The HYDRA family: epistemic correctness by construction

A fraud-detection system acting on a stale transaction pattern, a sensing system acting on outdated intelligence, and a clinical decision-support tool acting on a recalled guideline all fail for the same structural reason: their architecture has no mechanism to check that a belief is still valid before it drives an action. The failure is not in the reasoning engine — it is that the engine reasons over invalid inputs.

This matters most exactly where oversight cannot scale. When a system processes beliefs faster than a human can review them, the expected number of invalid beliefs reaching the reasoner grows with volume, and manual oversight degrades past a crossover point. HYDRA’s claim is that the fix must be structural, so it scales with volume instead of degrading with it.

The key move is to separate perception from belief commitment. In a naive agent, a sensor reading (including a language-model output) is written straight into the belief base. In HYDRA, that reading is only a proposal: it must pass an epistemic-validation gate before any belief is committed — the same propose-only seam ROOTFRAME enforces, applied to every perception boundary.

Freshness check

Is the belief temporally current, or has it gone stale relative to its source?

Contradiction check

Does it conflict with what is already held or with other incoming evidence?

Source triangulation

Is it sufficiently corroborated by independent sources, rather than spuriously echoed by one?

The same five-layer pattern — typed sensor, epistemic gate, a BDI/HTN cognitive engine, a bounded personality/behaviour profile, and a coalition coordinator — is instantiated across domains that share no domain knowledge:

• ▸Anti-money-laundering and fraud detection — the most mature instance; the write-up reports it as validated in production.
• ▸Adversarial red-team simulation for AML — models an evasive adversary, with a formal treatment of intent ("absent mens rea") used to reason about detection coverage.
• ▸Autonomous UAV swarm operations — a two-tier cognitive/reactive architecture that keeps human authority over irreversible actions, with mission orders validated before execution.
• ▸Clinical belief management — epistemic validation applied to recalled guidelines and contradicting evidence in decision support.

The family’s central theoretical result is a cross-instance universality claim: the pattern’s epistemic-safety properties are preserved under instantiation to any domain that satisfies the five-layer specification. The result is a property of the pattern under its stated model — not a guarantee that any particular deployment is safe in the world.