":"neighbor","note":"PROV-O and EARL are conceptual mappings. RFC 8785 is an optional serialization binding, not the format-neutral meaning of a score."},{"target":"WM-AI-003","type":"parent"}],"interaction":{"identity":{"applicability":"required","items":["Authoritative master-system identifier","Governed global identifier or IRI","UUID or ULID assigned by the adopting Dimension"]},"properties":{"applicability":"not-applicable","items":[]},"recognition":{"applicability":"optional","items":[]},"capabilities":{"applicability":"required","items":["Validate a procedure definition: Proposed, unimplemented local operation. Check references, applicable modes, formula domains and required failure policies.","Prepare a scoring request: Proposed, unimplemented local operation. Resolve permitted item, material and judge bindings into a request manifest.","Inspect a supplied score envelope: Proposed, unimplemented local operation. Check structure, status, provenance links and denominator disclosures against the chosen revision.","Record qualification evidence: Proposed, unimplemented local operation. Attach reviewed fixture and judge-calibration evidence to the exact procedure revision.","Assess revision comparability: Proposed, unimplemented local operation. Produce a semantic difference and comparability statement for a pair of definitions.","Record procedure retirement: Proposed, unimplemented local operation. Record authorized deprecation, successor and disposition references."]},"hazards":{"applicability":"optional","items":[]},"interfaces":{"applicability":"required","items":["PROV-O: The PROV Ontology"]},"context":{"applicability":"required","items":["No jurisdiction-specific legal compliance is asserted. Adopters determine privacy, rights, retention and review requirements.","Public guidance is used for conceptual design, not as a universal obligation or institutional endorsement."]}},"sources":[{"title":"AI RMF Playbook - Measure","url":"https://airc.nist.gov/airmf-resources/playbook/measure/","note":"NIST"},{"title":"Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile","url":"https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf","note":"NIST"},{"title":"Building Evaluation Probes into Agentic AI","url":"https://www.nist.gov/programs-projects/building-evaluation-probes-agentic-ai","note":"NIST"},{"title":"PROV-O: The PROV Ontology","url":"https://www.w3.org/TR/prov-o/","note":"W3C"},{"title":"Evaluation and Report Language (EARL) 1.0 Schema","url":"https://www.w3.org/TR/EARL10-Schema/","note":"W3C"},{"title":"JSON Canonicalization Scheme (JCS)","url":"https://www.rfc-editor.org/rfc/rfc8785.html","note":"RFC Editor"},{"title":"Date and Time on the Internet: Timestamps","url":"https://www.rfc-editor.org/rfc/rfc3339.html","note":"IETF"},{"title":"LLM01:2025 Prompt Injection","url":"https://genai.owasp.org/llmrisk/llm01-prompt-injection/","note":"OWASP"}],"openQuestions":["Restore independent external review and verify admitted source versions and claim support before any canonical or publishable-draft promotion.","Specify conditional instance constraints and executable fixtures for input joins, malformed responses, empty denominators, retries, semantic changes and lawful disposition.","Empirically qualify judge calibration, bias sensitivity, reference-label protocols, uncertainty methods and comparability for each intended task profile.","Pin external model, dataset, evaluation, audit and runtime bindings and test loss-aware interoperability mappings.","Nested instance constraints, operational scorer implementations and executable fixtures are not delivered.","Statistical estimator choice, sample-size planning, domain-specific validity and human reference-label protocols require qualified profiles.","Exact model, dataset, audit and runtime neighbor IDs and versions remain unpinned.","Multimodal, interactive, embodied and high-risk use cases require specialized profiles; this draft offers no operational evaluation recipes.","Live-source version checks, reproduction rights and independent external review remain incomplete."],"resources":{"spec":"/models/wm-ai-011-evaluation-harness-and-scorer/spec.yaml","agents":"/models/wm-ai-011-evaluation-harness-and-scorer/AGENTS.md","source":"https://github.com/ver-cy/world-models/tree/feat/mega-model-registry/research/runs/wm-ai-011"},"provenance":{"origin":"world-models research","builtFrom":["models/wm-ai-011-evaluation-harness-and-scorer/spec.yaml"],"providers":["Codex"],"researchStatus":"reviewable-draft","generatedAt":"2026-10-21T16:57:27Z","builder":"tools/build_cards.py@1.0.0"},"completeness":{"sections":{"classifiers":"filled","whatItIs":"filled","purpose":"filled","distinguishingFeatures":"derived","structure":"filled","agentConduct":"derived","ethics":"derived","owners":"filled","relations":"filled","interaction.identity":"filled","interaction.properties":"not-applicable","interaction.recognition":"missing","interaction.capabilities":"filled","interaction.hazards":"missing","interaction.interfaces":"derived","interaction.context":"filled","sources":"filled"},"notes":{"distinguishingFeatures":"Derived from boundary notes against neighbouring models.","agentConduct":"Derived from functions, policies, CRUD and access rules; prohibitions were not authored for agents as such.","ethics":"Sentences mentioning harm, privacy, consent or similar, collected from the specification.","interaction.properties":"Institutional or informational subject: no invented physical properties."},"score":0.75}}