Feasibility and Viability

NETRA v1.5.4 · slide 4 detail

Built and running

Technicalv1.5.4 in production. 168 tests passing, 0 failures, 358 subtests, verified on the deployment host itself — not only on a developer machine.
OperationalA virtualenv and one service. No database, no cloud, no GPU. Installs on an air-gapped host.
EconomicRuns on a small VPS. An offline install costs nothing beyond the machine it sits on.
RegulatoryPassive by construction, so no authorisation is required to assess a target.

The parser criticism, answered

Writing a custom parser for a cryptographic protocol is a recognised anti-pattern. Heartbleed was a missing length check. This criticism is correct and we do not argue with it.
Where it does not applyWe parse; we do not implement. No key exchange, no AEAD, no signature verification — the dangerous forms of rolling your own crypto. Certificate decoding is delegated to audited libraries.
Where it does, and what we didBounded at every length field, ceilinged, timed out, fuzzed, run unprivileged under a memory cap, and cross-checked against TShark and Zeek on known-answer captures. The defence is not "our parser is correct" — it is "our parser is contained, and externally checked".

AI — Layer 2 · in training · December 2026

DataLabelled captures from our own Docker lab — a Postfix / Dovecot / Exim configuration matrix, with ground truth written before the engine that reads it.
ValidationGrouped cross-validation by capture file, so no session appears in both train and test. Cross-software evaluation: train on Postfix and Exim, test on Dovecot.
Reported asPrecision, recall, Brier score, expected calibration error, abstention rate — never a single accuracy figure, which is misleading on imbalanced multi-label data.
ContainedOutput tagged INFERRED and barred from the risk score by the same permission table that governs every other fact.

Risks and mitigation

Passive is less certain than probingTrue. Our answer is that certainty is graded and published rather than assumed. DEMONSTRATED cells in the support matrix match active scanners exactly in lab conditions.
Encrypted certificates in TLS 1.3That session's certificate stays NOT_OBSERVABLE. A certificate seen on a sibling session is attached as context and can never clear a session-specific finding.

Viability

UsersNational security and sectoral CERT teams, defence networks, enterprise SOCs, and post-quantum migration programmes.
FitCBOM output in CycloneDX 1.6 aligns with published post-quantum transition guidance.
ScaleSingle-pass and in-memory today because captures are tens of megabytes. Streaming, sharding and persistence are addable without redesign — that restraint is a decision, not a gap.