Safety Screen

This skill teaches you how to evaluate hypotheses for dual-use biosafety risk and research integrity violations (evidence fabrication and phantom citations), and how to quarantine hypotheses that pose potential dangers or violate integrity standards. Safety screening is a mandatory part of every review — it is never skipped, even for seemingly benign hypotheses.

This protocol contributes records to a run that the supervisor ultimately publishes with oase hypex analyze; workers do not publish a separate result.

Dual-Use Biosafety Rubric

A hypothesis has dual-use risk when the research it proposes could be directly misused or could provide knowledge enabling the creation of biological, chemical, or radiological weapons, or other instruments of mass harm.

Concerning Topics

The following research areas require heightened scrutiny. A hypothesis involving any of these topics must receive careful safety evaluation:

1. Gain-of-function modifications — Proposals to increase the transmissibility, virulence, host range, or immune evasion capability of any pathogen. This includes computational predictions of gain-of-function mutations even without wet-lab work.
2. Synthesis of dangerous pathogens — Proposals to reconstruct, synthesize, or reverse-engineer select agents or other dangerous pathogens (e.g., variola, H5N1, reconstructed 1918 influenza).
3. Weaponizable delivery mechanisms — Research on aerosolization, environmental persistence, or targeted delivery methods that could be repurposed to disseminate biological or chemical agents.
4. Immune evasion engineering — Proposals to modify organisms or molecules to evade host immune responses, vaccines, or therapeutic antibodies — particularly when applied to pathogens rather than for therapeutic purposes (e.g., CAR-T engineering is therapeutic; engineering a virus to evade vaccine-induced immunity is dual-use).
5. Toxin enhancement — Research to increase the potency, stability, or deliverability of biological toxins (botulinum, ricin, etc.).
6. Antimicrobial resistance transfer — Proposals involving the transfer of antibiotic resistance genes to new organisms, particularly via mobile genetic elements, unless the explicit purpose is to study resistance mechanisms for mitigation.
7. Neurological or behavioral modification agents — Research on agents designed to alter cognition, behavior, or consciousness in ways that could be coercive or weaponized.
8. Environmental release of engineered organisms — Proposals for releasing genetically modified organisms with potential for uncontrolled spread or ecological disruption.

Risk Assessment Matrix

Use this matrix in conjunction with the safety scoring anchors from the review-rubric skill:

Research Integrity & Evidence Fabrication (Phantom Citations)

In addition to biosafety risks, scientific integrity is strictly enforced across the pipeline. Fabricating evidence, citing hallucinated literature (non-existent PMIDs or DOIs), or pairing real identifiers with fabricated titles undermines scientific reasoning and corrupts downstream evolutionary crossover.

Automated citation verification is performed using:

oase cite verify <run-dir>/hypotheses/H-XXXX.json --out raw/citations

The resulting manifest records per-citation status (verified, phantom, unverified) and an overall summary.

Quarantine Triggers

A hypothesis MUST be immediately quarantined if ANY of the following triggers are met:

1. Dual-Use Biosafety Risk (Safety Score ≤ 2)

Use the safety axis anchors from the review-rubric skill to assign a score from 1 to 5.

• Score ≥ 3: No quarantine required. Note any safety considerations in the review verdict and key criticisms, but the hypothesis proceeds normally.
• Score = 2: Triggers quarantine. Notable dual-use potential that requires institutional review. The hypothesis may still have scientific merit, but it cannot proceed through the standard pipeline without oversight.
• Score = 1: Immediate quarantine. Clear dual-use risk. The hypothesis is removed from the active pipeline.

When in doubt between scores 2 and 3, choose 2 (quarantine). False positives are recoverable — the supervisor can review quarantined hypotheses and release them if appropriate. False negatives cannot be undone.

2. Evidence Fabrication / Phantom Citations

A hypothesis MUST be immediately quarantined if the citation manifest (<run-dir>/citations/H-XXXX.json) meets ANY of these conditions:

• phantom_count >= 2, OR
• phantom_count / total_citations >= 0.50, OR
• total_citations > 0 AND verified_count == 0.

(Note: Citations with status: "unverified" due to network timeout or upstream API errors do NOT count toward phantom_count and do not trigger quarantine.)

Quarantine Procedure

When a hypothesis triggers quarantine (via safety score ≤ 2 or critical phantom citations), execute the following steps in order:

Step 1: Complete the Review

Finish scoring all axes and write the full review JSON. Do not truncate the review just because the hypothesis will be quarantined — the review record is needed for downstream audit:

• For dual-use biosafety concerns, document the safety rationale in the review's verdict and key_criticisms fields.
• For phantom citations / evidence fabrication:
  • Assign scores.correctness: 1.
  • Populate phantom_citations with all phantom lit_ids.
  • Set citation_manifest to "citations/H-XXXX.json".
  • Record the finding in key_criticisms: "CRITICAL PHANTOM CITATIONS: <N> phantom references detected in manifest. Correctness assigned 1 and hypothesis quarantined."

Step 2: Write the Review File

Write the review using hypex add-review --run <run-id> --run-dir <run-base> --for <hypothesis-id>. This assigns the next review ID, validates against review.schema.json, and writes atomically to reviews/H-XXXX.R-NN.json.

Step 3: Quarantine the Hypothesis

Use hypex set-status to update the hypothesis status and move it to quarantine:

hypex set-status --run <run-id> --run-dir <run-base> <hypothesis-id> quarantined

This updates the status field to "quarantined", re-validates against schemas/hypothesis.schema.json, and atomically moves the file from hypotheses/<hypothesis-id>.json to quarantine/<hypothesis-id>.json. This removes the hypothesis from the active pool so other agents (ranking, evolution) scanning hypotheses/ will not encounter it.

Step 4: Log the Quarantine Event

Write a quarantine log entry to stdout so the supervisor can detect it. Use the standard format corresponding to the quarantine trigger:

• For dual-use biosafety concerns:

  QUARANTINE: H-XXXX quarantined by <reviewer-name>. Safety score: <N>. Reason: <brief reason>.

  Example:

  QUARANTINE: H-0042 quarantined by reflection-agent-1. Safety score: 1. Reason: Proposes gain-of-function modification of H5N1 to increase aerosol transmissibility.

• For critical phantom citations / evidence fabrication:

  QUARANTINE: H-XXXX quarantined by <reviewer-name>. Reason: Critical phantom citations detected (<N> phantom references in manifest).

  Example:

  QUARANTINE: H-0042 quarantined by reflection-agent-1. Reason: Critical phantom citations detected (2 phantom references in manifest).

The supervisor monitors for QUARANTINE: prefixed lines in agent output to detect quarantine events.

Step 5: Record in Completion Report

When reporting completion to the supervisor, separately list any quarantine events with:

• The hypothesis ID
• The reason (dual-use biosafety score or critical phantom citation counts)
• A one-sentence explanation of the concern

Screening Workflow

Apply this screening workflow to every hypothesis you review, regardless of how benign it appears:

1. Automated Citation Verification & Phantom Check:
   • Run automated citation verification on the hypothesis:

     oase cite verify <run-dir>/hypotheses/H-XXXX.json --out raw/citations

   • Inspect the manifest summary (summary.phantom, summary.verified, summary.total).
   • If phantom_count >= 2, OR phantom_count / total_citations >= 0.50, OR (total_citations > 0 AND verified_count == 0):
     • Trigger immediate quarantine for evidence fabrication.
     • Complete review with scores.correctness: 1, phantom_citations, and citation_manifest.
     • Execute the quarantine procedure above.
2. Read the hypothesis — focus on mechanism, experiments, and predictions fields.
3. Check against the concerning topics list — does the hypothesis touch on any of the eight categories above?
4. If no concerning topics are triggered, assign safety score 4 or 5 using the calibrated anchors and move on.
5. If a concerning topic is triggered, apply the risk assessment matrix: a. Assess the pathogen class (if applicable). b. Assess the modification type. c. Assess the knowledge barrier — would the hypothesis, if published as a paper, provide actionable instructions for harm? d. Assess the scope and reversibility of potential harm.
6. Assign the safety score using the calibrated anchors.
7. If score ≤ 2, execute the quarantine procedure above.

Edge Cases

• Therapeutic research on pathogens (e.g., vaccine development, antiviral mechanism studies): Usually scores 3–4. The intent is defensive, and the knowledge produced is more useful for protection than for harm. Score 3 if the research involves BSL-3 organisms or could reveal evasion mechanisms.
• Antimicrobial resistance studies (for mitigation): Usually scores 3–4. Understanding resistance mechanisms is essential for public health. Score 2 only if the hypothesis proposes creating novel resistance transfer mechanisms rather than studying existing ones.
• Computational gain-of-function prediction: Scores 2–3. Even without wet-lab work, computational predictions of enhancing mutations can guide malicious actors. Score 2 if the predictions are specific enough to be directly actionable.
• Gene drives and environmental release: Scores 2–3. Self-propagating genetic modifications with ecological consequences require careful review. Score 2 if the proposal involves release in uncontrolled environments.

Examples

Example 1: Safe Hypothesis (Score 5)

"Gut microbiome composition correlates with cognitive decline rate in Alzheimer's patients, mediated by short-chain fatty acid levels."

• Observational study design
• No pathogen modification
• No dual-use concern
• Safety score: 5

Example 2: Minimal Concern (Score 4)

"CRISPR-Cas9 editing of APOE4 to APOE3 in iPSC-derived neurons reduces amyloid-beta accumulation."

• Standard gene editing techniques (therapeutic intent)
• BSL-1/2 work with human cell lines
• Minimal dual-use potential
• Safety score: 4

Example 3: Moderate Concern (Score 3)

"Immune evasion mechanisms of SARS-CoV-2 variants can be predicted by structural modeling of spike protein mutations."

• Touches on immune evasion (concerning topic #4)
• But: computational study, no wet-lab modification
• Knowledge is primarily useful for vaccine updates (defensive)
• Safety score: 3 — note the dual-use consideration in the review

Example 4: Quarantine Required (Score 2)

"Engineering Pseudomonas aeruginosa with enhanced biofilm resistance to identify novel antibiotic targets."

• Creates an organism with enhanced resistance (concerning topic #6)
• Intent is therapeutic (finding targets), but the engineered organism itself is more dangerous than wild-type
• Knowledge barrier: the methods section would describe how to create a more resistant organism
• Safety score: 2 — quarantine; requires institutional DURC review

Example 5: Immediate Quarantine (Score 1)

"Modifying H5N1 hemagglutinin to enable efficient aerosol transmission between ferrets as a model for pandemic risk assessment."

• Direct gain-of-function on a select agent (concerning topics #1 and #2)
• High scope of harm (pandemic potential)
• Irreversible if released
• Safety score: 1 — immediate quarantine

Example 6: Evidence Fabrication Quarantine (Phantom Citations)

Hypothesis proposes a mitochondrial pathway in lupus nephritis, citing two papers: PMID:99999999 (non-existent record) and PMID:42297600 with a fabricated title ("Direct Mitochondrial Repair in Podocytes via Synthetic Peptides").

• oase cite verify generates a manifest whose summary reports phantom: 2, verified: 0, total: 2.
• Quarantine triggers met: phantom_count >= 2, phantom_count / total >= 0.50, and verified_count == 0.
• Review completed with scores.correctness: 1, phantom_citations: ["PMID:99999999", "PMID:42297600"], and citation_manifest: "citations/H-0042.json".
• Hypothesis moved from hypotheses/H-0042.json to quarantine/H-0042.json with status: "quarantined".
• Log: QUARANTINE: H-0042 quarantined by reflection-agent-1. Reason: Critical phantom citations detected (2 phantom references in manifest).