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
Decision signals
Turns observations into weighted request evidence and recent visitor history.
Shared decision language
Keeps signal meaning and order consistent between learning and live review.
Hyx AI
Loads the approved version and returns a confidence score or an explicit failure.
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
Actor behavior
Windowed request volume, diversity, response shape, content markers, and behavioral ratios.
Request evidence
Path, query, payload shape, transport identity, rule score, bot structure, and first-hit intent.
Long-horizon reputation
Cross-site activity, identity churn, historical rates, independent ground truth, and entity context.
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
Emit
Detectors attach weighted observations to one request; exclusions can suppress known false-positive classes.
Score
Prior actor history and current request content form the vector before the backend, enabling a first-hit verdict.
Record
After completion, real status and final security evidence enter the actor window without feeding Hyx’s own block back as origin behavior.
Learn
The classifier reconstructs current-contract examples, weighs stronger evidence, and trains a reproducible candidate outside the edge.
Validate
Newest-traffic holdout, anchors, golden examples, calibration, and drift checks test the candidate at the serving operating point.
Shadow
The candidate scores live vectors without enforcement until time, traffic, decision-rate, and corroboration gates resolve.
Promote
Only an approved candidate advances the version pointer observed by the fleet.
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.
Contract
Wrong-width vectors and incompatible models are rejected instead of being silently reinterpreted.
Forward holdout
The newest traffic slice is held out so evaluation measures the direction production actually travels.
Independent evidence
Behavioral anchors and a separately managed golden corpus constrain false positives and recall at the live threshold.
Live shadow
A candidate must observe real traffic without enforcing, then satisfy volume, block-rate, delta, and corroboration ceilings.
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.
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.
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.
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.
Consistent decision signals
Hyx AI uses one ordered set of 62 signals during learning and live review.
Behavioral memory
Hyx AI combines current request evidence with a limited window of recent visitor behavior and completed challenges.
Review and record
Hyx AI reviews prior behavior and the current request first, then records the completed outcome for future learning.
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.
Decision contract
Hyx AI returns an attack confidence level, rejects incomplete inputs, records failures, and never treats a failed review as safe traffic.
Time-based checks
Hyx AI reserves the newest approximately 18% of dated examples for quality checks, with a documented fallback for older records.
Candidate gates
Anchor checks, serving-threshold validation, independent golden metrics, drift diagnostics, and live shadow outcomes precede publication.
Operational attribution
Serving records include engine, model version, probability, latency, and reason codes for flagged requests; promotion retains a previous-version rollback target.
References / Harvard style
Probability calibration
Platt, J.C. (1999) ‘Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods’, in Smola, A.J. et al. (eds.) Advances in Large Margin Classifiers. Cambridge, MA: MIT Press, pp. 61–74.
Open source ↗Reference / 02ML systems debt
Sculley, D. et al. (2015) ‘Hidden technical debt in machine learning systems’, in Advances in Neural Information Processing Systems 28, pp. 2503–2511.
Open source ↗Research archive
Return to the systems studies.
The model, the mechanism, and the evidence required to trust either.