INVESTIGATE · PREDICTIVE & BEHAVIOR ANALYTICS
Don’t just see them appear. Know when they break pattern.
Most systems tell you a subject showed up. Sentinel tells you when a subject who follows a regular rhythm stops following it — and when the history is too thin to make that claim honestly, it says so.
A subject’s regular rhythm, learned over 90 days — the deviation is the moment this breaks.
Get alerted when a subject breaks their normal pattern.
Sentinel watches subjects with an established rhythm for the break in that rhythm — raising a deviation alert when the gap since their last expected appearance runs well past their normal cadence.
- Deviation, not presence — the alert fires on the absence of an expected appearance, the inverse of a watchlist hit
- Two severities — a caution when the observed gap runs 2.5× the expected cadence; an escalation at 5×
- Auto-resolves — when the subject is sighted again, Sentinel closes the alert on its own; no stale alarms
- Confidence-aware — plate-only cameras and low-confidence matches are excluded from the pattern
- One alert per subject — at most one open deviation alert per person at a time, by design — never a flood
Click to enlargeCaution at 2.5×. Escalation at 5×. Auto-resolves the moment they are sighted again.
It won’t cry wolf on a subject it doesn’t understand.
Deviation alerts only fire for subjects with a genuinely established pattern — below the bar, the subject is silently skipped rather than generating a claim the math can’t support.
- Eligibility gate — alerts fire naturally only for subjects above a regularity and sighting-count threshold
- Silent skip, not a fabricated alert — below the bar, no off-pattern claim is made
- Eligibility shown in plain terms — the person's profile shows the score and count against the threshold, pass or fail
- Gate-aware empty states — an empty deviation list explains why there are no alerts, never just shows nothing
- The honesty contract — the same confidence-aware filtering that governs the rest of Sentinel governs this
Click to enlargePass or fail, shown plainly. The analyst always knows whether the system is watching or waiting.
Prediction accuracy
The AI that grades itself.
A prediction only earns trust if it’s measured. Sentinel keeps score on its own next-camera predictions — logging every one it fires, then checking whether the subject actually turned up where it said. The result is a self-measured, per-camera-pair hit-rate you can put in front of oversight: evidence the capability works on your estate, not a vendor’s word that it does.
- Self-scoring — every prediction it fires is logged with its outcome: it landed, it missed, or it timed out. Nothing is quietly dropped to flatter the number.
- Colour-graded, honestly — strong pairs read green, marginal ones amber, weak ones red, and a pair with no data shows “—” rather than a misleading zero.
- An onboarding answer — “your network is this accurate after this many weeks,” straight from the data as the model learns your estate.
- Finds its own weak spots — sort by lowest hit-rate and a moved or mis-mapped camera surfaces itself, before it costs an operator a trace.
- Built for oversight — it answers the reviewer’s hardest question (“does this actually work?”) with a figure the system measured on itself.
Click to enlargeThe full behavior toolkit
Deviation, subscriptions, history, and more.
Honest eligibility gates
A deviation alert only fires for a subject whose pattern is genuinely established — enough sightings, regular enough to be predictable. Below that bar, the subject is silently skipped. No "off-pattern" claim, because the math can't credibly make one. The person's profile shows the eligibility test in plain terms — score and count against the thresholds, pass or fail.
Deviation subscriptions
For a VIP, a person of interest, or a witness with sparse history, an operator can explicitly subscribe — watch them on relaxed gates with an optional justification and an expiry. Subscriptions can be set, paused, resumed, edited, removed, and they expire on their own. Every state change is audited with the operator and reason.
Behavior on every profile
Deviation is not a separate console to remember to open. It is a section at the foot of every person profile: eligibility, subscription controls, and a collapsible 90-day history of every deviation — each with its severity, status, source, observed gap, and how it resolved. Persons-list chips surface the off-pattern subject at a glance without opening anyone.
Complete deviation history
A dedicated deviations page gives the whole estate's off-pattern history in one place — every alert raised, filterable by status, severity, source, and time window, searchable by subject name. Stat cards summarize what's open, acknowledged, resolved, and dismissed. Act from the list — Acknowledge, Dismiss, or Trace — every action audited.
Permission-gated
Viewing deviation history is one permission; acknowledging, dismissing, and subscribing require another. Locked-out operators still see history but cannot act. A subscribed subject is clearly distinguished from one Sentinel flagged on its own — the record always shows whether Sentinel found this or an operator asked Sentinel to watch.
Predictive next-camera — Emerging
There's a natural next question after "they broke pattern": where are they likely to head? The foundations are in pilot — a transition model that suggests the next camera or two with a confidence figure and an ETA, and an honest "insufficient pattern history" signal when the data is too thin. Not claimed as a dependable capability until pilot-validated; ask us where it stands in your demo.
Click to enlargeHow it works
From a rhythm to a deviation, accountably.
Learn the rhythm
From the same confidence-filtered sightings that power every person profile, Sentinel computes a regularity score over a 7×24 activity map. Only subjects with a genuinely established pattern qualify for natural alerts.
Watch the gap
A background process checks expected-versus-observed cadence on a short cycle. When a subject's gap runs well past their normal, a deviation alert is raised — a caution at 2.5× expected, an escalation at 5×.
Surface it
The alert appears on the Command Center banner, on the subject's profile, as a chip in the persons list, and in the deviation history — never more than one open alert per subject.
Decide and resolve
An operator acknowledges or dismisses, with the action audited. When the subject is sighted again, Sentinel auto-resolves the alert. No autonomous action, ever.
Under the hood
Specifications
| Detection basis | Regularity score (Shannon-entropy) over a 7×24 sightings matrix — the same math as the Pattern-of-Life profile |
| Natural eligibility gate | Subject must clear both a regularity threshold and a 90-day minimum sighting count; below either, silently skipped (no fabricated alert) |
| Severity thresholds | Caution when observed gap ≥ 2.5× expected cadence; escalation at ≥ 5× |
| Auto-resolve | Alert closes automatically once the subject is sighted again |
| De-duplication | At most one open alert per subject at a time, idempotent by construction |
| Confidence-aware input | Plate-only cameras and sub-threshold matches excluded from the pattern math |
| Subscriptions | Explicit per-subject watch on relaxed gates; optional reason and expiry (24h / 7d / 30d / 90d / indefinite); set · pause · resume · edit · remove · auto-expire |
| Source tagging | Every alert tagged natural or subscription, so the record distinguishes the two |
| Discoverability | Command Center banner · person behavior section · persons-list chips (deviation / watching / paused) · dedicated deviation history page |
| History | Per-subject 90-day history on the profile; estate-wide history page with status / severity / source / window filters, name search, pagination |
| Lifecycle audit | Subscribe / pause / resume / edit / delete / expire and acknowledge / dismiss all written to the tamper-evident audit with operator + reason |
| Permissions | View vs. act split — viewing history is one permission; acknowledging, dismissing, and subscribing require another |
| Autonomy | None — every alert is resolved by an operator; no automatic enrolment, camera control, or notification cascade |
| Predictive next-camera | Emerging / in pilot — transition-model next-camera suggestion with confidence + ETA and an honest "insufficient history" signal; not claimed as a dependable capability until pilot-validated |
| Self-measured accuracy | Sentinel logs every next-camera prediction and grades its own hit-rate per camera pair (origin → destination, by hour) over rolling 7- and 30-day windows — hit / miss / timed-out all counted, colour-graded, zero-data pairs shown as "—", open to oversight |
| Deployment | Cloud-managed, on-premise, or fully air-gapped — one codebase |
Specifications describe shipped platform capabilities except where explicitly labeled emerging or in pilot; we'll confirm exactly what's proven versus in trial during your demo.
The rest of Investigate
Deviation is one read on a subject. There’s more.
Persons & Pattern-of-Life
The 7×24 activity map, regularity score, co-occurrence network, and anchored-vs-mobile distribution that the deviation engine reads from. The differentiator this builds on.
Learn moreAlerts & Threat Detection
Watchlist matching in seconds, an auto-computed threat level, and counter-surveillance pattern detection for the live room.
Learn moreBriefing Mode
Every operator's shift starts with an auto-generated brief: the headline, what changed overnight, and the top actions for the day — the natural home for an overnight deviation.
Learn moreSee the off-pattern alert fire on live data.
Request demo access and we’ll send you a private, pre-loaded environment — open a subject with an established rhythm, watch a deviation alert surface on the Command Center, subscribe to a person of interest, and walk the full deviation history. We’ll show you exactly what’s shipped and what’s still in pilot, with nothing overstated.