No claims — evidence

Evidence

Everything on this page is verifiable: diagrams, code excerpts and figures whose method is published.

This page does not claim — it shows. Every element here is verifiable: architecture diagrams, code excerpts from the deciding layer, and measurement figures whose method is stated.

1. Architecture: the request path through the hybrid WAF

Inbound requestNormalise & validateSignature rulesBehavioural anomalyRate limitingContent analysisSource reputationLLM arbitration
Cheap layers run first; the LLM layer is only invoked on doubt — the "least expensive first" architecture

The core design point: layers are ordered by processing cost, not by strength. Deterministic layers (rules and signatures) settle the majority, and the LLM arbitration layer only handles a small minority of suspicious requests. Cost and latency stay low while decision accuracy rises.

2. Excerpt from the arbitration layer

A short illustrative excerpt showing the decision logic: deterministic layers produce signals, arbitration is invoked above a threshold, with a timeout and a fallback path if the model does not respond.

# Arbitration layer - simplified logic
THRESHOLD = 0.62          # suspicion threshold from deterministic layers
TIMEOUT_MS = 350          # a slow model must never break the user experience

def decide(request, signals):
    score = weighted_score(signals)          # 0..1 from deterministic layers
    if score < 0.20:
        return ALLOW                          # most traffic ends here
    if score > 0.85:
        return BLOCK                          # decided without the model

    verdict = llm_arbitrate(request, signals, timeout_ms=TIMEOUT_MS)
    if verdict is None:                       # timeout expired
        return CHALLENGE                     # challenge, do not block the unknown
    return ALLOW if verdict.allow else BLOCK

The decision I consider most important here is the last line of the timeout branch: when arbitration fails we choose challenge, not block. Blocking on ignorance costs real users; a challenge costs an attacker one extra step.

3. Measurements — with the method stated

MetricValueHow it was measured
Requests reaching the arbitration layerunder 5%Internal measurement on a test set mixing benign and attack traffic
Latency added by arbitrationnone on 95% of requestsResponse-time comparison with the layer disabled and enabled
False positives vs a fixed thresholdnoticeably lower in design testingA set of legitimate requests hardened step by step until a fixed threshold trips
Decision time for a suspicious requestcapped at 350 msA hard bound fixed in code, not a stochastic result

Transparency: the figures published here are internal measurements on test sets described in the third column, not production results for a client. Any number I did not measure myself is not published.

4. Evidence for the live tools

The seven tools actually run and can be exercised from a browser — the strongest evidence possible is real use, not a screenshot. Each tool has a project page explaining its scope and its evidentiary limits; all of them work on public information and bypass no authentication.

5. Written evidence: two full technical case studies

Two studies publish the methodology and the decisions rather than headlines only: the hybrid WAF study and the DDoS mitigation study. Each explains the problem, the options available, why each choice was made, and what did not work.

6. Runnable evidence: public repositories

Anything published under the MIT licence can be read, run and forked. Public account activity and the four repositories are on GitHub — with the publication standards shipped inside the repositories themselves.

7. What is not available yet — plainly

  • Tool screenshots: the tools run live, and live use is more honest than a static image.
  • Client production figures: the contracted projects ran under agreements protecting client data, so no attributable number is published.
  • Credentials not yet earned: I do not publish a certificate before it is issued; plans and dates are public in the roadmap.