preprint · arXiv:wm.2604.0001 · [cs.NI] · web infrastructure Submitted 02 Apr 2026 · Revised 21 Apr 2026 Matrix Systems Innovations Ltd · Cardiff · 15598218
W
WebMatrix Systems
AI-driven web infrastructure · est. MMXXIV
Cardiff · CF10 Companies House 15598218
Incorporated 27 Mar 2024
arXiv:wm.2604.0001 · [cs.NI] · cs.SE · cs.DC v2

The web infrastructure layer that thinks before it routes.

The WebMatrix engineering desk1 · Three directors of record1 · Cardiff edge team1, 2

1 Matrix Systems Innovations Ltd, Cardiff CF10, United Kingdom · 2 Edge PoPs: Cardiff, Manchester, Dublin, Amsterdam, Frankfurt

Abstract

WebMatrix Systems is an AI-driven web infrastructure layer for the enterprise. We describe a control plane that unifies an adaptive edge cache (MatrixEdge), a reasoning-based observability graph (MatrixObserve), and a policy-as-reasoning runtime (MatrixGuard) under a single data model, such that an SRE at 03:14 sees one explained paragraph rather than four uncorrelated dashboards.

Across fourteen production engagements we report a median reduction in mean-time-to-explanation from 28 minutes to 6 minutes for a four-region P1 with twelve services and two upstream dependencies. We outline the data model, the calibration discipline behind each surface, and the conditions under which the platform's inference layer is — deliberately — silent.

Keywords edgeobservabilityWAFpolicy-as-codeSREincident-explanationotlpkernel-isolation

The thesis

For two decades, web infrastructure has been a stack: a CDN below, a WAF beside, an observability platform overhead, and an alerting layer above that. Each layer is the property of a different vendor; each speaks a different data model; each emits its own dashboard. The aggregate is functional — most of the web works — but it is functional in the way a switchboard is functional: by holding everything together with human attention.

The WebMatrix thesis is that the integration is the product. MatrixEdge knows the route. MatrixObserve has the trace. MatrixGuard sees the request. When the three live on one control plane and one data model, the platform can reason about an incident in the way an experienced SRE reasons — by joining the route, the trace and the request together — and produce, at the moment the page fires, one paragraph that names what happened and what is on fire.

Proposition 1.1(Integration over assembly) Let E, O, G denote the edge, observability and policy surfaces respectively. The operational value of E ⊕ O ⊕ G on a common data model exceeds the value of E ∪ O ∪ G when joined post-hoc by humans, by a factor that grows with incident complexity. The 28 → 6 minute reduction (§4) is one measurement of this proposition.

One platform. Three things it does well.

MatrixEdge — adaptive edge cache, routing, TLS

An edge cache whose TTLs are not human guesses but a per-route, per-hour, per-region reinforcement loop. The loop optimises for a declared tuple — origin cost, p99 tail latency, hit ratio — under a constraint set the operator writes once. We publish the loop, the constraints and the calibration plots. We do not publish the cache-key, which is yours. Read the MatrixEdge note →

MatrixObserve — observability with reasoning

Logs, traces, RUM and synthetic checks join into one graph. The inference layer reads the graph at incident time and produces a one-paragraph explanation rather than a wall of correlated lines. The eval is run monthly and the eval set is public. Read the MatrixObserve note →

MatrixGuard — policy as reasoning

Security policy as a natural specification, compiled to enforcement, validated by a simulator before promotion. The compiler is deterministic; the simulator runs against the last 30 days of production traffic; the runtime emits an explanation per block rather than a regex match. Read the MatrixGuard note →

SurfacePrimary verbData classEval
MatrixEdgeRoute, cache, terminate TLSHTTP · DNS · TLSPer-route, hourly
MatrixObserveExplain incidentsLogs · traces · RUM · syntheticPublic eval, monthly
MatrixGuardEnforce policy with explanationRequests, sessions, headersSimulator, pre-promotion

Table 1 The three surfaces, one data model. Eval surfaces are documented in §Docs and the per-surface evaluation appendices.

Three things we shipped this quarter that we want you to read.

  1. arXiv:wm
    2603.0001

    Policy as reasoning: writing WAF rules without writing regex.

    The MatrixGuard specification language, the compiler, the simulator and the runtime. Why a policy you can read is worth one you can grep, and what fifteen months of production traffic at a Cardiff retailer taught us about the eval.

    Released 03 Mar 2026 · MatrixGuard · engineering desk
  2. arXiv:wm
    2602.0003

    Observability with reasoning: from log dumps to incident paragraphs.

    How MatrixObserve takes a P1 — twelve services, four regions, two upstream dependencies — and produces a one-paragraph explanation in the time it took the on-call to pour coffee. With the public monthly eval and the things the system is honestly bad at.

    Released 17 Feb 2026 · MatrixObserve · engineering desk
  3. arXiv:wm
    2602.0001

    MatrixEdge auto-tune v2: per-route TTL adaptation, online.

    The reinforcement loop that decides per-route, per-hour, per-region TTLs from observed traffic and downstream cost. How it is calibrated, why the dashboard is public, and what the loop will not do for you no matter how confidently you ask.

    Released 03 Feb 2026 · MatrixEdge · engineering desk

Talk to engineering.

We do not run a sales motion. The address below reaches the engineering desk; the first reply is written by the engineer the question would land on, not by a front-line responder. A working session is forty-five minutes against your origin, your traces, your policy file. The output is a written technical note you keep whether or not the engagement continues.

BibTeX · @techreport{webmatrix2026, author = {Matrix Systems Innovations Ltd}, title = {{WebMatrix: one control plane for adaptive edge, observability and policy}}, year = {2026}, institution = {Cardiff, UK}, number = {wm.2604.0001}}