An AI safety result for multi-agent systems. Why do some agent collectives improve with more agents, more debate or a stronger model, while others do not? The productive question is which resource limits a collective, and which intervention removes that limit. The paper decomposes collective intelligence into three ceilings: information (a team cannot know more than its starting data plus what its tools can carry — talking and voting add nothing to that budget, and adding agents raises it only through more tool access), reasoning (what bounded agents can use is at most what they hold; debate and compute close the gap but never pass it, at a cost from linear to infeasible), and communication (evidence split across agents is unusable until enough is exchanged — about n/2 bits for some n-bit tasks even with unlimited computation). The decomposition explains published findings that otherwise look contradictory, yields a diagnostic for which ceiling binds, and makes predictions a single capability axis does not. Proofs are short; every finite example is verified exactly in public code.
Independent research institute & security lab
Shrewd Research.
The metre, for AI safety.
We test the claims AI systems depend on, prove where defenses fail, and publish evidence others can verify.
The mission
An independent standard for AI safety claims.
A measurement only means something if it can be checked independently. We bring that principle to AI safety: research the limits, test real systems, and make the evidence available to others. The same method travels across the fields where AI is now deployed.
Shrewd Research is completely independent — no venture money, no corporate sponsor, no product to protect. When we say a claim holds or fails, the only interest behind that answer is the evidence.
- Healthcare
- Agentic systems
- Multi-agent teams
- Model evaluation
- Digital forensics
- Infrastructure
- Security education
What we do
Visit the labResearch
Formal results and adversarial measurement that find the limits of accepted defenses.
02Verify
Independent testing of AI infrastructure, evaluations, and security claims.
03Build & disclose
Open models and tools, plus reported vulnerabilities in production software.
Newest first
Latest research
Presents QFIRE, an inline, provider-agnostic prompt firewall for healthcare AI agents that combines positive-security scope constraints, an asynchronous detector graph, and de-obfuscation — and releases QFIRE-HealthBench, where generic injection detectors recover only 0.40 recall while QFIRE's scope+PHI chain reaches 0.83.
Proves that continuity, utility preservation, and complete safety cannot coexist for prompt-injection wrapper defenses, and characterizes the boundary where every such defense must fail.
Have a claim worth testing?
Tell us what needs to be verified.