Chapter 3
A genome is not just another data field, and an edit to it is not just another intervention. This chapter explains how WIA-BIO-018 treats genetic information and gene editing: how data is protected and shared, what happens when sequencing reveals something nobody was looking for, and how the standard separates acceptable somatic therapy from heritable germline change.
Chapter 2 described the general machinery of human-subjects protection: disclosure, comprehension, capacity, IRB review and the extra safeguards owed to vulnerable groups. All of that still applies when a study sequences DNA or edits it. Genomics adds problems that the general machinery was not designed for. A consent form signed by one person can expose information about parents, siblings and children who never signed anything. A dataset stripped of names can still point back to an individual. A variant found by accident can carry life-changing news. And an edit made to an embryo does not stay with the person who consented; it passes to descendants who cannot be asked.
WIA-BIO-018 addresses these problems in two sections of the specification. Section 6 covers genetic data ethics: protection levels, security controls, discrimination, incidental findings and secondary use. Section 7 covers gene editing: the somatic/germline distinction, the conditions under which somatic editing is acceptable, the arguments around germline editing, and a six-step decision framework. The SDK expresses the same ideas in three record types, GeneticData, IncidentalFinding and GeneEditingEvaluation, which this chapter introduces at the level of meaning; Chapter 4 treats them as data structures and Chapter 5 traces the code that evaluates them.
Most personal data describes one person at one moment. Genetic data breaks both of those assumptions, and three properties explain why the standard gives it a section of its own.
Every person shares roughly half of their inherited DNA with each parent, each full sibling and each child. A pathogenic variant found in a research participant is therefore a probabilistic statement about relatives as well. This is the reason the spec's incidental-findings protocol includes a step labelled "Family Implications" and why IncidentalFinding carries an optional familyImplications field. Consent is given individually, but the information disclosed is partly collective, and a system that handles genetic data has to anticipate questions from people who are not participants.
A leaked password can be changed; a leaked credit-card number can be cancelled. A leaked genome cannot be reissued. Whatever it reveals about disease risk, ancestry or biological relationships stays true for the life of the person and, in part, for their descendants. The practical consequence is that the cost of a breach does not decay over time, and retention and access decisions have to be judged against a horizon of decades rather than months.
The genome is itself an identifier. Removing the name from a whole-genome file removes a label, not the identity. Published research has shown that surnames can be inferred from Y-chromosome markers by comparison with public genealogy databases (Gymrek et al., Science, 2013), and since 2018 law-enforcement investigations in the United States have publicly used consumer genealogy databases to identify suspects through distant relatives. Neither technique requires access to the original research record; both only need the DNA sequence and public reference data. For this reason the word "anonymous" must be used with great care for genomic data.
Section 6.1 of the spec defines four levels, and the SDK encodes them as the DataProtectionLevel type. The table below lists them together with the way they should be read for genomic material.
| Level | SDK value | Spec definition | Practical reading for genomic data |
|---|---|---|---|
| 1 | anonymous | No identifiers, cannot re-identify | Realistic for aggregate results (allele frequencies, summary statistics). Rarely achievable for individual-level sequence data, because the sequence itself identifies. |
| 2 | de-identified | Coded, can re-identify with key | The normal state for research datasets: direct identifiers replaced by a code, key held separately by a custodian. Legally this is often pseudonymised data, still personal data under GDPR. |
| 3 | confidential | Identified but protected | Clinical records, return-of-results workflows and re-contact lists, where the identity is needed to act on the data. |
| 4 | public | Openly shared | Appropriate only when participants explicitly consented to open release, and when the re-identification consequences were explained to them. |
Two points are worth drawing out. First, the numbering is not a ranking of strictness: Level 4 is the least protected, while Level 1 is the least identifiable. A system should not assume that a higher number means more protection. Second, the protection level describes a copy of the data, not the participant. The same genome may exist as a Level 3 record in the clinical return-of-results system, a Level 2 record in the analysis environment, and a Level 1 contribution to a published frequency table. Each copy needs its own GeneticData record so that its controls and permitted uses can be audited separately.
The spec's "Required Protection" list for genetic data sets a baseline of five controls. They are unremarkable individually; the value is in requiring all of them together and recording which are in place.
| Control (spec §6.1) | What it means in practice | Where it is recorded |
|---|---|---|
| Encryption: AES-256 minimum | Data at rest encrypted with AES using a 256-bit key (AES is specified in NIST FIPS 197). Key management matters more than the algorithm: keys belong in a key-management service, not beside the data. | GeneticData.encryption |
| Access control: role-based, audit logs | Access is granted to roles (analyst, genetic counsellor, data manager), not to individuals ad hoc, and every read of individual-level data is logged. | GeneticData.accessControls |
| Storage: secure servers, off-site backup | Backups inherit the same encryption and access rules; an unencrypted backup undoes the rest of the list. | System documentation (Chapter 7) |
| Transmission: encrypted channels only | Transport encryption for every transfer, including internal transfers between the sequencing lab and the analysis cluster. | System documentation (Chapter 7) |
| De-identification: remove 18 HIPAA identifiers | The HIPAA Privacy Rule "Safe Harbor" method (45 CFR 164.514(b)(2)) lists 18 identifier categories, from names and small geographic units to dates and device serial numbers. | Protection level set to de-identified or anonymous |
Safe Harbor was designed for clinical records in which the identifying information sits in labelled fields. Removing those fields from a genomic dataset is necessary but not sufficient. The sequence remains, and as Section 3.1 showed, the sequence can be linked to people through relatives and public databases. HIPAA itself offers a second route, "Expert Determination", in which a qualified expert assesses the actual re-identification risk of a specific dataset in its intended context. For individual-level genomic data this risk-based approach is the more honest one, and the book recommends that BIO-018 implementations record which method was used rather than only asserting that data is de-identified.
Because technical de-identification cannot fully protect individual-level genomes, the research community relies heavily on governance controls in addition to technical ones: controlled-access repositories in which a data access committee reviews each request, data use agreements that forbid re-identification attempts, and restrictions on onward sharing. None of these is named explicitly in the BIO-018 spec, but they are how the spec's dataSharingConsent flag becomes enforceable in practice. A true value means the participant agreed to sharing; it does not mean the data may be posted openly.
The following example shows a GeneticData record for a coded exome dataset held in a research analysis environment. The identifiers and control names are illustrative.
{
"dataId": "GD-2026-0147",
"participantId": "P-00412",
"dataType": "exome",
"protectionLevel": "de-identified",
"encryption": "AES-256-GCM at rest; TLS in transit",
"accessControls": [
"role:research-analyst (read, coded data only)",
"role:key-custodian (re-identification key, separate system)",
"audit:all-reads-logged"
],
"secondaryUseConsent": true,
"dataSharingConsent": false,
"returnOfResults": "actionable"
}Read as an ethics record, this says: the participant allowed secondary research use but not sharing outside the institution; only actionable findings will be returned; and the key that links P-00412 to a person is held by a separate custodian. An auditor can check each of those claims against system logs, which is the point of writing them down.
Section 6.2 of the spec lists five prohibited uses of genetic information: employment discrimination, health insurance discrimination, life insurance discrimination ("varies by jurisdiction"), educational discrimination and social stigmatisation. It then names GINA, GDPR and "local genetic privacy laws" as legal frameworks. This list describes what a responsible research program should never enable. It is not a description of what the law actually forbids everywhere, and the gap between the two is exactly what consent forms must explain honestly.
The Genetic Information Nondiscrimination Act of 2008 has two operative parts. Title I restricts health insurers from using genetic information in eligibility and premium decisions. Title II restricts employers (those with 15 or more employees) from using genetic information in hiring, firing and other employment decisions. GINA does not cover life insurance, disability insurance or long-term-care insurance. A consent form that tells US participants "the law protects you from genetic discrimination" without that qualification overstates the protection.
The General Data Protection Regulation defines genetic data (Article 4(13)) and places it among the special categories of personal data whose processing is prohibited unless a specific condition in Article 9 applies, such as explicit consent or scientific research with appropriate safeguards under Article 89. GDPR is a data protection law rather than an anti-discrimination law: it controls how genetic data is collected, used and shared, and it gives the data subject rights of access and erasure, but it does not by itself regulate how an insurer prices a policy based on information the applicant disclosed voluntarily.
Many countries add their own rules. Korea's Bioethics and Safety Act, for instance, prohibits discrimination on the basis of genetic information in areas such as education, employment, promotion and insurance, and regulates genetic testing institutions. Other countries rely on voluntary insurance-industry agreements rather than statute. The BIO-018 system does not need to encode every national rule, but the consent form must be written for the jurisdiction in which participants are enrolled.
| Area of potential harm | GINA (US) | GDPR (EU) | Implication for consent text |
|---|---|---|---|
| Health insurance | Covered (Title I) | Controls processing of genetic data; not an insurance-pricing rule | State the protection and its source law. |
| Employment | Covered for employers with 15+ employees (Title II) | Processing restricted; employment law is national | Mention the employer-size limit where relevant. |
| Life, disability, long-term-care insurance | Not covered | Depends on national law | Say plainly that protection may not exist. |
| Education, social stigma | Not covered | Indirectly, through limits on processing and disclosure | Describe the confidentiality measures instead of promising legal protection. |
Design rule. The spec's list of prohibited uses is a policy commitment for the research program. The consent form must separate what the program promises (for example, never releasing identifiable results to employers) from what the law guarantees. Participants weigh these risks when deciding whether to join, so an inflated description of legal protection undermines the voluntariness of the decision.
The spec defines incidental findings as "clinically significant findings beyond research scope". A study sequencing exomes to investigate a cardiac trait may, for example, detect a pathogenic variant in a cancer-predisposition gene. Such findings are not rare in genome-scale sequencing, because every analysis examines thousands of genes even when the research question concerns a handful. Clinical genomics often distinguishes incidental findings (found by chance) from secondary findings (deliberately searched for in a defined list of genes); the American College of Medical Genetics and Genomics (ACMG) maintains such a list of genes for which it recommends reporting in clinical sequencing. BIO-018 uses one term for both, and the record type is the same.
GeneticData.returnOfResults, with the values all, actionable, requested or none.Section 6.3 sorts findings into three return categories. The table restates them and adds the field of IncidentalFinding that captures each condition.
| Category | Conditions (spec §6.3) | Fields that express the conditions |
|---|---|---|
| Should return | Clinically significant AND actionable (treatment or prevention available) AND high validity (confirmed) AND participant wants to know | significance is pathogenic or likely-pathogenic; actionable: true; clinicallyConfirmed: true; participantWantsToKnow: true |
| May return | Significant but not actionable; OR uncertain significance; OR participant specifically requested | actionable: false, or significance: "uncertain", or a recorded specific request |
| Should not return | Non-actionable; OR low validity; OR participant declined to know | participantWantsToKnow: false overrides all other fields; unconfirmed results are not returned as clinical facts |
The matrix is not a strict partition as written: "significant but not actionable" appears under may return, while "non-actionable" appears under should not return. The sensible reading is that a finding that is significant and not actionable falls to may return if the participant wants it and to should not return otherwise. A second point concerns variants of uncertain significance. They are listed as "may return", but in practice most programs do not return them, because they cannot guide clinical action and may be reclassified later. A protocol that chooses to return them should explain in the consent form that they are uncertain and may change.
The IncidentalFinding.significance field uses the five-tier terminology from the ACMG/AMP variant-interpretation guidelines (Richards et al., Genetics in Medicine, 2015): pathogenic, likely-pathogenic, uncertain, likely-benign and benign. Using this shared vocabulary means that the classification recorded by the laboratory can be carried into the ethics record without translation. It also means that the BIO-018 system should record which laboratory made the classification, and when, since classifications are periodically revised as evidence accumulates.
shouldReturnIncidentalFinding has a boolean shouldReturn field. The spec does not say how it is computed, and the SDK does not compute it; it is simply stored. The book recommends deriving it from the other fields so that the decision is reproducible and auditable, rather than letting it be set by hand. A minimal implementation of the spec's matrix could look like this:
import type { IncidentalFinding, GeneticData } from './types';
type ReturnCategory = 'should-return' | 'may-return' | 'should-not-return';
export function classifyReturn(
f: IncidentalFinding,
policy: GeneticData['returnOfResults'],
specificallyRequested = false
): ReturnCategory {
// The participant's refusal to know overrides every other consideration.
if (f.participantWantsToKnow === false || policy === 'none') {
return 'should-not-return';
}
// Research-grade results are never returned as clinical facts.
if (!f.clinicallyConfirmed) return 'should-not-return';
const significant =
f.significance === 'pathogenic' || f.significance === 'likely-pathogenic';
if (significant && f.actionable && f.participantWantsToKnow === true) {
return 'should-return';
}
if (
(significant && !f.actionable) ||
f.significance === 'uncertain' ||
specificallyRequested
) {
// Honour the participant's chosen policy for this group.
if (policy === 'all') return 'may-return';
if (policy === 'requested' && specificallyRequested) return 'may-return';
return 'should-not-return';
}
return 'should-not-return';
}Two design choices in this function are deliberate. An absent participantWantsToKnow (the field is optional) does not produce should return: the matrix requires an affirmative wish. And an unconfirmed finding is not returned; it is first sent for clinical confirmation, after which the record is updated and the function is run again. The "right not to know" is part of respect for persons, and the code must not let a well-meaning default override it.
An illustrative record for a confirmed, actionable finding looks like this:
{
"findingId": "IF-2026-0031",
"participantId": "P-00412",
"gene": "BRCA2",
"significance": "pathogenic",
"actionable": true,
"shouldReturn": true,
"participantWantsToKnow": true,
"clinicallyConfirmed": true,
"familyImplications": "First-degree relatives have a 50% chance of carrying the variant; cascade testing offered via the participant",
"geneticCounselingProvided": true
}The 50% figure here is not a statistic about any study; it is the Mendelian probability that a first-degree relative shares a single autosomal dominant variant with a heterozygous carrier.
Genomic datasets are expensive to generate and valuable long after the original study ends. Biobanks exist precisely so that samples and data collected once can support many future questions. Section 6.4 of the spec sets five requirements for such secondary use:
This is why the tenth disclosure element in Chapter 2, "future use of data and samples", carries so much weight. What participants agreed to at enrolment determines which later studies can proceed without contacting them again.
| Model | What the participant agrees to | Strength | Weakness |
|---|---|---|---|
| Specific consent | One named study | Maximum clarity about use | Every new question requires re-contact or a waiver |
| Broad consent | A described category of future research, under governance | Makes biobanking feasible; the revised US Common Rule provides for it explicitly (45 CFR 46.116(d)) | Participants cannot foresee specific future uses; depends on trust in governance |
| Tiered consent | A menu of options (for example, disease research yes, commercial use no, international sharing yes) | Respects differing preferences | Each tier must be enforced on every data release; more complex to administer |
In the EU, GDPR's purpose-limitation principle (Article 5(1)(b)) states that further processing for scientific research is not considered incompatible with the original purpose, provided the safeguards of Article 89 are in place. This softens but does not remove the need for a lawful basis and for ethics review of each new use.
The GeneticData record offers only two booleans, secondaryUseConsent and dataSharingConsent. That is enough to encode the core of §6.4 as a gate that every request for data must pass, with the richer detail of tiered consent held in the consent record itself:
type SecondaryUseCheck =
| { allowed: true; conditions: string[] }
| { allowed: false; action: 'recontact' | 'obtain-irb-approval' | 'deny' };
export function checkSecondaryUse(
data: GeneticData,
req: { irbApproved: boolean; withinConsentScope: boolean; externalRecipient: boolean }
): SecondaryUseCheck {
if (!data.secondaryUseConsent || !req.withinConsentScope) {
return { allowed: false, action: 'recontact' }; // §6.4 item 5
}
if (!req.irbApproved) {
return { allowed: false, action: 'obtain-irb-approval' }; // §6.4 item 2
}
if (req.externalRecipient && !data.dataSharingConsent) {
return { allowed: false, action: 'deny' };
}
const conditions = ['Purpose limited to the approved protocol'];
if (data.protectionLevel === 'confidential') {
conditions.push('Release a coded (de-identified) copy, not this record'); // §6.4 item 3
}
return { allowed: true, conditions };
}Re-contact itself is an ethical act that needs care. Participants may have changed address, lost capacity or died; some asked at enrolment never to be contacted again. A re-contact attempt should therefore be logged as an event, with its outcome, and never used as a way to deliver unsolicited results.
Section 7 of the spec opens with the distinction that organises the whole field of gene-editing ethics.
| Aspect | Somatic | Germline |
|---|---|---|
| Target | Body (non-reproductive) cells | Reproductive cells, embryos |
| Heritability | Not inherited | Inherited by offspring |
| Ethical status | Generally acceptable | Controversial / restricted |
| Current use | Clinical trials and approved therapies | Research only in most jurisdictions |
| Reversibility | Effects confined to the individual | Irreversible for the lineage |
| Who bears the risk | The consenting patient | Future persons who cannot consent |
| SDK value | editingType: "somatic" | editingType: "germline" |
The last two rows are added by this book. The first is the ethical crux: somatic editing fits the ordinary model of research, in which the person who bears the risk is the one who consents. Germline editing does not. The spec's "Current Use" row says "Clinical trials" for somatic editing; since late 2023 this has moved further, with regulators in the United Kingdom and the United States approving the first CRISPR-based therapy (exagamglogene autotemcel) for sickle cell disease. That approval does not change the ethical analysis, but it does mean that somatic editing is now a clinical reality rather than a hypothetical.
Section 7.2 lists six conditions, all of which must hold:
Each maps directly to a field of GeneEditingEvaluation: purpose; ethical.benefitsOutweighRisks; ethical.noAlternatives; ethical.consentObtained; scientificValidity.preclinicalEvidence together with mechanismUnderstood and safetyDemonstrated. Long-term monitoring is not an input flag; it is something the decision imposes as a condition.
The spec names five concerns, four of which reappear as numeric fields in GeneEditingEvaluation.risks, each on a 0–1 scale:
offTargetEffects): the editing tool cuts or modifies DNA at sites other than the intended one, possibly disrupting another gene.immuneResponse): immune reactions to the editing components, the delivery vehicle or the edited cells.mosaicism): only some cells are edited, or cells are edited differently. For ex vivo therapies, where cells are edited outside the body and then returned, this appears as incomplete editing efficiency; in embryos it is a more serious concern, because the resulting person may carry a mixture of edited and unedited cell lines.longTermUnknown): consequences that only appear years later.ethical.equitableAccess because it is a justice concern rather than a safety risk.Section 7.2 also prints a "Somatic_Risk_Score" formula that sums probability-times-severity terms and divides by a "Normalization_Factor". The spec never defines that factor, so the formula cannot be computed as written. It should be read qualitatively: risk grows with both the likelihood and the seriousness of each harm, and it must be judged against the expected benefit. The SDK takes a simpler route, averaging the four 0–1 risk fields and treating an average above 0.5 as requiring enhanced mitigation; Chapter 5 traces that logic.
Appendix A.1 of the spec analyses a hypothetical CRISPR-based therapy for sickle cell disease. Its numbers are illustrative weights chosen by the spec's authors, not data from any trial. Laid out as a table, the calculation is:
| Item | Probability | Severity / weight | Score |
|---|---|---|---|
| Benefit: cure of severe genetic disease | not applied | 10 | 10 |
| Benefit: improved quality of life | not applied | 8 | 8 |
| Benefit: reduced healthcare burden | not applied | 6 | 6 |
| Benefit: generalisable knowledge | not applied | 4 | 4 |
| Total benefit | 28 | ||
| Risk: off-target effects | 0.05 | 8 | 0.4 |
| Risk: immune response | 0.1 | 6 | 0.6 |
| Risk: treatment failure | 0.2 | 4 | 0.8 |
| Risk: long-term unknown | 0.3 | 5 | 1.5 |
| Total risk | 3.3 | ||
| Ratio | 28 / 3.3 ≈ 8.5 → "favourable; approve with enhanced monitoring" | ||
The arithmetic is correct. The method deserves scrutiny, because it illustrates how easily a single number can carry hidden assumptions:
None of this makes the analysis useless. Writing the weights down forces reviewers to state and debate their assumptions, which is better than an unexamined gut feeling. But the ratio is a structured argument for an IRB to examine, not a verdict. Chapter 8 returns to this limit of quantified ethics.
Monitoring horizon. When the SDK approves a somatic therapeutic proposal, it attaches "Long-term follow-up (minimum 5 years)". That is the SDK's default, not a regulatory standard. The US FDA's guidance on long-term follow-up after administration of gene therapy products recommends follow-up of up to 15 years for products such as genome-editing products, depending on their risk profile. Implementations should treat 5 years as a floor and apply the applicable regulator's requirement.
Germline editing changes eggs, sperm, their precursors or early embryos, so that the change is carried by every cell of the resulting person and can be passed on to their children. The spec's summary of the current consensus is a single line: a moratorium on clinical applications. That position is reflected internationally. The Council of Europe's Convention on Human Rights and Biomedicine (the Oviedo Convention, 1997) permits interventions on the human genome only for preventive, diagnostic or therapeutic purposes and only if they do not aim to introduce a modification in the genome of any descendants. Korea's Bioethics and Safety Act prohibits gene therapy on embryos, eggs, sperm and fetuses. In 2018 a researcher in China announced the birth of twins from edited embryos; the work was widely condemned, and he was later criminally convicted. The Third International Summit on Human Genome Editing (London, 2023) concluded that heritable human genome editing remains unacceptable at this time, and in 2021 the World Health Organization issued recommendations on governance and oversight of human genome editing.
| Against (spec §7.3) | For, in limited cases (spec §7.3) |
|---|---|
| Unknown risks: long-term, multi-generational effects cannot be observed in advance. | Prevent suffering: eliminate a severe genetic disease from a family line. |
| No consent: the people most affected do not yet exist. | Family choice: reproductive autonomy of prospective parents. |
| Irreversibility: a mistake cannot be undone for descendants. | Scientific progress: understanding of early human development. |
| Enhancement: a slippery slope to non-therapeutic uses. | Global health: reduce the burden of inherited disease. |
| Justice: access limited to the wealthy could widen inequality. | |
| Eugenics: the history of state-sponsored "improvement" programmes. |
A point that often decides the argument in practice: for most couples at risk of passing on a serious single-gene disorder, preimplantation genetic testing, which selects unaffected embryos without editing any of them, is already available. The cases in which germline editing would be the only way to have an unaffected genetically related child are few. This is one reason the "no alternatives" condition weighs so heavily against clinical germline editing today.
The spec does not close the door completely. It lists five conditions under which a future use might become acceptable: a serious genetic disease with no other options, proven safety from extensive preclinical evidence, international consensus, robust oversight, and a strictly therapeutic purpose. All five must hold, and at present the international-consensus condition is not met. The spec also lists uses that are prohibited outright: enhancement of normal traits, sex selection for non-medical reasons, cosmetic modifications, performance enhancement and intelligence augmentation. The boundary between therapy and enhancement is not always sharp, which is why the SDK asks for a declared purpose and why an IRB must examine the declaration rather than accept it.
Laboratory research on edited embryos that are never transferred for pregnancy is a different activity from clinical germline editing, and many jurisdictions allow it under strict conditions. The SDK reflects this with the purpose: "research" value: for germline research it adds the conditions "Laboratory research only - no clinical application", "No implantation of edited embryos" and "Embryo destruction within 14 days". The 14-day limit is a long-standing international convention for embryo culture. The International Society for Stem Cell Research, in its 2021 guidelines, removed it from the category of absolutely prohibited research and recommended case-by-case review of proposals to go beyond it, while national laws in many countries still fix the limit. It should be recorded as the applicable legal or institutional limit rather than as a universal rule. (Chapter 5 notes a defect: in the current SDK, these research conditions are computed but not returned.)
Section 7.4 turns the principles into a process. The table shows each step, the questions it asks, and the fields of GeneEditingEvaluation that record the answers.
| Step | Questions (spec §7.4) | Recorded in |
|---|---|---|
| 1. Classify the intervention | Somatic or germline? Therapeutic or enhancement? Serious disease or minor condition? | editingType, purpose, target |
| 2. Assess scientific validity | Preclinical evidence? Mechanism understood? Safety demonstrated? | scientificValidity.* |
| 3. Evaluate risk-benefit | Benefits to the individual? Risks acceptable? Alternatives available? | risks.*, ethical.benefitsOutweighRisks, ethical.noAlternatives |
| 4. Consider justice | Equitable access? Vulnerable populations protected? Global implications? | ethical.equitableAccess; population data in the IRB submission |
| 5. Ensure oversight | IRB approval? Regulatory compliance? Public engagement? | regulatoryCompliance, internationalConsensus |
| 6. Obtain consent | Informed consent? Capacity confirmed? Voluntariness assured? | ethical.consentObtained, linked consent record (Chapter 4) |
The framework ends with three possible outcomes: proceed if all steps are satisfied and benefits exceed risks; defer if uncertainty is too high; reject if ethical principles are violated. In the SDK these become the values of GeneEditingDecision.decision, which the type declares as approve, approve-research-only, defer and reject. The decision always comes with a rationale and, where relevant, conditions, oversight bodies and monitoring requirements, so that the reasoning survives alongside the result.
Applied to the combinations of type and purpose, the framework and the current SDK produce the following outcomes. Where the two differ, the table says so.
| Therapeutic | Enhancement | Research | |
|---|---|---|---|
| Somatic | Approve if every validity, consent and risk check passes; otherwise defer or reject with conditions | Reject | approve-research-only after the same validity and consent checks, with the condition "No clinical application without separate approval" |
| Germline | Defer (moratorium; awaits international framework) | Reject | Intended: laboratory research only, no implantation, embryo culture limit. Current SDK: generic defer (see Chapter 5) |
The following record describes a hypothetical somatic, ex vivo therapy for sickle cell disease. The risk values are illustrative inputs on the SDK's 0–1 scale, not measured rates.
{
"evaluationId": "GEE-2026-004",
"editingType": "somatic",
"purpose": "therapeutic",
"target": "sickle cell disease (ex vivo edited autologous stem cells)",
"scientificValidity": {
"preclinicalEvidence": true,
"mechanismUnderstood": true,
"safetyDemonstrated": true
},
"risks": {
"offTargetEffects": 0.2,
"immuneResponse": 0.1,
"mosaicism": 0.3,
"longTermUnknown": 0.5
},
"ethical": {
"consentObtained": true,
"benefitsOutweighRisks": true,
"noAlternatives": true,
"equitableAccess": false
},
"regulatoryCompliance": true
}The average of the four risk values is (0.2 + 0.1 + 0.3 + 0.5) / 4 = 0.275, below the SDK's 0.5 threshold, and all validity and consent checks pass, so the SDK returns approve. Because equitableAccess is false, the decision carries the condition "Develop plan for equitable access to therapy". That condition is the SDK's way of keeping justice on the table even for a therapy that is otherwise acceptable, and it is a good example of why a decision record must be read in full rather than reduced to its first word.
shouldReturn should be derived from recorded facts, with the participant's wish not to know taking priority.Chapter 4 now turns from ethics to data structures, and shows how the consent, risk, genomic and safety records fit together into a single auditable model.