FID-105 · Open question
When Do Safety Evaluations Change Deployment Decisions?
When a safety evaluation crosses a threshold or leaves consequential uncertainty, what public evidence shows that the result affected a release, restriction, mitigation, or decision to gather more evidence?
Why the question remains open
A published score, a commitment to act on a score, and an implemented response are different records. The decision may also change because of an incident, a new mitigation, a revised test, or an exception to the policy. Independent reviewers need a way to reconstruct what happened and to distinguish evidence that informed the decision from gaps in the public account.
Working hypothesis
A proposition to test, not a finding.
A review that links the versioned policy, evaluation, decision owner, action and follow-up record will distinguish documented influence from coincident timing more reliably than a review of system cards or threshold tables alone.
Proposed method
How the question could be tested
- 01Select a small, prespecified set of comparable release decisions from publicly documented safety frameworks. Include at least one reported safeguard or release change and at least one similar case where the decision evidence is incomplete. Preserve the policies and evaluator reports as they applied at the time, including later changes as separately dated records.
- 02Build an evidence table separating a stated trigger, test configuration, reported result, authority to act, stated rationale, observed implementation and any retest. Distinguish a developer account, independent corroboration, and an unknown outcome. Record competing explanations and changes to the benchmark or policy.
- 03Compare independent reviewers' conclusions using a framework or system card alone and using the full linked evidence. Score reproducibility of the classification and the questions that remain unresolved. Where the public evidence allows it, include a counterexample to an apparent score-to-action link.
Needed controls
What must constrain the study
- 01Predefine the sampling frame and include ambiguous and negative cases. Public disclosure is selective, so missing evidence cannot establish inaction.
- 02Distinguish attribution to an evaluation from verified implementation, and implementation from effectiveness. Timing or a developer's explanation alone cannot establish a causal effect or the evaluation's net social value.
- 03Record when publication of evaluation details may enable replication, gaming or capability transfer. Use public, non-sensitive evidence in the first study.
Relationship to existing work
FID-078 tests whether an evaluation predicts failure. FID-022 develops evidence requirements for faith-sector procurement, and FID-077 reconstructs incidents. This call studies the evidence linking safety evaluations to organizational decisions and documented follow-through across releases.
Expected outputs
Artifacts the work should produce
- 01A small sourced casebook and reusable evidence-to-decision record.
- 02A comparison of which conclusions independent reviewers can reproduce.
- 03Examples of remaining unknowns that a more informative disclosure or agreed access could resolve.
Open questions
Uncertainties the protocol must resolve
- 01Which parts of the decision record can an outside reviewer corroborate?
- 02How should a reader handle moving thresholds, multiple triggers, and exceptions?
- 03When can publication increase the decision value of an evaluation without revealing sensitive capabilities or test details?
Related calls
Continue through this research area
FID-069
Verifiable Delegation and Revocation in Multi-Agent Networks
How can people and institutions verify which human, organization, agent, or sub-agent is acting; what authority it received; what limits apply; and whether that authority has been narrowed or revoked across a multi-principal agent network?
FID-071
Confidential Agent Memory and Cross-Context Disclosure
How do persistent memory, summaries, retrieval stores, tool traces, delegation, and exports cause confidential context to influence or leak into unrelated sessions, roles, tasks, or organizations? Which technical controls make purpose limitation, deletion, and revocation testable?
FID-077
Independent Agent Incident Investigation and Evidence Sufficiency
What operational evidence lets independent investigators reconstruct an agent incident, distinguish competing explanations, and identify which interventions could have changed the outcome?
Open question
Open work
Primary need: technical governance, case study methods, evaluation policy, and independent evidence review
- Reconstruct the reported May 2025 Claude Opus 4 safeguard decision in 1–2 hours. Return a one-page record linking the applicable policy, stated rationale, reported action, implementation evidence, and one open question.
- Audit the citations or find independent evidence that would change the classification.
New to the Contributor Network? Share a short private introduction once, even if you join several calls. You can join the public discussion while introducing yourself.
Join the Contributor Network ↗