Labs

With real human data, mouse dynamics separate a person from a simple script almost perfectly and drop to a third of the hit rate against a smooth replay of the human movement itself; keyboard dynamics fall to chance. Passive observation cannot tell a competent copy from the original, so a passive signal serves to decide when to challenge, never as proof of presence.

Method: Three public sets of human data (free mouse, fixed-text keyboard and free-text keyboard) against scripts at four levels of sophistication, plus a smooth replay of real human traces, validated by a gate that checks whether it preserves the microstructure of the movement. Metric at the operating point: human hit rate at a false alarm rate of at most 1%, evaluated per session and across people, with the worst person as the indicator. No keystroke content captured; no pooling of sets into a single class.

Source: Hummand measurement bench (private repository). Human data: Balabit Mouse Dynamics Challenge (mouse), Killourhy and Maxion, CMU (keyboard, fixed text), and Dhakal et al., Aalto 136M (keyboard, free text). Measured on 2026-06-27.

Comparing face embeddings under homomorphic encryption is numerically free: the error sits seven to eleven orders of magnitude below what would change a biometric decision, reproducible a month later. The cost lies elsewhere: the available multiparty scheme requires every party to decrypt, the loss of one key is unrecoverable by cryptographic means, and performance in the reference environment was never measured. That is why the vault is research, and the product uses centralised custody.

Method: Validation gate with six criteria fixed before running, on CKKS in OpenFHE 1.5.1 with Python bindings, in a locked environment. Five measurements passed: bindings, multiparty protocol with three parties over one hundred runs, precision of distance and inner product over ten thousand pairs of dimension 512, and reproducibility after 28 days. The performance criterion was not measured under the planned protocol and the independent verification criterion was not done.

Source: Hummand validation gate and reproducible code (research repositories, private). OpenFHE v1.5.1, openfhe-python 1.5.1.0.24.4, Ubuntu 24.04, Python 3.12. Measurements of 2026-05-03, 2026-05-15 and 2026-05-31.

Six sentences for the supervision of AI systems: machines will act faster than any person can follow, and the design starts from there; delegation has a human origin, a declared limit and a way back; an alert is the moment human attention is called; proof over trust; stating the limit is part of the product. Nothing here was measured: it is a position.

Method: Position document, with no measurement. Derived from the TheAlert start document (19 September 2026), published before the technical proof by the founder's decision on 21 September 2026. This note is a declared exception to the criterion for notes: it records what TheAlert promises to sustain, not what was learned with a method.

Source: Manifesto at thealert.io. TheAlert start document and Hummand's dated decision record (internal). Agents line at labs.hummand.com.br/en/notas/mandate-presence-and-circuit-breaker-for-agents and labs.hummand.com.br/en/notas/a-gate-that-warns-and-a-gate-that-requires.

On a deterministic test bench with seven shapes of call, the gate that only warns let six through. The executor that requires a signed receipt let one through, the correct one. The advisory gate stopped exactly the caller that was going to stop anyway.

Method: Deterministic test bench, with no language model and no network. Seven caller behaviours, from correct to replay, receipt from another request, expired receipt and action under a revoked mandate, against two gates that run the same verification and differ only in what they do with the answer. Receipts signed with a test key, reproducible result, locked by an automated test.

Source: Hummand test bench (private repository). Design described at labs.hummand.com.br/en/notas/mandate-presence-and-circuit-breaker-for-agents.

A gate that advises an agent depends on the agent's cooperation. An executor that requires a signed receipt depends on nothing, and an agent's authority can be revoked by a human act, with no access to the machine it runs on.

Method: Design on top of the existing act primitive: scoped mandate, delegated act and signed humanproof/1 receipt, verifiable offline. A reference executor and adapters for four agent frameworks, with a deterministic test bench comparing an advisory gate against a requiring one.

Source: Hummand specification, implementation and test bench (private repositories). Receipt format at labs.hummand.com.br/en/notas/the-humanproof-1-format. Public key at gateway.hummand.com.br/.well-known/humanproof/keys.json.

A signed receipt that carries only hashes and that a third party verifies offline, with standard cryptography and per-client chaining, is enough for a regulator to check an act without trusting Hummand.

Method: Specification of the format on top of RFC 8785, ECDSA P-256 and SHA-256; a single implementation in Rust used by the server, by the command-line verifier and by this page; fixed test vectors checked on every change.

Source: Hummand specification and implementation (private repositories); verifier at labs.hummand.com.br/en/verificar; public key at gateway.hummand.com.br/.well-known/humanproof/keys.json.