indentia.ai

Human oversight & HITL

Autonomy at the edge. Authority with the human.

AI agents that act in regulated environments cannot operate as black boxes. Indentia builds explicit gates into every agent workflow — a person decides where the agent looks, what it produces, and when it goes out the door. Below a confidence threshold, the agent simply refuses to guess.

Walk through a HITL scenario Back to the platform

Four explicit gates

Every agent run passes through human review four times.

Gate 1 — Scope & intent

Before the agent starts work, a human confirms the goal, the scope, the data sources in play and the time budget. The agent cannot self-extend its scope.

Gate 2 — Retrieval review

After retrieval, a reviewer can see which sources were consulted, with what permissions, and can prune or add. The agent does not proceed until retrieval is approved.

Gate 3 — Draft review

The agent's draft output is shown alongside its citations and confidence scores. A reviewer accepts, edits or rejects. Edits are tracked and feed back into the agent's memory.

Gate 4 — Release

Final release happens under explicit human authorisation, with releasability checks against need-to-know labels and audit-log entry signed by the releasing identity.

Safety & release controls

Refusal, releasability, and a paper trail that holds up.

AbstentionGate & confidence thresholds

Agents must signal "I don't know" when retrieval evidence is thin or contested. Below a configurable confidence threshold, the answer is withheld and the question is routed to a human — not bluffed.

Releasability & need-to-know

Every output is checked against the consumer's clearance and the artefact's releasability label before it is shown. Need-to-know enforcement happens at the platform, not at the reader's discretion.

STANAG 4559 metadata

Intelligence products are exported with STANAG 4559 metadata so they integrate with existing collection-management and dissemination systems. Lineage and classification travel with the product.

NATO Cursor-on-Target (CoT)

Tactical-display outputs are emitted in CoT XML, ready for downstream situational-awareness platforms. The same agent reasoning, in a format the tactical edge already understands.

Decision log & audit trail

Every gate transition, every approval, every override is signed and immutably logged. Replay the decision chain that led to any released output — months later, without losing context.

Citations on every claim

No output without a source. Every paragraph, every number, every recommendation links back to the records that support it. "The AI said so" is never a final answer.

Auditability is the product

Replay any decision. Months later.

The decision log is the heart of trust. Every gate transition, every approval, every model and prompt version, every retrieval source — signed at the moment, immutable afterwards. When the auditor, the regulator or the inspector-general comes back six months from now, the answer to "why did this output go out?" is one query, not a forensic project.

Show me the audit trail