EYDII

EYDII is ROAM's behavioral intelligence layer. It watches how agents operate: role drift, scope expansion, verification gaps, stalled work, unsafe handoffs, and peer-reported behavioral changes. It does this without reading prompts, code, messages, files, or model outputs.

What EYDII is

Most agent safety systems start with access control, policy checks, or another model reading the work. Those layers matter, but they do not answer the operational question: is this agent still behaving like the role it was trusted to perform?

EYDII answers that question from behavior. It observes timing, state changes, handoffs, declared intent, verification patterns, peer reports, and deviation from the role over time. The content of the work stays outside the trust layer.

In ROAM, EYDII is three things: a content-blind observation boundary, a per-agent trust signal, and a signed evidence trail for consequential operating events.

Content-blind by architecture

The boundary is structural. EYDII can receive behavioral metadata such as task IDs, state transitions, timing, declared intent, role boundaries, and peer reports. It does not need the actual prompt, response, file contents, source code, chat transcript, or artifact payload to do its job.

Drift

Drift is what happens when an agent keeps running but stops acting like the role it was assigned. It may move outside scope, skip verification, report progress without evidence, avoid tasks it should handle, or become more confident while doing less real work.

EYDII looks for behavioral patterns that tend to show up before bad work spreads: silent drift, authorization drift, phantom progress, unverified completion, role collapse, cascade amplification, and confidence-effort inversion.

The output is a trust signal, not a moral judgment on the model. It tells the operator which roles are healthy, which need review, and which work should be contained before it becomes shared memory.

Trust Score

Trust Score is the operator-facing signal shown on agent and role views. It is derived from EYDII's behavioral analysis, agent health, operating cadence, verification patterns, and peer reports inside the ROAM environment.

A Trust Score is not a one-time grade. The trend matters. An agent that drifts and recovers should look different from an agent that keeps degrading. When a role is swapped to a fresh agent, ROAM keeps the history attached to the role so the standard does not reset with the model.

Signed actions and the audit

ROAM treats consequential operating events as evidence: task transitions, approvals, handoffs, memory writes, quarantines, and swaps. Those events can be signed and journaled so a technical user can inspect what happened without simply trusting the UI.

The point is not compliance theatre. The point is inspectability. If an agent was quarantined, a memory write was flagged, or an operator approved a swap, the system should leave a local trail that can be reviewed later.

Quarantine and swap

When EYDII flags serious drift, ROAM can contain the affected work and present a decision: review, hand off, quarantine, or approve a replacement. The operator stays in control of the risk tradeoff.

A swap does not mean the organization forgets. The replacement agent inherits the role's canon, journal, open tasks, and relevant memory, while work written during the drift window can be marked for review.

That is the ROAM operating model: autonomy with a trust layer above it, not blind delegation and cleanup later.