Re-scan what you saved. Surface what changed.

What is Watchtower in Nerxus One?

Watchtower is the scheduled re-scan layer on opportunities you already saved, built to surface diffs instead of re-telling last month's diagnostic. The buyer is the marketer. The shop on the row is your customer. Watchtower does not introduce itself.

Radar owns what to watch and how often. A watch record holds detector families or specific signals, a cadence, alert routing, and a spend cap. Saved Opportunities is the library. The watch view is that library filtered to monitored rows plus a feed of deltas.

The job is unattended, so the implementation is paranoid on purpose. A sweep key authorizes starting a sweep and nothing else. Each tenant scan still uses that tenant's credentials. If the database is missing, the worker throws rather than report SUCCESS with no snapshot. A silent no-op is the one outcome the function may never produce.

M08 is the scheduler on paper. Inngest functions exist: a sweep, then one scan per business. Interactive Radar scans get the wider fleet. Watchtower concurrency is tighter, eight at a time, one per business id, because nobody is staring at a progress bar. Yield to the human.

Saved views matter because the same library holds New, Contacted, In pipeline, and Won. Won and still monitored is how a retainer stays a product instead of a hope. New drift this week is how Monday starts without a notes app. Filters combine: score band, tag, category, watched-only, has-drift.

Watchtower does not introduce itself to the shop. Radar owns what to watch. Relay is the only agent that sends outreach, and only after ComplianceGate. A delta that mails the owner by itself is a different product. Supervised is the default.

How a scheduled re-scan diffs what you already saved

A scheduled re-scan diffs what you already saved by taking a new snapshot at the watch depth and comparing it to the baseline stored from the last successful run. Reasons the job can record are baseline, not-comparable, no-change, changed, and failed. Only changed writes a delta you can sell.

The event payload is a job id and a business id. Depth, target, region, and watch config reload from the assigned job. That is deliberate. A stale event cannot choose a Deep run you never approved. Work facts are job-bound.

Each delta card shows signal key, before, after, confidence, and detected-at. License lapsed. Permit filed in the service radius. Review velocity cooled. Receptionist role reposted. AI-invisible to cited, or the reverse, when a probe sample exists. Coverage lost is its own reason: a family that used to resolve and now cannot is not a clean absent.

Trigger outreach and Update fix plan sit on the card. Both persist first. Relay still drafts. Forge still opens from fulfillmentActionId if the new state maps to a fix. You still approve a send. Watchtower is a read that creates a dated finding, not a new sender.

not-comparable is the honest cousin of failed. Depth changed, artifact family missing, or a detector version moved. The job should not force a fake before and after across incompatible snapshots. Wait for a clean pair. Baseline is the first successful snapshot. It is not drift. It is the ruler.

Persist first still applies before a watch exists. You cannot watch a pin you never saved. Opening Watch from a diagnostic header writes the record on the same business uuid Scout already kept. A Maps URL is not a subject.

Why timing is a finding, not a reminder

Timing is a finding because a dated change on a business you already understand is a sentence you can send this week. A reminder is a ping that the row still exists. Marketers already have calendars. They do not need another digest that says remember Austin dentists.

Public-record drift is the overlooked half. License boards, permit portals, UCC lapse windows, entity changes, health inspections where an adapter exists. The owner did not tweet these. They still change the week. A pulled kitchen permit next to a cafe you already diagnosed is a different conversation than no website, again.

Hiring reposts are the other half. Two snapshots of the same receptionist role are a buying window if you sell inbound voice. Role-to-service still maps the DEMO. The date on the second post is why the email is not creepy leftover mail from March.

Pulse freshness and signal decay are cousins. A live signal is one that has not aged out of its profile. Watchtower writes the new checked-at that decay can read. ICP minimumRecord can demand a live signal before a row reaches an outreach queue. Timing that is not stored is a story you told yourself.

UCC lapse windows are calendar math: filed date plus about five years minus a notice window. That date is a refinance conversation if you sell that work. Entity change is a new owner or a new location. Health inspection drift is a restaurant week. Storm playbooks are scoped NOAA-style watches you compose, not a national siren we turn on for everyone.

A calendar ping without a diff is a digest. The product you are buying is the stored change. If Monday's feed cannot name before, after, and detected-at, you are looking at a reminder dressed as monitoring.

What license, permit, and rating drift look like on a watch

License, permit, and rating drift look like before and after values on a named detector, not a weather icon. license-drift after a board lookup: active to lapsed, or a site that still claims bonded next to a board that says expired. Confidence is high on the status when the adapter covered the jurisdiction. Unknown(no_coverage) is the honest miss.

permit-lifecycle can move from none to filed, or from filed to stalled when a portal shows inactive past a threshold with no final inspection. Job values on permits are self-reported. Do not quote them as revenue. Quote the date and the address. The DEMO, if you have one, is still the site or the intake path, not a loan.

Rating and review-velocity diffs come from listing sources you actually connected. Google Places can move a star or a count. Yelp and Tripadvisor remain monitor or assisted. A velocity change you cannot source stays off the card. Echo can draft a reply later if they are a client. Collection still cannot gate.

Storm and seasonal watches are scoped playbooks you can ask Nova to compose: watch my Austin dentists for storm damage. The chat can preview match counts before you commit. That composition is not the same sentence as the overnight loop already ran last night. The watch record is what you are buying. The dispatch habit is opt-in and supervised first.

Review-response-rate diffs need a source that actually returns owner replies. Google can move. Yelp stays monitor or assisted. Do not alert on a platform you only imagined. Photo-count ticks of one are library, not a wedge, unless you sell photography and the listing went from two pictures to none.

AI-invisible to cited, or the reverse, needs a fresh admitted probe sample. Watchtower will not invent a k/n flip from a failed adapter. Coverage lost is its own reason: a family that used to resolve and now cannot is not a clean absent.

How exponential backoff stops a dead site from burning spend

Exponential backoff stops a dead site from burning spend by closing the subject after a failed attempt instead of reclaiming it every time a lease expires. last_scan_at advances from the attempt. next_attempt_at grows on top of the cadence. After five consecutive failures the subject suspends, and the watchlist is supposed to say watchSuspended.

Before that rule, a business that could never be scanned was re-claimed on the tenant's provider budget because success was the only clock and the claim lease was designed to expire. That is a quiet invoice. The failure handler exists so the loop stops.

A failed run stores no snapshot. Nothing was measured, so nothing is comparable, so nothing may be reported as a change. If the failure handler cannot reach the database it throws, and the run goes red. A subject silently stuck claimed is the other lie.

Retries on the function are two. Then onFailure is the only close-out. Poison-pill quarantine after repeated engine failures is a sibling idea on the fetch fleet. Watchtower should not DDoS one dentist because a thousand tenants watched the same domain. Politeness is global per domain, not per tenant.

Spend caps still sit on the watch record and in Settings. The tighter one wins. When a check would exceed, the job should queue or skip and say so. Silent overspend is how a dead site becomes an invoice you cannot explain.

What Watchtower will not invent from a failed scan

Watchtower will not invent drift, opportunity dollars, or a cleaner listing from a failed scan. Partial artifacts can still run detectors on what exists. Unreachable stays unknown. The feed does not get a fake lapse to keep the badge count interesting.

Conflicting sources stay conflicting. The engine does not resolve a board versus a homepage in the basement and then alert you as if the basement were a fact. Coverage lost is the truthful middle: we used to know, we do not know now.

Modeled dollars on an old diagnostic do not refresh themselves because a watch ticked. If you need a new range, you re-quantify after a successful snapshot. Atlas schedules can show score deltas between report versions when a compose ran on fresh findings. That delta still needs a real run underneath.

The sweep key is a platform token, not a Places key. Absent sweep key means the sweep does nothing and says so. It never falls back to unauthenticated work. Tenant keys still meter the actual lookups. Unknown adapters stay unknown.

How a delta becomes outreach, a DEMO, or a retainer conversation

A delta becomes outreach, a DEMO, or a retainer conversation by being a dated finding on a business you already have a right to talk about. New prospects get a Relay draft that names the change. Existing clients get a Steward conversation or an Atlas schedule that shows the delta on the next branded memo.

I noticed your listing dropped and the permit next door filed last Tuesday, so I already refreshed the DEMO. Here is the live link. That sentence needs a stored delta, a fulfillmentActionId if you built something, and a send that passed ComplianceGate. It does not need a motivational digest.

Update fix plan is for when the new state maps to work you did not sell yet. A lapse can open a labeled compliance draft. A booking button that disappeared can open the booking tool again. A citation share that fell, when a probe sample exists, can open schema and service pages. DEMO first. DEPLOY still gated.

Pulse can move the row or bump freshness. Forecast still sums FEE, not a poetic monthly from the alert. Health bands stay bands. Do not print the client's revenue because a watch fired.

Actioned on a delta is how you mark that you already drafted or already fixed. The feed should not nag the same lapse every morning after you worked it. Duplicate cards with the same checked-at are a bug, not a hotter signal.

Who should put a business on watch

You should put a business on watch when you already spent a diagnostic and you would take the work if the date moved. Hot prospects you have not sent yet. Open deals in Pulse. Won clients whose license, reviews, or GBP you are paid to mind.

You should not watch a metro. That is Scout plus an ICP, run again. Watchtower is not a second discovery engine. Density of watches on one tenant is a spend question. The cap on the watch record exists because Deep adapters on a daily cadence will find the bottom of a key.

Agencies running many books care about isolation. Watches are tenant-scoped. There is no path from one client's monitored dentist into another client's feed. v1 does not resell white-label sub-accounts. Tenant isolation is still the story.

If you only needed a weekly PDF of rank, a rank tool is lighter. If you needed a calendar reminder, a calendar is lighter. Watchtower is for marketers who sell the change.

Run now is how a careful tenant starts. The governance plane and the watch records can exist while overnight dispatch is still a build. Ask what lastWatchScan says on the row. Do not take a slogan for a log.

How Watchtower sits next to Pulse freshness and report schedules

Watchtower sits next to Pulse freshness and report schedules as the writer of dated signal changes, not as a third CRM. Pulse decay profiles decide whether a finding is still live enough to lead an email. Atlas schedules regenerate a branded artifact and can print a score or dollar delta when compose ran on new findings.

signal_changes is the store automation also wants. If writers are missing in an environment, Health bands can look seeded. This page will not tell you the Health board is a live derivative if the writers are not there. Ask the product what lastWatchScan says on the row in front of you.

Regulatory Watchtower Report is a lens in the report catalog. It packages license, permit, UCC, and bond drift for a leave-behind. M05 does not re-scan. It composes. If the watch is stale, the report is stale, and the timestamp should say so.

Nova can answer watch my Austin dentists for storm damage with a scoped playbook preview. Turn off the storm playbook should pause a real record. What did you do last night should only narrate runs that happened. Dry-run remains the honest first enable.

Common mistakes when treating Watchtower as a digest

The first mistake is enabling daily Deep on every saved lead because monitoring sounded like value. Cadence should match volatility. TLS can wait. Permits and ratings should not. Spend caps are not decoration.

The second mistake is sending every delta. A photo count that ticked up by one is not a wedge. A lapse, a stall, a repost, a probe sample that flipped, a rating drop with volume: those are sentences. The rest is library.

The third mistake is leaving suspended rows in the story you tell a client. Suspended means we stopped paying to fail. Fix the site, the key, or the jurisdiction, then resume. Do not invoice a watch that has been red for a month.

The fourth mistake is confusing a reminder with a finding. If the card cannot show before and after, you do not have Watchtower. You have a ping. Get started if you want into the next batch. The public site will not invent a monitoring SKU price.

What you configure when you put a business on watch

You configure families or specific signals, a cadence, alert routing, and a spend cap when you put a business on watch. Watch from a row, from the diagnostic header, or from a bulk action on selected opportunities. The record is what M08 and the Inngest sweep are supposed to consume.

Cadence is daily, weekly, or monthly in the spec UI. Daily Deep on a permit adapter will find the floor of a key. Weekly Standard on GBP-ish L0 plus a license check is how most retainers should start. Monthly is for slow facts: domain expiry, some SoS fields. Match volatility.

Alert routing is in-app, an email digest to you, or both. The digest is for the marketer. It is not a letter to the owner. If you want the owner to see drift, Atlas schedules a client-facing pack with a delta, after they are a client or after you chose to show them a prospect report.

Spend cap is local to the watch. A global monthly cap still sits in Settings. The tighter one wins. When a check would exceed, the job should queue or skip and say so. Silent overspend is how monitoring becomes an invoice you cannot explain.

States on the row should be readable: watchlist-only, scheduled, processing, backoff, suspended, error. Backoff means we failed and we are waiting longer. Suspended means we stopped. Processing means a claim is live. If the UI shows a green monitoring chip on a suspended row, that is a defect.

How a sweep claims work without double-scanning a shop

A sweep claims work without double-scanning a shop by leasing a subject for a short window, emitting one job per business, and refusing a second concurrent function on that business id. The 30-minute claim lease exists so a dead worker does not hold the row forever. The close-out on failure exists so the lease cannot become a forever bill.

Job-bound reload is the other half. The event is not allowed to carry depth or detector scope as mutable truth. If you change the watch config after a job was queued, the running function should read the assigned job, not a wish you typed later in chat.

WATCHTOWER_SWEEP_KEY is a platform token that starts a sweep. It is not a Places key and not a license-board key. Rotation is an owner operation. Absent key means the sweep idles and says so. Tenants still pay their own lookups when a scan actually runs.

Global artifact cache still helps. The public homepage of a cafe is stored once. Tenant rows store derived signals. A thousand watches on the same domain should not fetch a thousand times. Politeness locks are per domain globally. Watchtower is not a license to DDoS a dentist because your playbook is eager.

Idempotency keys on durable steps keep a retry from writing two deltas for one snapshot. If you see duplicate drift cards with the same checked-at, that is a bug, not a hotter signal. Actioned on a delta is how you mark that you already drafted or already fixed. The feed should not nag the same lapse every morning after you worked it.

Does Watchtower email a shop when a rating drops?

Watchtower does not email a shop when a rating drops. A delta can stage a Relay draft that names the change. Sending still passes ComplianceGate on a domain you verified. Supervised is the default. The owner is your customer. The feed is yours.

Rating diffs come from listing sources you actually connected. Google Places can move a star or a count. Yelp and Tripadvisor remain monitor or assisted. A velocity change you cannot source stays off the card.

Echo can draft a reply later if they are a client. Collection still cannot gate. A digest to you is not a letter to the owner. If you want the owner to see drift, Atlas schedules a client-facing pack after they are a client or after you chose to show them a prospect report.

Photo-count ticks of one are library, not a wedge, unless you sell photography and the listing went from two pictures to none. Send the deltas that are sentences. Leave the rest in the feed.

What is the difference between a reminder and a watch delta?

A reminder is a ping that the row still exists. A watch delta is a dated before and after on a named detector, with confidence and a detected-at time. Marketers already have calendars. They do not need another digest that says remember Austin dentists.

Only the changed reason writes a delta you can sell. Baseline is the first successful snapshot. It is the ruler, not drift. not-comparable and failed write no sellable change. Coverage lost is the truthful middle: we used to know, we do not know now.

Pulse freshness and signal decay are cousins. A live signal is one that has not aged out of its profile. Watchtower writes the new checked-at. ICP can demand that live signal before outreach. Timing that is not stored is a story you told yourself.

If the card cannot show before and after, you do not have Watchtower. You have a ping. Get started if you want into the next batch. The public site will not invent a monitoring SKU price.

Can I watch a whole metro the way Scout searches one?

You cannot watch a whole metro the way Scout searches one. Watchtower re-scans businesses you saved and chose to watch. A metro hunt is Scout plus an ICP, run again. Density of watches on one tenant is a spend question.

The cap on the watch record exists because Deep adapters on a daily cadence will find the bottom of a key. Weekly Standard on listing facts plus a license check is how most retainers should start. Daily Deep on every saved lead is how monitoring becomes an invoice you cannot explain.

Nova can answer watch my Austin dentists for storm damage with a scoped playbook preview. That composition is not the same sentence as the overnight loop already ran last night. Turn off the storm playbook should pause a real record.

Intelligence, when preview is on, can roll a category times metro cell. That is not a watch. A metro does not go on watch. A business does.

What happens after five failed scans on the same site?

After five consecutive failed scans on the same site, the subject suspends, and the watchlist is supposed to say watchSuspended. Exponential backoff grew next_attempt_at on top of the cadence. last_scan_at advanced from each attempt. The loop stops paying to fail.

A failed run stores no snapshot. Nothing comparable means nothing may be reported as drift. Partial artifacts can still run detectors on what exists. Unreachable stays unknown. The feed does not get a fake lapse to keep the badge count interesting.

Suspended means we stopped. Fix the site, the key, or the jurisdiction, then resume. Do not invoice a watch that has been red for a month. Do not leave suspended rows in the story you tell a client.

If the failure handler cannot reach the database it throws, and the run goes red. A silent SUCCESS with no snapshot is the one outcome the function may never produce. Absent sweep key means the sweep idles and says so.

How does a won client stay on watch without a second import?

A won client stays on watch because Steward inherits the same business uuid. Won and still monitored is a saved view. Recurring Atlas reports are a sibling, not a replacement. The watch record still holds families, cadence, alert routing, and a spend cap.

There is no second import when they pay. Leftover findings become upsells. License, reviews, and GBP you are paid to mind stay on the same row. A stale diagnostic with a confident dollar is how you send last month's permit as if it were this morning.

Regulatory Watchtower Report is a lens in the report catalog. It packages license, permit, UCC, and bond drift for a leave-behind. Atlas composes. It does not re-scan. If the watch is stale, the report is stale, and the timestamp should say so.

v1 does not resell white-label sub-accounts. Watches stay tenant-scoped. Two books share a feed only if they share a login. Isolation is the story you are buying.