v4.4.0–v4.4.2 — PromptKing Now Governs the Difference Between Tagged and Ambient AI

Enterprise AI now runs in two fundamentally different behavioral modes. PromptKing v4.4.0 through v4.4.2 treats them differently — because they are.

What shipped

v4.3.9 — Ambient Intelligence introduced trigger_type filtering across the Seat Fleet. For the first time, operators can filter seats by how they were triggered: human-initiated, scheduled, autonomous, or unattributed. Autonomous seats now surface with a violet left-border and an AUTONOMOUS badge. The Unowned Cost card on Mission Control shows the dollar value of behavioral entities with no owner seat — the cost that belongs to nobody on paper.

v4.4.0 — Behavioral Baseline Fingerprinting extended this foundation into the governance layer. Every behavioral entity now accumulates a multi-dimensional baseline fingerprint that explicitly captures operating mode as a dimension. A reactive tagged agent and an ambient proactive agent form baselines at different rates, with different sample requirements, and with different maturity thresholds — because their behavioral variance profiles are fundamentally different.

v4.4.1 — Drift vs Intended Variation applied mode-specific governance to the drift classification engine. The thresholds are different by design:

Reactive tagged mode: noise window 30 minutes, confirmed drift at 4 hours. Ambient proactive mode: noise window 60 minutes, confirmed drift at 12 hours.

A deviation that would be confirmed drift for a reactive agent is still noise for an ambient one. This is not a looser standard for ambient agents — it is a more honest standard, reflecting the reality that ambient agents have wider natural behavioral variance between sessions.

When confirmed drift is detected on an ambient proactive agent, the evidence requirements for intended variation are stricter: a matching policy simulation record or explicit human annotation is required. The system will not classify ambiguous ambient behavior as intended variation. Without verified evidence, it classifies as insufficient_context.

v4.4.2 — Rollback Primitive completed the loop with mode-aware recovery. When an ambient proactive agent drifts and an operator initiates a rollback, the system automatically flags all unverified external actions taken during the drift window as manual_required. An ambient agent with tool access may have written files, sent messages, or modified external state. The governance record reflects that honestly — it does not pretend those actions can be cleanly undone.

What the operator sees

The /dashboard/behavioral-baseline surface now shows both operating modes side by side:

An ambient proactive agent — be000001 in the demo org — with a mature baseline at 87% maturity, 54 samples, confirmed drift detected, $420 unrecoverable cost estimated, archetype review flagged.

A reactive tagged agent — be000002 — forming at 31% maturity, 8 samples, baseline building.

The governance story for each is different. The evidence requirements are different. The rollback implications are different. PromptKing surfaces those differences explicitly rather than collapsing them into a single risk score.

The multi-buyer moment

IT and Security: The trigger_type filter in Seat Fleet answers the question your team has been asking since you deployed ambient agents: which seats are running without a human in the loop, and how much are they costing?

Compliance and Legal: The mode-specific drift thresholds and evidence requirements for intended variation mean the governance record is defensible. When an auditor asks why a behavioral change was classified as acceptable, the answer is documented: policy simulation match within the observation window, or human annotation, or insufficient_context — never a judgment call without evidence.

CFO: The Unowned Cost card and the confirmed_drift business impact attribution now separate cost that was authorized and governed from cost that accumulated during drift. That is the accountability story finance has been waiting for.

The EU AI Act connection

Article 14 requires human oversight capability for high-risk AI systems. Ambient proactive agents — running for hours or days with tool access and minimal human check-ins — are exactly the systems Article 14 was written for.

The rollback primitive's audit trail, the mode-specific evidence requirements for intended variation, and the manual_required compensating action logging for ambient agents are all designed to satisfy Article 14 documentation requirements. The governance record exists. It is honest. It is exportable.

That is what ships today.