v4.4.0 + v4.4.1 — Temporal Integrity Layer

Visibility tells you what's happening right now. The temporal integrity layer tells you whether what's happening now is consistent with what was always supposed to happen.

These two releases ship the first and second layer of that stack.

What changed

v4.3.9 — Ambient Intelligence shipped the data foundation: trigger_type filtering across the Seat Fleet, autonomous seat detection with violet badge and row styling, and the Unowned Cost card on Mission Control — surfacing behavioral entities with no owner seat and attributing their cost directly.

v4.4.0 — Behavioral Baseline Fingerprinting is the first release that gives PromptKing a memory of what normal looks like. Every behavioral entity now accumulates a multi-dimensional fingerprint — tool usage patterns, trajectory shape, temporal rhythm, operating mode — over a meaningful observation window. A maturity_score (0–1) is honest about data sufficiency. No baseline claims stability until the data supports it.

v4.4.1 — Drift vs Intended Variation converts that memory into governance intelligence. The drift engine evaluates three required components before classifying anything as drift: meaningful deviation from baseline, measurable business impact, and duration beyond the noise window. A change with no business impact is not drift. A self-correcting deviation is not drift. Drift is change that matters.

What the operator sees

On /dashboard/behavioral-baseline:

• Maturity status in plain language — "Baseline established — drift detection active" or "Baseline forming — initial patterns visible, not yet reliable." Never raw enum values.
• Drift classification with full evidence — which deviation dimensions triggered, what the business impact is, how long the deviation has persisted, and the recommended action.
• CFO-readable impact — "Token spend increased 34% above baseline over 6.2 hours across 3 consecutive sessions. Persistence threshold exceeded. Estimated $420 unrecoverable cost if unchecked."
• Archetype review flag — when confirmed drift is detected on an entity linked to a seat, a banner recommends human review of that seat's archetype classification.

The multi-buyer moment

CFO: For the first time, AI cost deviation has a business impact number attached — not a token count, not a percentage. A dollar figure with a recommended action.

CIO / IT: Agent behavioral drift is now classified, not just detected. The difference between normal_variation, intended_variation, and confirmed_drift is the difference between a noise alert and an incident.

Legal / Compliance: Intended variation requires verified evidence — a matching policy simulation record or a human annotation. Without it, the system classifies insufficient_context, not intended_variation. The governance is honest about what it doesn't know.

Security: The cycle rate spike signal — the +187.5% loop frequency increase that surfaces in the demo — is the early warning for runaway agent behavior before it becomes a billing incident or a data exposure event.

What's next

Sprint 03 — Rollback Primitive — closes the temporal integrity loop. Baseline tells you what normal looks like. Drift tells you when it changed. Rollback gives you the recovery path — with an immutable audit trail that satisfies Article 14 of the EU AI Act.

That's the full governance story. See it running now.