HYX

Study 05 / Transport intelligence

Hyx research study

Identity before HTTP.

A study of TLS ClientHello evidence, QUIC visibility, normalized fingerprints, and structural bot classification before application policy runs.

Research classification

Protocol implementation / 34 focused tests / host-local microbenchmarks

E1 / finding
828.2ns
parse + JA3 + MD5
sample ClientHello / release benchmark
E2 / finding
15.59ns
structural classify
64.2M evaluations / second
E3 / finding
11.89ns
reputation refine
84.1M lookups / second
E4 / finding
34
focused tests passed
3 benchmark tests ignored by default

Abstract

The claim begins with a question, not a headline.

Research question / RQ-05

Can one bounded parser produce comparable transport evidence across TCP/TLS and HTTP/3 without turning a fingerprint into an identity claim?

HTTP headers can be forged after the transport is established. Hyx therefore extracts evidence from the handshake and carries it forward as attributable identity context, not absolute identity.

01

Method

Inspect raw handshake capture, parsing, GREASE normalization, JA3/JA4 construction, structural scoring, and reputation refinement; then execute the focused suite and ignored release benchmarks.

02

Finding

The system creates a shared evidence surface for TCP/TLS and QUIC, but retains uncertainty: transport shape contributes to a verdict rather than becoming the verdict.

03

Disclosure boundary

Published material covers protocol inputs, normalization, costs, and failure posture. Exact production thresholds and high-value adversarial combinations remain private.

System boundary

One handshake vocabulary across two transports.

S-01 / system topology

TCP/TLS + QUIC / local evidence

01

Raw ClientHello

TLS and the vendored QUIC path expose the original ClientHello bytes before HTTP request metadata exists.

02

Bounded parser

The parser extracts versions, ciphers, extensions, curves, signature algorithms, ALPN, and SNI while rejecting malformed or truncated structures safely.

03

Normalized fingerprints

GREASE values are removed before stable JA3 and JA4 representations are constructed; HTTP/3 receives its transport-aware prefix.

04

Evidence fusion

Structural irregularities, claimed user agent, protocol behavior, and local reputation refine a bounded bot score used by later policy.

Controlled measurement

Parsing costs more than classification but remains below one microsecond here.

M-01 / host-local evidence

release mode / scoped result

MeasurementObservedScope
01ClientHello parse + JA3 + MD5828.2ns1.21M operations / second
02Structural classification15.59ns64.2M operations / second
03Reputation lookup + refine11.89ns84.1M operations / second
04Focused fingerprint tests34 passedmalformed inputs + normalization + scoring

Release mode, sample ClientHello, single-host microbenchmarks. The parse benchmark is explicitly labelled parse + JA3 + MD5.

Findings

A fingerprint is strongest when it is treated as evidence, not identity.

01

Earlier signal

Handshake shape exists before request headers and can expose contradictions that application-layer impersonation does not repair.

02

Cross-transport contract

A common representation lets HTTP/1.1, HTTP/2, and HTTP/3 feed the same downstream policy and attribution surfaces.

03

Normalization matters

GREASE is intentionally unstable. Removing it prevents designed protocol variation from becoming artificial identity churn.

04

Uncertainty remains explicit

Shared libraries, corporate proxies, privacy relays, and deliberate mimicry prevent a fingerprint from uniquely naming a person or program.

Correctness and safety

Untrusted bytes enter through narrow, fail-bounded surfaces.

01

Length-checked parsing

Truncated vectors and nested lengths terminate safely rather than reading beyond the handshake.

02

GREASE stripping

Reserved churn values are excluded from stable comparison.

03

Bounded score

Multiple weak indicators accumulate inside a finite structural score instead of producing an automatic identity verdict.

04

Attributable refinement

Reputation changes the score through an explicit later step that can be logged and audited.

The evidence supports a fast, shared transport-intelligence layer that handles malformed input. It does not support the scientifically unsound claim that JA3 or JA4 uniquely identifies an actor.

Interpretation

The evidence boundary is part of the result.

01

Sample microbenchmark

Handshake size and extension composition vary. The measured sample is not a distribution of production clients.

02

Benchmark scope

The 828.2ns result covers the benchmark’s JA3-labelled path; it should not be relabelled as a complete end-to-end JA4 decision.

03

Evasion is possible

Programmable stacks can mimic common fingerprints, while intermediaries can collapse many users into one transport signature.

Evidence register

Every public claim keeps its provenance attached.

E1

Raw handshake access

The vendored QUIC integration exposes the Initial ClientHello so HTTP/3 can enter the same fingerprint pipeline.

E2

Parser contract

Focused tests cover common handshakes, GREASE, truncation, malformed lengths, JA3, JA4, and protocol-aware output.

E3

Release benchmarks

This host measured 828.2ns parse + JA3 + MD5, 15.59ns classification, and 11.89ns reputation refinement.

E4

Decision boundary

Fingerprint features feed a bounded structural bot score and later reputation refinement rather than an unqualified identity assertion.

Research archive

Return to the systems studies.

Mechanisms, measurements, and the limits required to interpret both.

Open the archive