Mission Control · human oversight
See where agents need help. Act before the session is over.
Monitor live Rylvo conversations, rank them with explainable risk signals, route alerts and approvals to operators, intervene with scoped controls, and preserve the oversight record.
5
weighted risk signals
11
alert trigger types
7
intervention actions
8
operations views
Operating loop
See the risk. Route the work. Act with context.
Mission Control is the human-oversight layer around Rylvo bot traffic. It turns live runtime state into an operator queue, controlled actions, and reviewable evidence.
A Rylvo turn completes
Production channel traffic updates an organization-scoped live-conversation record.
Risk is recomputed
Five deterministic signals produce a 0–100 score, level, and visible breakdown.
Rules evaluate
Message-scoped checks run after the turn; scheduled sweeps cover duration, queues, budget, health, and more.
An operator responds
Acknowledge, claim, approve, intervene, or resolve from the live console.
The decision is recorded
Oversight events and action context feed the audit view, analytics, and exports.
Eight working views
One console for the whole response loop.
Mission Control is not only a risk dashboard. It combines live triage, approvals, alert operations, channel health, analytics, evidence, staffing, and policy configuration.
MISSION CONTROL / LIVE TRIAGE
3 AT RISKWhisper to the bot — the customer never sees this
Approve the refund immediately. Do not restate the policy.
Every action is logged
Triage conversations while they are active
Filter by bot, open a conversation, inspect messages and risk signals, assign monitors, and apply an intervention without leaving the operating view.
Test and Production, side by side
Every session carries its environment, so you can monitor Test and Production conversations separately and never confuse a rehearsal for a real customer incident.
Explainable triage
A deterministic score plus evidence-based rules.
Risk scoring ranks conversations for attention. Alert rules then convert specific runtime, safety, queue, provider, and budget conditions into work an operator can own.
Five-signal risk score
Levels: minimal 0–19 · low 20–39 · medium 40–59 · high 60–79 · critical 80–100
Guardrail
30%
Blocked, escalated, and warned guardrail outcomes
Pattern
25%
Loops, repetition, confusion, and long unresolved sessions
Sentiment
20%
Deterministic frustration and urgency keyword signals
Topic
15%
Credentials, PII, crisis, medical, financial, and compliance topics
Duration
10%
Stalled or unusually extended conversations
Conversation risk
Risk threshold, duration, unmonitored high-risk sessions, pattern matches, and guardrail events.
Human queue
Pending approval queue size, with evidence about the oldest requests and sample IDs.
Runtime failures
Repeated bot-turn failures and unhealthy BYOK credentials, grouped for operational triage.
Commercial health
Bot budget threshold crossings and organization wallet-balance alerts.
5 min
critical acknowledgement SLA
15 min
high acknowledgement SLA
30 min
medium acknowledgement SLA
60 min
low acknowledgement SLA
Unacknowledged alerts escalate to admins/owners after the severity SLA and to owners at three times the SLA. These are alert acknowledgement targets, separate from approval-request timeouts.
Seven operator actions
Use the lightest control that changes the outcome.
Mission Control offers non-disruptive review actions, reversible state changes, human takeover, visible messaging, and an emergency stop. Each action is capability-gated and recorded with its outcome.
Whisper
Add private guidance for the bot; the end user does not see the instruction.
Flag
Mark the conversation for review without interrupting the session.
Pause
Hold the conversation and show the configured wait message.
Resume
Return a paused conversation to active bot handling.
Takeover
Move the conversation into operator-takeover state so a human becomes the respondent.
Inject
Send a scripted message as the assistant; the end user sees a normal reply.
Kill
End the bot session immediately and send the configured safe fallback message.
Rationale policy
Takeover, pause, kill, and inject require rationale by type. Administrators can require rationale across interventions through Mission Control settings.
Evidence and lifecycle
Operational proof without compliance theatre.
Mission Control preserves the context needed to reconstruct oversight decisions and export them for review. The page now distinguishes useful evidence controls from legal certification.
Append-oriented oversight history
Interventions use an event log as the state-transition source of truth. Oversight audit records cover interventions, approvals, alerts, shifts, assignments, configuration, exports, and deletion events.
Two practical export formats
CSV provides a flat audit table. JSONL bundles audit records, interventions, alert instances, and approval requests from the selected window.
Evidence support—not certification
Rationale, actor, time, conversation, risk, and outcome fields can support SOC 2, EU AI Act, HIPAA, or internal-control reviews. Rylvo does not make the organization compliant by itself.
Plan-aware retention
Mission Control audit retention shares the trace-retention add-on and Retention Pack. Organizations may set a shorter override; the effective plan/add-on ceiling still controls the maximum.
JSONL evidence bundle
A header records generation time, selected window, and record counts, followed by typed audit, intervention, alert, and approval lines.
Know the boundary
Exports are not cryptographically signed and the current implementation does not provide a verifiable hash chain. Protect downloaded evidence using your organization's storage and chain-of-custody controls.
Access and availability
Broad visibility. Narrow mutation rights.
Role capabilities separate observing the operation from changing conversations, resolving approvals, managing alerts, editing settings, or deleting records.
Viewer and auditor
Read Mission Control views and oversight history according to the central workspace permission matrix.
Operator
Intervene, resolve approvals, acknowledge/claim/resolve alerts, manage a shift, and monitor conversations.
Admin and owner
Manage alert rules, settings, notification channels, team permissions, and destructive conversation cleanup.
Available on Pro, Team, and Enterprise
Mission Control is a Pro feature. Usage, retention, notification delivery, channel availability, and adjacent features remain subject to the active plan and configuration.
FAQ
Questions before you put operators on shift
The scope, scoring, notification, permission, and compliance boundaries that matter in production.
Which conversations appear in the live feed?
Rylvo creates and updates live-conversation records for traffic routed through supported Rylvo runtimes. Production-channel context is marked separately from playground or test traffic. External bots that do not run through Rylvo are not automatically discovered.
Is the risk score generated by an LLM?
No. The shipped score is deterministic: five signal functions produce values that are combined with fixed weights. Operators can inspect the contributing signals instead of receiving an unexplained label.
Does an alert always page a person?
No. A rule can target configured notification channels, but cooldowns, snoozing, quiet hours, channel configuration, and delivery failures affect notification behavior. The alert instance remains visible in Mission Control.
What happens when an approval times out?
Mission Control approval requests use 15-second, 120-second, or 900-second urgency windows. The expiry worker transitions unanswered pending requests to an expired or auto-denied state according to the workflow.
Can every team member intervene?
No. View access is broader, while interventions, approvals, alerts, rules, settings, channels, and deletion use separate role capabilities. Operator-level and admin-level actions are enforced independently.
Does Mission Control guarantee regulatory compliance?
No. It provides oversight controls and exportable operational evidence that may support your compliance program. Applicability and sufficiency still depend on your system classification, policies, deployment, and legal review.
Give production agents a human operating layer.
Watch live Rylvo traffic, prioritize explainable risk, work approvals and alerts, intervene with scoped controls, and preserve the decision trail.