How it works · the data journey

From the wire to an answer
you can trust.

Every fact inside CrossConnect makes the same journey: it enters from the network or your keyboard, gets staged as an observation, is confidence-scored before it is trusted, becomes part of the source of truth, is computed on, and finally leaves as a cited answer or an alert. Here is that journey, drawn out end to end, with no packet ever captured.

The whole journey on one line

Six stages, every time.

Whether a fact arrives over SNMP or is typed in by an operator, it passes through the same six stages. Each stage has one job, and the boundary between them is where trust is earned.

THE TRUST BOUNDARY 01 Enters SNMP, flows, mDNS,traps, manual, API 02 Staged observed,not yet trusted 03 Validated confidence-scoredbefore it is trusted 04 Trusted source of truth+ audit chain 05 Computed scores andrankings 06 Output answersand alerts
Two colors tell the whole story. Navy is what the platform has only observed; teal is what has become trusted. The single amber moment in the middle is the validation gate, where each observation is confidence-scored before it is trusted.
01 Enters

Nine front doors, one rule: observe, never capture.

Data arrives many ways. The platform reads switch-derived signals, listens for the announcements AV gear already broadcasts, accepts traffic summaries, and takes what operators type in. It never captures a packet or inspects a payload, it reads state the network already exposes.

ENTRY POINTS SNMP / LLDP discovery sweep NetFlow / sFlow receiver2055·6343 mDNS listener5353 DHCP fingerprint67 SNMP trap receiver162 Inbound event API Manual entry (UI / REST) Running-config collection External connectors collector Ingestion decode the protocol resolve sender → device stamp the tenant Staging layer one row per observation Switch-derived signals only. No packet capture. No payloads read.
Many sources, one ingestion path. Each front door decodes its own protocol, resolves the sender to a known device, and stamps the tenant, then hands a clean observation to staging.
SNMP / LLDP sweep

The scheduled discovery worker polls each device for its interfaces, neighbors, serial, sensors, and more. The backbone of how the network is found.

NetFlow / sFlow UDP 2055 · 6343

Routers export traffic summaries to the built-in receiver, decoded into who-talked-to-whom, with no packet payloads ever read.

mDNS listener 5353

Joins the multicast group and hears the service announcements AV gear already broadcasts (Dante, NDI, AirPlay), the strongest signal for typing AV endpoints.

DHCP fingerprint UDP 67

Reads the option-55/60 fingerprint from relayed DHCP requests to recognize a device family, corroborating AV classification.

SNMP traps UDP 162

Devices push real-time events (link down, PSU fail, lamp hours). Each becomes a confidence-scored observation for review.

Manual & API

Operators document gear through the UI or REST, and external systems push assertions to the inbound event API. Both are stamped with who and when.

02 Staged

Observed, but not yet believed.

Raw observations land in append-only staging tables, one family per kind of fact. Nothing here is treated as truth. Each row carries when it was seen, the newest one per natural key is the operative one, and old rows are purged automatically. This is the platform's short-term memory of reality.

ingest Discovery staging append-only · newest row per key wins automatically purged after 14 days 15 families to validationconfidence comes next Interface · Neighbor · Endpoint · IP · VLAN · PoE · mDNS · DHCP · PTP · IGMP · TrafficFlow
Short-term memory, kept honest. Fifteen families of observation, append-only so the change history has something to compare against, and self-cleaning so the tables never bloat.
Append-only by design

A new sweep does not overwrite the last one, it adds rows. That history is what powers the operator-facing “what changed since we last looked” view, and the audit trail behind it.

Self-purging

A daily sweep drops observations past the retention window (14 days by default), so staging stays a rolling picture of recent reality rather than an ever-growing log.

03 Validated

Confidence earned before anything is trusted.

A staged observation is not trusted just because it was seen. CrossConnect scores how believable it is by how well it corroborates across sources, labels it with that confidence, and only then commits it to the source of truth. Nothing is ever invented, and a value the platform is unsure of stays visibly flagged.

Staged observationwhat was seen Confidencecorroborate, score Confirmed · 2+ sources agree Inferred · a single source Unconfirmed · stays flagged Source of truthlabeled with confidence Believability is earned by corroboration, never assumed, and every fact carries the confidence it was committed with.
Confidence, never invented. Two independent sources that agree score Confirmed; a single source is Inferred; something that resolves to nothing stays Unconfirmed and flagged. Every fact is committed with the confidence it earned.
Confidence scoring

Believability is earned by corroboration: agreement across sources lifts an observation toward Confirmed; isolation keeps it low and visible.

Never invented

Where the platform is unsure, it labels the value Inferred or Unconfirmed and shows the next check, rather than presenting a guess as a fact.

No silent writes

Every fact that enters the source of truth is written through the audit chain, so what changed, when, and from which observation can always be traced.

04 Trusted

The source of truth, and a history that cannot be rewritten.

Once validated, a fact becomes part of the canonical model: devices, interfaces, cables, addresses, circuits, and the services they deliver. And every change to that model is written into a tamper-evident, hash-linked chain, so who changed what, and in what order, can always be proven.

THE CANONICAL MODELthe one source of truth every screen and the assistant read Device Interface Cable IP / Prefix Circuit Network service TAMPER-EVIDENT AUDIT CHAINevery change is hash-linked to the one before it change #41hash ← #40 change #42hash ← #41 change #43hash ← #42 chain verified Alter any record and its hash no longermatches the next link. Tampering is detectable.
One model, one ledger. The canonical entities are what every screen and the assistant read; the hash chain underneath makes the history provable, which is what lets an auditor take the platform's word for it.
05 Computed

One snapshot, many answers.

The intelligence layers do not store new facts, they compute from the source of truth. Each reads a snapshot and returns a score or a ranked list as a pure function, which is why the same evidence can power data quality, compliance, maturity, and the “what should I fix first” queue all at once.

snapshot Compute engine pure functions of the source of truth cached 90s · always consistent with the model One ranked answerworst-first Data quality · Compliance · Maturity · Hotspots · Readiness · Capacity · Reachability · Segmentation · Threats · CVEs · AV posture · PTP
Compute, do not duplicate. Because the analytics are pure functions of the snapshot, they stay consistent with the model, and a short cache means one computation serves every viewer for the window.
06 Output

It leaves as something you can act on, and check.

At the end of the journey, data becomes an answer. The assistant replies in plain language with the records it used. Webhooks and SIEM sinks carry changes to your other tools. Reports and the API export the model. And every screen reads the same single source of truth.

Source of truth+ computed layers AI assistant · cited answers Webhooks · signed, 50+ events Outbound sinks · SIEM / chat Reports · PDF / CSV / JSON REST API UI views Youact on it, and verify it AI writes never apply on their own,they wait for your confirmation.
Cited, signed, exportable. The assistant proves its answer with records; webhooks and sinks are signed and safe-by-default; and the AI never changes the network without a human pressing confirm.
End to end, in one breath: a switch is polled over SNMP, its new neighbor is staged as an observation, confidence-scoring corroborates and commits it to the source of truth, the audit chain records the change, data quality and the topology recompute, and the assistant can now answer “what is plugged into this switch?” and cite the connection it just learned.
The journey, in one idea

Observe, earn trust, then answer.

CrossConnect treats every fact as an observation until it earns confidence, records the moment it becomes trusted, and computes everything else from that one trustworthy model, so the answer you get at the end is one you can check all the way back to the wire.