RESPOND · COUNTER-SURVEILLANCE
Spot the watcher watching you.
Hostile reconnaissance has a shape: a subject who doublebacks on themselves to check for a tail, or loiters at a fixed point to stake out a location. Sentinel detects both with deterministic rules — not a black-box score — and shows you exactly which events fired each one. A marquee capability for critical-infrastructure and public-safety control rooms.
Deterministic detection over identified subjects — every signal traceable to the events behind it.
Two behaviours. One honest definition each.
Counter-surveillance does not guess. It runs two deterministic rules over a subject’s recent sightings — each defined in plain terms an operator can repeat back.
- Doubleback — a subject seen at camera A, then B, then A again within the window; each leg must take long enough to be a real transition, not two cameras catching the same moment
- Loitering — the same subject seen repeatedly at one camera inside the window; more visits escalate the severity
- Identity-anchored — the rules reason over a named person record, never anonymous motion
- No black-box score — there is no model confidence to argue with; the rule either fired or it did not
- A third rule (passing here more often than usual for this subject) is deliberately held back until per-subject baselines can make it honest
Click to enlargeA doubleback flagged on a single subject — with the cameras and the rule that fired it stated plainly.
Every detection cites the events that fired it.
The difference between a usable signal and a liability is whether you can say why. Each counter-surveillance detection carries the exact event ids behind it, so an operator — or an auditor — can follow the chain to the underlying frames.
- Each detection links to the precise events that triggered the rule — not a summary, the actual sightings
- Trace straight to the subject on the Command Center map and watch the flagged route play out
- Plate-only cameras are excluded from the face-pattern math; sub-confidence matches never count
- Re-firing escalates severity and refreshes last-seen — it never duplicates the card or quietly downgrades
- Nothing autonomous — no auto-watchlist, no alert cascade, no camera takeover; the operator decides
Trace a flagged subject across the network — the doubleback rendered as a route.
The lifecycle
See it, decide it, record it.
A counter-surveillance detection is not an alarm that demands action — it is a signal that asks for a decision. Every detection offers the same four moves, and every move is written to the audit with who did it and when.
Acknowledge
Mark a detection as seen without closing it. The card stays in the active inbox so the next operator on shift inherits the same picture — an explicit "I have eyes on this," recorded with who and when.
Dismiss
Close a detection that turned out to be benign, with an optional reason that is written to the audit. Dismissing is a deliberate act — and a dismissed subject does not immediately resurface on the same window of events.
Trace
Jump straight from the detection to the subject on the Command Center map and follow their route across the camera network — turning a flagged pattern into a live picture in one click.
Mute
Suppress further detections on a subject who is known-OK — a receptionist at their post, a guard on a patrol loop. One click from the card; reason prompted; reversible. The audit keeps the record.
Watch the right people, in the right places.
Run unscoped and the rules are noisy on a normal population — a worker walking to the break room doublebacks, a receptionist loiters at their post. So Sentinel ships four orthogonal levers to point the detector exactly where it earns its keep.
- Org-wide mode — watch everyone and mute the known-OK, or watch only explicitly subscribed subjects
- Per-person subscribe / mute — enrol or silence a single subject, with reason, expiry, and pause/resume
- Watchlist flag — guarantee a person of interest is always monitored, regardless of mode
- Sensitive-zone cameras — narrow detection to the cameras covering restricted areas only
- The levers compose — e.g. watch everyone, but only in your secure zones, with a few staff muted
Click to enlargeFlag a watchlist subject as monitored and the detector watches them regardless of org mode.
Four ways to scope
Orthogonal levers that compose.
Org-wide mode
Choose the default for everyone: watch every identified subject and mute the known-OK ones, or watch only subjects an operator has explicitly enrolled. One toggle sets the posture for the whole deployment.
Per-person subscribe / mute
Enrol a single subject for monitoring, or mute one who generates noise. Each entry carries an author, an optional reason, and an optional expiry — and can be paused without losing its history.
Watchlist flag
Flip a switch on a watchlist entry and that subject is always monitored, regardless of org mode. The system errs toward watching a flagged person of interest — the flag wins over a mute if both are set.
Sensitive-zone cameras
Mark the cameras that cover restricted areas — server rooms, vaults, executive floors — and detection narrows to just those. Answer "show me counter-surveillance behaviour in my secure zones only," for everyone.
Continuity across shifts, on the subject.
Open any subject and the counter-surveillance card shows where they stand — subscribed, muted, or neither — alongside a 90-day detection history, so an operator inherits context instead of starting cold.
- A context strip names the org default, so an operator sees what subscribe or mute will actually do for this subject
- Three states made plain — subscribe to monitoring, mute from monitoring, or do neither
- Pause preserves the audit trail; remove deletes the entry — the distinction is explicit
- A 90-day detection history with the same severity and status pills as the main inbox
- Read-only operators still see the state and history; the action buttons are simply hidden
Click to enlargeThe org-default context strip means an operator always knows what a click will do before they make it.
How it works
From a face to a flagged pattern.
Identity anchors the movement
The AI vision pipeline turns faces into numeric signatures and rolls matching appearances into one person record. Counter-surveillance reasons over a subject’s recent sightings — never anonymous motion — so a detection always names who.
Two rules run on a cadence
On a fixed cron, deterministic rules scan recent sightings: a doubleback (camera A → B → A within the window) and loitering (the same subject seen repeatedly at one camera). Plate-only cameras and sub-confidence matches are excluded before the math runs.
A detection lights up — quietly
A match opens a card in the counter-surveillance inbox with its severity, the subject, the cameras involved, and the exact events that fired it. Nothing pages, nothing auto-watchlists, nothing takes camera control. The signal surfaces; the operator decides.
The operator works it, and it is recorded
Acknowledge, dismiss with a reason, trace on the map, or mute the subject — each action audited, each detection kept on the subject’s 90-day history so the next shift has continuity rather than a cold start.
The honest picture
Useful where it is pointed — and we say so.
The two rules are mathematically exact, but exactness is not the same as relevance. Run them across a whole building and they will fire on innocuous behaviour — the office worker on a coffee run, the guard on a patrol loop, the queue at a security gate. That is not a bug in the math; it is the nature of a rule that does not yet know this subject’s normal.
So we are direct about it: counter-surveillance is a precision instrument for narrow, high-suspicion subjects — a watchlisted person of interest doubling back on themselves, a stranger loitering at a vault door — not a dragnet for the general population. The four scoping levers exist precisely to point it there, and a future per-subject baseline rule will calibrate against each person’s own pattern. We would rather ship an honest tool you can trust than a confident one you cannot.
Under the hood
Specifications
| Detection model | Deterministic rules over event history — no ML, no data-inferred thresholds; every detection is explainable |
| Doubleback rule | Camera A → B → A within a 30-minute window; each leg must take ≥ 30 s (sub-30 s = two cameras seeing the same moment, not a real transition) |
| Loitering rule | Same subject, same camera ≥ 5 visits within 30 minutes → medium; ≥ 10 → escalates to high |
| Explainability | Each detection stores the exact triggering event ids — an auditor can follow the chain back to the underlying frames |
| Severity | Low · Medium · High; re-firing escalates severity and refreshes last-seen — never duplicates the card, never downgrades |
| Idempotency | At most one open detection per (subject, rule); a dismissed detection does not resurface on the same overlapping window |
| Lifecycle | Acknowledge · Dismiss (with reason) · Trace on map · Mute — all audited with operator and timestamp |
| Scoping — org mode | Watch everyone (mute the known-OK) or watch subscribed subjects only; mode change takes effect on the next cron tick |
| Scoping — per person | Subscribe / mute with author, optional reason (≤ 500 chars), optional expiry (≤ 365 days), pause/resume without deletion |
| Scoping — watchlist | A per-entry flag forces always-monitored for that subject, independent of org mode; the flag wins over a mute |
| Scoping — sensitive zones | Flag cameras covering restricted areas; once any camera is flagged, detection restricts to those cameras org-wide |
| Per-person history | 90-day detection history on the subject’s detail page, with severity and status, for cross-shift continuity |
| Honest filtering | Plate-only cameras excluded from face-pattern math; sub-confidence matches excluded — carried from the platform’s confidence contract |
Specifications describe shipped platform capabilities; we'll confirm the configuration that fits your deployment during your demo.
Keep working the case
Counter-Surveillance is one tool in the Respond set.
Cross-Camera Tracking
Follow a flagged subject’s route across your camera network on a map, color-locked so they are unmistakable on every screen — the natural next step after a counter-surveillance detection fires.
Learn morePredictive & Behavior
Get alerted when a subject with an established rhythm breaks their normal pattern — the deviation analytics that run alongside counter-surveillance.
Learn morePersons & Pattern-of-Life
The identity-level record that anchors every detection — a subject’s 7×24 activity map, regularity score, and where they are anchored, all on one page.
Learn moreSee a doubleback flagged on live data.
Request demo access and we’ll send you a private, pre-loaded environment. Open the counter-surveillance inbox, read a detection, follow the cited events to the frames behind them, trace the subject across the map, and set the four scoping levers for your own deployment.