HYX

Study 03 / edge intelligence

Hyx research study

Learning without the hot path.

A production study of how Hyx observes, scores, learns, challenges itself, and changes models without putting training in the request path.

Research classification

Repository invariants / controlled suites / primary literature

E1 / finding
62
canonical features
one ordered contract / training + serving
E2 / finding
60s
actor behavior window
first-hit content + bounded recent history
E3 / finding
18%
newest temporal holdout
forward generalization / no future leakage
E4 / finding
69
focused Hyx AI tests passed
decision signals + scoring + safety

Abstract

The model is one component. Trust comes from the loop around it.

Research question / RQ-03

Can an edge intelligence system adapt to hostile traffic while keeping request-time inference bounded, training evidence attributable, and model change reversible?

01 / Method

Trace a request through Hyx AI review, then follow the result through learning, quality checks, review mode, and release.

02 / Finding

Hyx separates fast evidence use from slow evidence revision. The edge consumes an immutable model; the classifier earns the right to replace it through multiple independent gates.

03 / Publication boundary

This study reports contracts, lifecycle, safety properties, and test evidence. It does not expose model artifacts, private corpora, detector inventories, or reconstructable thresholds.

System boundary

Four crates. Two planes. One attributable decision.

S-01 / intelligence topology

edge plane / learning plane

01

Decision signals

Turns observations into weighted request evidence and recent visitor history.

02

Shared decision language

Keeps signal meaning and order consistent between learning and live review.

03

Hyx AI

Loads the approved version and returns a confidence score or an explicit failure.

04

Hyx AI learning

Builds examples, learns candidate versions, checks them separately, reviews them on live traffic, and controls release.

Edge plane / synchronous

Observe → construct → score → enforce

No training, corpus mutation, or remote feature lookup is required to score a request. The edge reads local state and a loaded model snapshot.

Learning plane / asynchronous

Collect → label → train → challenge → publish

The classifier is a separate service. It can fail, pause, or reject a candidate without becoming an availability dependency for request delivery.

Feature contract

A feature means the same thing before training, during scoring, and after attribution.

F-01 / canonical vector

62 ordered values / append-only growth

Block / 0116

Actor behavior

Windowed request volume, diversity, response shape, content markers, and behavioral ratios.

Block / 0223

Request evidence

Path, query, payload shape, transport identity, rule score, bot structure, and first-hit intent.

Block / 0316

Long-horizon reputation

Cross-site activity, identity churn, historical rates, independent ground truth, and entity context.

Block / 047

Appended context

Transport/header consistency and host-class context added without reindexing earlier features.

Invariant / width + order

62

Legacy vectors may be padded when meaning is preserved. Oversized or incompatible vectors are rejected rather than guessed.

Why the contract matters

Train/serve skew is treated as a correctness defect, not a model-quality footnote.

The shared schema, transformations, sampled serve-time vectors, width checks, and model-version attribution all exist to make divergence observable before it silently changes decisions.

Evidence loop

Current evidence scores now. Completed evidence teaches later.

L-01 / closed-loop lifecycle

no training on request threads

01

Emit

Detectors attach weighted observations to one request; exclusions can suppress known false-positive classes.

02

Score

Prior actor history and current request content form the vector before the backend, enabling a first-hit verdict.

03

Record

After completion, real status and final security evidence enter the actor window without feeding Hyx’s own block back as origin behavior.

04

Learn

The classifier reconstructs current-contract examples, weighs stronger evidence, and trains a reproducible candidate outside the edge.

05

Validate

Newest-traffic holdout, anchors, golden examples, calibration, and drift checks test the candidate at the serving operating point.

06

Shadow

The candidate scores live vectors without enforcement until time, traffic, decision-rate, and corroboration gates resolve.

07

Promote

Only an approved candidate advances the version pointer observed by the fleet.

08

Rollback

The prior version remains addressable, allowing operational reversal without rebuilding or redeploying the edge.

Model safety

A candidate must disagree safely before it is allowed to agree in production.

01

Contract

Wrong-width vectors and incompatible models are rejected instead of being silently reinterpreted.

02

Forward holdout

The newest traffic slice is held out so evaluation measures the direction production actually travels.

03

Independent evidence

Behavioral anchors and a separately managed golden corpus constrain false positives and recall at the live threshold.

04

Live shadow

A candidate must observe real traffic without enforcing, then satisfy volume, block-rate, delta, and corroboration ceilings.

05

Reversible publish

Only the shadow-approved artifact advances the edge-visible pointer; the prior model remains a rollback target.

Sculley et al. (2015) identify hidden feedback loops, data dependencies, and changing external conditions as common production ML failure modes. This architecture addresses them with explicit contracts, attribution, shadow comparison, and rollback.

Inference runtime

The edge receives a probability. Policy retains authority over the action.

R-01 / Hyx AI decision service

approved version / explicit failure / attributable output

Hyx AI review

Approved version, ready locally.

The active Hyx AI version is loaded locally so requests can be reviewed without waiting on another service.

Output contract

A probability or no result.

Wrong width, absent model, or prediction failure returns no score and increments failure telemetry. Failure is never recoded as confident benign traffic.

Decision separation

Inference does not equal enforcement.

The model emits P(malicious). Host-aware policy combines the operating threshold with independent structural and signal evidence to choose observe, challenge, or block.

01 / attribution

model version

02 / attribution

inference latency

03 / attribution

reason features

Hyx AI combines a consistent set of decision signals with a controlled release process. The same meanings are used during learning and live review, and updates must pass independent checks before release.

Interpretation

Architecture can constrain risk. It cannot manufacture ground truth.

01

No public accuracy claim

This study does not publish a current fleet-wide precision, recall, or latency headline. Those values depend on live traffic, operating threshold, corpus state, and model version.

02

Labels remain evidence

Human labels and independent ground truth are strongest; weak labels remain imperfect. Weighting and gates reduce circularity but do not turn inferred labels into certainty.

03

Production test boundary

The production Hyx AI dependency was unavailable on this host. Shared decision and scoring tests pass, while the complete production-linked test still requires the deployment environment.

The defensible result is architectural: Hyx keeps inference local and bounded, shares feature meaning across planes, records attributable outcomes, and requires independent offline and live evidence before a new model can become authoritative.

Evidence register

Every systems claim maps back to a tested contract or lifecycle boundary.

E1

Consistent decision signals

Hyx AI uses one ordered set of 62 signals during learning and live review.

E2

Behavioral memory

Hyx AI combines current request evidence with a limited window of recent visitor behavior and completed challenges.

E3

Review and record

Hyx AI reviews prior behavior and the current request first, then records the completed outcome for future learning.

E4

Focused Hyx AI suites

The shared decision signals, scoring, and safety checks report 69 passing tests, two generators ignored, and zero failures on this host.

E5

Decision contract

Hyx AI returns an attack confidence level, rejects incomplete inputs, records failures, and never treats a failed review as safe traffic.

E6

Time-based checks

Hyx AI reserves the newest approximately 18% of dated examples for quality checks, with a documented fallback for older records.

E7

Candidate gates

Anchor checks, serving-threshold validation, independent golden metrics, drift diagnostics, and live shadow outcomes precede publication.

E8

Operational attribution

Serving records include engine, model version, probability, latency, and reason codes for flagged requests; promotion retains a previous-version rollback target.

Research archive

Return to the systems studies.

The model, the mechanism, and the evidence required to trust either.

Open the archive