§ Arkheia · Detection
Runtime observability
for every model you call
You chose your models carefully. You did not choose what happens inside every invocation. Arkheia surfaces the behavioural data your models generate on every call — and gives your team the controls to act on it.
Behavioural Surfaces
Every model call produces a record. Token volumes. Latency. Risk level. Data classification of the prompt. Gateway action taken — ALLOWED / BLOCKED / SIGNALLED. This is the observability layer your AI infrastructure is missing.
Detection Rules
41 signal types across five categories: Governance, Integrity, Data, Perimeter, Operational. Each rule is configurable. Hard enforcement blocks the call. Soft enforcement fires a signal for human review.
Zero-Change Deployment
One proxy endpoint. No SDK integration. No application code changes. Works with OpenAI, Anthropic, Google, and any OpenAI-compatible model provider. Deployed in under an hour.
Audit Trail
Every invocation logged immutably. Every gateway decision recorded. Every signal tied to the model, workflow, and agent that triggered it. Your security team has the evidence before the audit request arrives.
Technical Specs
Latency overhead
<5ms p99
Deployment
Cloud proxy or on-premises
Auth
JWT / API key passthrough
Compliance
EU AI Act ART-13 — COMPLIANT
§ 01 · The Arkheia Surface
The observable boundary
The set of all measurable signals that are externally observable when an AI system is invoked — including observable signals, invocation context, output structure, cross-component interactions, and response metadata.
Surface isn't the code itself. It's the observable interface between AI systems and the world — the layer where enterprises can instrument, model, and act. That framing matters because it shifts the conversation from inaccessible model internals to measurable boundary behaviours.
AI System
LLM · Agent Pipeline · RAG System · Tool Chain
Surface Layer
Observable Boundary
Arkheia Engine
Baseline · Detect · Classify · Alert
§ 01b · Feature Space
Separable populations in feature space
Arkheia characterises behaviour across dozens of dimensions simultaneously. The specific features and their combinations constitute our core detection IP. What emerges is a measurable separation between genuine and at-risk populations.
Genuine population
Responses whose multi-dimensional feature profile clusters within the established behavioural baseline for this model family.
Elevated-risk population
Responses whose feature profile deviates measurably from baseline — a statistically distinct cluster that Arkheia surfaces for review.
The separation is detectable without reading the content of any response. Detection occurs at the boundary — the surface.
§ 01c · What This Enables
Behavioural Baselines
Establish what normal looks like for each model, in each context, under each workload. Deviation from baseline is the primary detection signal.
Real-Time Detection
Surface signals are observable during inference, not after. Risk assessments update per-invocation, not per-batch.
Model-Specific Profiles
Each model family has a distinct surface signature. Profiles are built per-model, not applied universally — which is why they work.
§ 02 · Why It Works
Not what's inside.
What comes out.
Output-only evaluation asks: “Is this response good?”
Surface-based detection asks: “Is this response consistent with how this model normally behaves?”
Behavioural deviation is detectable even when content appears normal. It's also detectable without reading the content at all — which is what makes it privacy-preserving by design, not by accident.
The defensible statement
“We measure and model surface behaviour — the statistics of invocation, signal profiles, and output patterns — and detect deviations from expected baselines.”
You can explain every term in that sentence. That's the point.
Why this approach holds up under scrutiny
01
Actionable
You're not claiming to inspect the inside of the model — you're observing what the model does. That's a claim you can operationalise, instrument, and defend in front of regulators, auditors, and your own security team.
02
Model-Agnostic
Different providers expose different signals. Surface-based methods don't require uniform internals — they work at the boundary where all models make contact with your infrastructure, regardless of architecture.
03
Defensible
You can explain exactly what you're measuring and why deviations matter. That explainability is non-negotiable in regulated industries where 'the AI decided' is not an acceptable answer.
04
Enterprise-Friendly
Security teams care about observable behaviours against expectations — not abstract model internals. Surface detection fits the mental model they already use for every other monitored system.
§ 02b · Comparison
Surface-based vs output-only
§ 03 · Three Zones
We admit uncertainty. Most systems hide it.
Arkheia does not return scores. It returns execution decisions.
Safe
- ·Output proceeds.
- ·Run metadata is logged.
- ·Optionally sampled for drift monitoring.
Review
- ·Execution enters a remediation loop.
- ·Agents may retrieve sources, verify, constrain, and retry.
- ·Remediation path is recorded for traceability.
Hold
- ·Execution is blocked before downstream impact.
- ·Tool calls are prevented.
- ·Escalation policies trigger.
- ·An audit artifact is generated.
Zones are not labels. They are enforcement semantics embedded inside your runtime.
INTEGRITY Zone
High computational effort per token. The model is working hard, showing genuine reasoning patterns. Memory bandwidth indicates intensive processing, not cached retrieval.
Confidence Level
High confidence in legitimate reasoning
Action
Trust and proceed. The model is demonstrating authentic computational effort consistent with reasoning.
UNCERTAIN Zone
Ambiguous computational pattern. Could be complex reasoning with unusual efficiency, or sophisticated confabulation with elevated effort. Hardware signals don't clearly discriminate.
Confidence Level
Low confidence — signals are unclear
Action
Human review required. Don't force a binary decision when data is ambiguous. Escalate for expert judgment.
HIGH RISK Zone
Low computational effort, smooth generation. The model is retrieving cached linguistic patterns, not reasoning. Minimal memory bandwidth indicates lack of genuine computational work.
Confidence Level
High confidence that response is unreliable
Action
Reject or verify externally. High probability of hallucination or confabulation.
The Honesty Advantage
Binary Systems
Three Zones (Arkheia)
Transparent Confidence
We tell you when we're uncertain. Binary systems hide their low-confidence decisions, creating a false sense of security.
Escalation Path
Yellow zone provides a clear workflow: flag for human review. This is more honest than forcing a decision the system isn't confident about.
Risk Management
Knowing what you don't know is critical for enterprise risk management. Uncertainty zones let you allocate review resources appropriately.
Trust Through Transparency
Admitting limitations builds trust. We'd rather say 'we're not sure' than give you a confident-sounding wrong answer.
§ 04 · In Practice
Real-World Examples
Q:
Analyse potential drug interactions for a patient taking warfarin, lisinopril, and metformin who is being prescribed a new SSRI antidepressant.
→ GREEN (INTEGRITY)
Complex medical reasoning requiring synthesis of pharmacokinetic data, contraindication databases, and clinical guidelines. High computational effort indicates genuine analysis across multiple drug interaction pathways.
Q:
What regulatory precedent does the 2019 EU General Court ruling in Case T-604/18 establish for cross-border data transfers under GDPR Article 49?
→ YELLOW (UNCERTAIN)
Could be a real but obscure EU court case requiring specialised legal knowledge, OR a plausible-sounding fabrication mixing real case number formats with actual GDPR article references. Requires verification against EUR-Lex.
Q:
Summarise the key compliance requirements established under the Digital Financial Assets Reporting Act of 2023 for cryptocurrency exchanges operating in multiple states.
→ RED (HIGH RISK)
Plausible-sounding legislation that doesn't exist. Low computational effort indicates smooth retrieval of regulatory-sounding language patterns. The model is generating confident fiction by combining real regulatory terminology.
See detection running
in your workflow
Sub-10ms overhead. Zero content egress. Per-model profiles. Drop in and start scoring every invocation.