Technical Approach

NETRA v1.5.4 · slide 3 detail

The pipeline

M1 — IngestTCP reassembly. Flows keyed by 5-tuple plus a generation counter, so a reused port cannot merge two conversations. Builds the byte→frame provenance map that makes every later claim citable.
M2 — Protocol IDSMTP / IMAP / POP3 identified by payload grammar. The port is a hint, never proof.
M3 — STARTTLSState machine producing upgraded / failed / rejected / not_used / suspected. Verifies TLS records actually follow the 220.
M4 — TLS factsVersion, cipher, forward secrecy, key-exchange group, post-quantum group, downgrade delta, sentinel, Fallback SCSV, GREASE.
M5 — X.509Chain, expiry, key size, signature algorithm — judged against the capture's own timestamp, never the wall clock.
CorrelationSessions regrouped by server identity. Constraint propagation yields a support matrix: DEMONSTRATED / EXCLUDED_FIRM / EXCLUDED_LIKELY / UNDETERMINED.
Epistemic classifierEvery fact stamped OBSERVED / DEDUCED / INFERRED / NOT_OBSERVABLE. An untiered fact raises an exception rather than taking a default.
M6–M9YAML rules → AI ranking → scoring (observed_risk, evidence_coverage, HNDL) → canonical JSON.

Parsing traps we handle

Version trapTLS 1.3 reports itself as 1.2 in the legacy field. The truth is in supported_versions (RFC 8446 §4.2.1).
Downgrade sentinel444F574E47524401 in ServerHello.random means a 1.3-capable server was talked down (RFC 8446 §4.1.3).
Fallback SCSV0x5600 marks a retry after a failed handshake (RFC 7507).
GREASE16 reserved values must be ignored, never counted as capabilities (RFC 8701).
HelloRetryRequestA fixed random value marks a retry, not a real ServerHello.

Backend

Nothing is storedOne capture, one pass, in memory. No database, nothing kept between runs.
You watch it runEach module reports completion, streamed live over SSE.
The file is hostileSize limit, parse ceilings, timeouts, memory cap, unprivileged user.
One source of truthA single canonical JSON. Dashboard, HTML, PDF and CBOM are views of it.

AI layer

It ranks, it never rulesOrders sessions for a human to check first. Cannot create or change a finding.
Output is labelledTagged INFERRED and structurally barred from the risk score.
Switch it offTurn the AI layer off in the header — every number stays identical.
The AI layer's containment was designed before any model was trained. That order matters: a prediction that can move a measured number destroys the distinction between the two, and it cannot be recovered downstream.