Safety Exchange — Specification 001
The working contract for coordinating bounded safety signals without sharing raw evidence or transferring enforcement authority.
Turnkeeper Labs · applied research
Applied research on measurement and evidence quality.
Turnkeeper Labs studies how machine learning can help people interpret risk and preserve provenance — without turning model output into automatic enforcement.
Synthetic-first · Held-out closed unless stated · Reviewable evidence
Research areasSynthetic
Signal interpretation
Help reviewers understand bounded safety signals in context.
Model calibration
Evaluate where models help, fail, or require specialist judgment.
Human review systems
Give reviewers clearer evidence without automating accusations.
Fig. 01 — bounded · synthetic · reviewable
Research method
How we turn uncertainty into reviewable evidence.
Research begins with the conditions people actually face—not with a model looking for a use.
Frame a real review problem
Work with practitioners to identify the decision, evidence, and human responsibility involved.
Test with bounded evidence
Prototype with synthetic or privacy-minimized data that can be inspected and stress-tested.
Publish limitations before claims
Document failure modes, boundaries, and unresolved questions before describing progress.
Open questions
Questions we're testing now.
Where AI can strengthen human review—and where it must not take over.
010203040506
Paper 002 · research planTwo Signals, One SourceRead Paper 002 →One Bad LinkWhat changes downstream when one identity connection is wrong?
A synthetic benchmark study of contamination, decision changes, transferred risk, and recovery after one false identity link.
Follow the researchTwo Signals, One SourceWhen does apparent corroboration become double-counted evidence?
Source dependence, provenance displays, and confidence inflation under synthetic lineage controls.
Open the studyWardBenchCan reconstruction tell independent sources from derived multiplicity?
Frozen adversarial cases scoring naive multiplicity against lineage-aware reconstruction.
Open WardBenchModel calibrationWhen should reviewers trust—or challenge—a model signal?
Synthetic calibration runs test where models help, fail, or require specialist judgment.
View the roadmapEvidence reconstructionHow can evidence preserve context without expanding collection?
Privacy-minimized patterns and counterevidence help specialists review what a bounded signal means.
Explore Ward evidence systemsExecution authorizationHow can recommendations remain separate from enforcement?
Synthetic demonstrations bind a recommendation to an exact, reviewable effect without granting it authority.
Open the demonstrationNo experiment authorizes enforcement on its own.
Evidence from the work
Progress should be inspectable.
Evaluation methods
Synthetic calibration plans, research gates, and the questions still being tested.
Working demonstrations
How a recommendation becomes a bounded request that still requires explicit authorization.
Open research components
The public building blocks that make provenance and review easier to verify.
Boundaries before breakthroughs
People remain responsible for consequential decisions.
Evidence before claims
Experiments remain clearly labeled until code, tests, and operating evidence support them.
Privacy minimized by design
Research starts with bounded data and does not expand collection by default.
People decide
Models may surface evidence or recommendations. Consequential judgment stays human.
Is
Bounded, synthetic, reviewable research designed to strengthen human judgment.
Is not
A surveillance system, live accusation engine, or source of autonomous enforcement.
Bring us a safety-review problem worth solving.
Work with Turnkeeper Labs on research that can be bounded, inspected, and improved with the people responsible for protecting children.
Turnkeeper Labs is an initiative of Turnkeeper.