Eve library

Music Shield Legal Team Guide

Guide

Revision 1

On this page

This guide is for in-house counsel, outside counsel, rights clearance teams, and legal operations staff reviewing Music Shield risk reports before publication or in response to third-party claims.

Music Shield provides structured evidence, legal-risk signals, and immutable audit records. It does not replace counsel's judgment. Treat every automated classification as a starting point for legal review, not as a final legal opinion.

Legal review has four goals:

  1. Decide whether a generated work can be released in the configured jurisdictions.
  2. Preserve the evidence needed to defend or explain the release decision.
  3. Identify remediation steps when publication is blocked or conditional.
  4. Feed new legal developments, claims, and review outcomes back into threshold calibration.

A legal review is required when:

  • the scan decision is high_risk, reject, or equivalent release-blocking output;
  • a jurisdiction report returns high or critical;
  • the strictest-jurisdiction aggregation says releaseSafe: false;
  • a human reviewer escalates to legal;
  • a third-party claim, Content ID match, demand letter, DMCA notice, or platform complaint arrives;
  • a new case decision or legislative change affects a deployed threshold.

Review Packet#

Before reviewing the merits, confirm the packet includes:

  • generation id, scan id, assessment id, and any publication id;
  • full scan result with overallRisk, overallScore, confidence, elementScores, compoundMatches, jurisdictionAssessments, and recommendations;
  • all matched reference work ids, external identifiers, rights metadata, and protectability annotations;
  • audio excerpts, A/B comparison links, heatmaps, and isolated element clips used by reviewers;
  • human reviewer decisions, rationales, conditions, escalations, and overrides;
  • provenance export in Markdown for reading and JSON for downstream evidence management;
  • training data manifest and opt-out compliance issues when model provenance is relevant;
  • case law citations attached to the risk report, including source URLs.

Do not clear a matter when a required jurisdiction report, reference-work metadata record, or provenance chain is missing. Missing jurisdiction assessments are treated as release blocking in global distribution review.

Interpreting Risk Assessments#

The main risk report is shaped by RiskAssessment in @themis/risk-scoring.

Field Legal use
assessmentId Stable id for the legal matter, review notes, and exports
generatedAt Timestamp used for policy and law-in-effect review
overallRisk Top-level minimal, low, medium, high, or critical classification
overallScore Normalized 0-1 score behind the top-level classification
confidence Signal quality; low confidence usually means manual evidence review
elementScores Element-specific risk for melody, lyrics, vocal, harmony, spectral, etc.
compoundMatches Multi-element similarity against the same reference work
jurisdictionAssessments Per-jurisdiction legal test outputs
recommendations Operational actions such as hold, remediate, escalate, or retain evidence

Risk levels are ordinal:

Risk level Legal posture
minimal No material similarity signal. Confirm provenance exists and retain the packet.
low Weak signal. Publish only if configured policy allows and no metadata is missing.
medium Reviewable signal. Require legal or specialist review for important releases.
high Release-blocking by default. Require remediation, license, or legal sign-off.
critical Presumptive block. Prepare legal hold, remediation, and senior counsel review.

The score is not a damages estimate, infringement probability, or replacement for a legal test. It ranks evidence strength across analyzers. Use the underlying element, jurisdiction, and precedent fields to explain the decision.

Element Scores#

elementScores are confidence-weighted similarity findings for protected or semi-protected musical elements.

Element Legal review focus
melody Highest protectability. Review contour, phrase placement, hooks, access.
lyrics Literary expression. Review exact, translated, and paraphrased overlap.
vocal Timbre, delivery, and voice-adjacent identity signals.
structure Arrangement and selection/sequence patterns.
spectral Production, sample, or recording-similarity evidence.
rhythm Usually weaker alone; stronger when iconic or combined with other facts.
harmony Often thin protection; watch distinctive selection and arrangement.

For each element, review:

  • score and confidence;
  • referenced work and matched segment id;
  • explanation text from the analyzer;
  • whether the element is protectable, commonplace, licensed, public domain, or otherwise low-risk;
  • whether several weaker elements form a stronger compound match.

Compound Matches#

compoundMatches identify the same reference work appearing across several elements. A low or medium harmony/rhythm signal can become legally important when it appears with melody, lyrics, structure, or production similarity.

Prioritize compound matches by:

  1. number of matched elements;
  2. aggregate score;
  3. protectability of the matched elements;
  4. whether the matched segments align in time;
  5. commercial importance of the matched reference work;
  6. reviewer confidence and A/B evidence quality.

Jurisdiction-Specific Reports#

Music Shield applies jurisdiction-specific analyses and then aggregates them under a strictest-jurisdiction release rule.

Strictest-Jurisdiction Aggregation#

aggregateMultiJurisdictionRisk defaults global distribution to US, EU, and UK. high and critical are release-blocking by default. If a configured jurisdiction is missing, the report inserts a release-blocking missing_jurisdiction_assessment finding.

Review these aggregation fields first:

  • configuredJurisdictions - the territories covered by the release decision;
  • strictestJurisdiction - the jurisdiction driving the overall result;
  • releaseSafe - whether any configured jurisdiction blocks release;
  • missingJurisdictions - reports that must be generated before clearance;
  • perJurisdictionBreakdown - all individual risk levels, scores, tests, and findings.

Clearance for global distribution requires every configured jurisdiction to be non-blocking or separately cleared with documented conditions.

United States#

The US report uses arnstein_probative_similarity_access_substantial_similarity_with_ninth_circuit_extrinsic_intrinsic.

Important fields:

  • arnstein.probativeSimilarity - copying signal based on similarity strength;
  • arnstein.access - access evidence, including AI-specific presumption;
  • arnstein.strikingSimilarityApplied - high similarity can compensate for weak access evidence;
  • arnstein.substantialSimilarity - improper appropriation signal;
  • ninthCircuit.extrinsicScore - objective protectable-element analysis;
  • ninthCircuit.intrinsicReviewRequired - ordinary-audience review should be escalated;
  • ninthCircuit.thinCopyrightApplied and virtualIdentityRequired - low protectability elements require a stricter similarity showing.

US legal review should document:

  • access evidence or absence of access evidence;
  • protectable expression separated from commonplace or functional material;
  • whether the result is driven by melody/lyrics or by thin elements;
  • whether intrinsic review, senior musicologist review, or outside expert review is needed.

European Union#

The EU report uses eu_authors_own_intellectual_creation_with_dsm_articles_3_4_tdm.

Important fields:

  • originality.standard - authors' own intellectual creation;
  • originality.originalityPassed - originality threshold finding;
  • originality.protectedExpressionScore - protected expression strength;
  • originality.qualitativeTakingScore - qualitative taking assessment;
  • originality.elementFindings - free creative choices, personal stamp, and commonplace flags by element;
  • tdm.articleApplied - Article 3, Article 4, or no TDM exception;
  • tdm.exceptionAvailable - whether text-and-data-mining conditions are met;
  • tdm.rightsHolderOptOut and tdm.optOutEffect - opt-out evidence;
  • tdm.requirements - lawful access, retention, security, and purpose checks.

EU legal review should document:

  • whether the source material reflects protectable creative choices;
  • whether the detected taking is qualitative, not just quantitative;
  • TDM lawful access, opt-out status, and retention limitations;
  • member-state or national advice needed for high-value releases.

United Kingdom#

The UK report uses cdpa_section_16_substantial_part_qualitative_test.

Important fields:

  • copyingInference.passed - whether copying can be inferred from the signals;
  • substantialPart.passed - whether a substantial part is implicated;
  • substantialPart.qualitativeScore - importance of the copied part;
  • substantialPart.quantitativeScore - amount of the reference work implicated;
  • substantialPart.distinctivePartCopied - distinctive-part evidence;
  • substantialPart.wholeWorkOrNearWholeWorkCopied - near-whole-work flag;
  • substantialPart.elementFindings - element-level qualitative and quantitative evidence.

UK legal review should document:

  • why the copied part is or is not substantial;
  • whether the copied portion is distinctive even if short;
  • whether commonplace material reduces the qualitative score;
  • whether compound similarity creates cumulative substantial-part risk.

Moral Rights#

When moral-rights analysis is present, review attribution, integrity, and reputation findings separately from economic copyright risk. Moral-rights strength varies by jurisdiction. Residual risk can remain even when consent, waiver, or license evidence exists.

Escalate when:

  • attribution is missing, altered, or disputed;
  • the output distorts or mutilates a recognizable work or performance;
  • the release could prejudice the honor or reputation of a creator;
  • the relevant jurisdiction has strong integrity rights.

Case Law Citations#

Risk reports may include LegalPrecedentCitation records from the case-law database. Each citation contains:

  • caseId, caseName, year, court, and jurisdiction;
  • outcome, such as infringement, no_infringement, settled, or reversed;
  • keyElement, thresholdApplied, and similarityDescription;
  • estimatedSimilarity, similarityDistance, and relevanceScore;
  • damages, when available;
  • citationText and sourceUrl.

Use citations as analogies, not as automated holdings. For each cited case:

  1. Open the sourceUrl and verify the citation is still accessible.
  2. Confirm the jurisdiction is relevant to the planned distribution territory.
  3. Compare the cited element to the current matched element.
  4. Check whether the cited outcome is precedential, settled, reversed, or fact-specific.
  5. Note whether the current report has stronger or weaker access evidence, protectability, qualitative taking, or compound similarity.
  6. Record why the citation supports clearance, remediation, blocking, or escalation.

When the report has no relevant citation, document that absence. Lack of a citation does not mean low legal risk; it may mean the element, score, or jurisdiction does not map cleanly to stored case law.

The provenance chain is the canonical record for legal discovery, regulatory audit, and internal incident review.

ProvenanceChain includes:

  • generation - prompt, model id/version, checkpoint, seed, sampling parameters, source pipeline, stages, operators, and artifact ids;
  • scans - scan configuration, analyzers, jurisdictions, thresholds, matches, element scores, risk level, decision summary, diagnostics, and scan hash;
  • decisions - automated, human, escalation, modification, rejection, approval, and override records;
  • publication - distribution channels, published timestamp, and release artifact ids;
  • optional chain hash and per-record content hashes.

Export formats:

  • Markdown: human-readable report for counsel, business owners, and narrative review.
  • JSON: machine-readable export for e-discovery, evidence vaults, litigation tools, and external counsel.

The Markdown export includes generation parameters, training data provenance, scan results, the decision chain, and publication status. The JSON export includes exportVersion: provenance-export-v1, the raw chain, training data manifest, and opt-out issues.

Discovery Packet Checklist#

For legal discovery or a serious claim, preserve:

  • Markdown provenance export;
  • JSON provenance export;
  • decision audit report;
  • raw generated audio and all release artifacts;
  • normalized scan audio;
  • matched reference excerpts and metadata records;
  • analyzer diagnostics and confidence values;
  • reviewer notes, A/B comparison artifacts, and isolated element clips;
  • legal precedent citations and source PDFs or URLs;
  • training data manifest and opt-out issue report;
  • publication channel records and takedown/remediation history.

Apply litigation hold before regenerating, deleting, remastering, or replacing artifacts. If remediation creates a new version, keep the original chain and create a linked chain for the remediated work.

Decision Audit Trails#

Decision records are immutable once written. The Postgres audit table stores decision_id, scan_id, generation_id, content_hash, timestamp, decision_type, reviewer_id, and record_json.

Legal teams can retrieve:

  • a single decision by decision id;
  • decision history for a scan id;
  • decision history for a generation id;
  • a decision audit report from generation through scans, decisions, and publication.

Human decisions require:

  • reviewerId;
  • non-empty rationale;
  • conditions when approval is conditional;
  • original decision and override justification for human_override.

Review the content hash whenever a record appears inconsistent with the provenance export. A hash mismatch means the record is not legally reliable until the integrity issue is investigated.

Use these standards when recording legal advice or approval:

Outcome Required documentation
Clear for release Risk level, jurisdictions, evidence reviewed, provenance export id
Clear with conditions Required edits, territories, expiration, license terms, retained docs
Remediate before release Elements to change, reference work ids, required rescan
Block release Blocking jurisdictions, legal basis, evidence packet location
Escalate to outside counsel Open legal questions, deadlines, evidence packet, business owner
Override prior decision Original decision, override justification, authority, expiration

Every legal rationale should be specific enough that another reviewer can reconstruct the decision without private context. Avoid rationales like "approved by legal." Prefer concise findings tied to jurisdiction, element, score, citation, license, remediation, or provenance evidence.

Third-Party Claims#

Post-publication monitoring correlates claims with published works.

Claim outcomes:

  • record_false_negative - claim against a work that scanned as pass or advisory;
  • review_human_approval - claim against a flagged work that a reviewer approved or conditionally approved;
  • investigate_untracked_work - claim does not map to a tracked published work.

For each claim:

  1. Register litigation hold for the work and related artifacts.
  2. Pull the provenance chain, audit report, scan result, and review decisions.
  3. Compare the claim's matched work to the original scan matches.
  4. Determine whether the claim is a false negative, an approval-review issue, or a metadata/linkage issue.
  5. Record legal response deadline, claim source, claim type, matched work, and source URL.
  6. Feed validated false negatives and approval-review outcomes into threshold tuning.

Claims against approved flagged scans require review of the original human rationale. Confirm that all approval conditions were satisfied before publication.

The legal landscape monitor ingests new case decisions and legislative changes. Alerts are addressed to legal_team and include severity, jurisdiction, source URL, and threshold implications.

Review each alert for:

  • precedential weight and jurisdiction;
  • affected engines and current thresholds;
  • observed legal threshold, if supplied;
  • recommended action: lower_threshold, raise_threshold, or human_review;
  • urgency: low, medium, high, or critical;
  • whether release policy, reviewer training, or risk report language must be updated.

Critical or binding updates should trigger a threshold review and a sample audit of recently cleared works in affected jurisdictions.

Access And Retention#

Legal evidence is sensitive. Limit access to the smallest set of counsel, reviewers, and operators required for the matter.

Minimum controls:

  • preserve immutable provenance exports and content hashes;
  • store discovery packets in an access-controlled evidence location;
  • keep external counsel transfers logged;
  • retain claim, clearance, and override materials under the applicable legal or audit retention policy;
  • avoid editing reviewer rationales after decision recording; add a new decision or supplemental memo instead;
  • preserve source URLs or PDFs for cited precedent when relying on them.

Before signing off:

  • Packet includes scan, risk report, provenance export, and audit trail.
  • Configured distribution jurisdictions match the planned release.
  • No required jurisdiction report is missing.
  • Strictest jurisdiction is non-blocking or separately cleared.
  • Element and compound matches have been reviewed against protectability.
  • Case law citations are relevant, verified, and explained.
  • Reviewer rationales and conditions are complete.
  • Training data and opt-out issues were reviewed when relevant.
  • Publication, remediation, block, or escalation decision is recorded.
  • Evidence packet location and retention requirement are documented.