Applied AI for the physical world.
Independent research into the AI-safety problems that matter: open, verifiable, and neglected by the labs racing past them.
Independent California 501(c)(3) · EIN 41-4991887 · research published open-source
Measuring multi-agent collusion, before it measures us.
As AI systems begin to negotiate, price, and coordinate on our behalf, they can learn to collude in ways no single-agent test would catch. ColludeBench is a pre-registered, timestamped benchmark that measures it directly.
ColludeBench
A pre-registered, RFC 3161-timestamped benchmark for multi-agent LLM collusion. Agents run repeated pricing games across controlled conditions; the harness measures whether a network of models compresses prices supra-competitively and whether communication amplifies the effect.
Pilot / Stage-2b results. Full protocol and data in the repository.
Research that anyone can re-run.
Beyond the flagship benchmark, HHA maintains a suite of open-source reinforcement-learning environments for high-stakes clinical and embodied decision-making. Every one is published, versioned, and installable.
AnestheSim
RL for anesthesia dosing.
VentiSim
RL for mechanical ventilation.
GlucoSim
RL for glucose management.
OncoSim
RL for radiation-therapy planning.
CardioSim
RL for cardiac electrophysiology.
NeuroSim
Brain-computer interface environments.
VascularSim
Microbot vascular navigation.
PeptideGym
Peptide-design environments.
Fundable research tracks.
Each program is a TRL-mapped track with defined milestones, structured so a funder can back a specific, verifiable outcome rather than a vague mandate.
Evaluation and adversarial testing
Reproducible evaluation frameworks with binary-testable criteria, red-team methodology, and multi-perspective stress-testing.
Multi-agent safety and governance
Failure-cascade propagation, unintended coordination, and constraint erosion, with policy translation for democratic oversight. ColludeBench is the anchor program here.
Embodied AI systems
AI-driven control across scale tiers, and the manufacturing engineering that bridges laboratory prototypes to deployable systems.
RL for medical decision-making
Open reinforcement-learning environments for clinical applications, distributed as installable packages.
Methodology infrastructure
Cross-domain pipelines from research question to re-derivable artifact, with reference designs per category.
Independent by design.
HHA is a small, independent research group. The people are the method, and the governance is deliberately built so the research answers to the evidence, not to a commercial incentive.
Independence and governance. HHA is an independent California 501(c)(3). Research is published open-source, and every primary result is designed to be re-derived by a third party in a different toolchain.
Where research produces a commercializable reference design, that design is licensed at arm's length to commercial partners, and any such license is subject to conflict-of-interest review. Commercial revenue never sets the research agenda.
Reference designs.
The second frontier: putting learned intelligence onto the smallest possible hardware. These are research artifacts, published methods, not products.
Cassette
A gait-coaching reference design engineered against an 8 KB on-sensor (ISPU) memory budget, pairing on-sensor classification with a phone-side coaching model. On-device fit is currently predicted from a datasheet-derived model; hardware bench validation is pending.
Reference designs are licensed at arm's length to commercial partners under conflict-of-interest review. Royalties flow back to support open research, but the commercial path only ever picks up designs the research has already published. HHA sells nothing here.
Every claim resolves to an artifact.
The through-line across both frontiers: nothing is asserted that a third party cannot re-derive. That discipline is the product.
Criteria-driven
Research questions decompose into discrete, binary-testable criteria before any investigation begins.
Pre-registered
Hypotheses, endpoints, and stopping rules are locked and RFC 3161-timestamped before data collection.
Cross-toolchain
An independent implementation in a different language reproduces every primary numeric result.
Open by default
Papers, code, protocols, and timestamps ship publicly. The record is the receipt.
Fund research the labs are racing past.
HHA is grant-funded and independent. The open, verifiable safety work here exists because someone chose to fund the neglected question instead of the crowded one. That can be you.
Also open to: research collaboration · reference-design licensing · joining the team