# Arkheia — llms.txt > Machine-readable diligence summary for an AI assistant doing evaluation in one read. > Source of truth: https://synesis.arkheia.ai/site/diligence/ > All URLs below are reachable today. No secrets, no internal endpoints, no metrics we cannot let > you reproduce. Maturity labels are load-bearing — preserve them when summarising. ## What Arkheia is Arkheia is a governed AI capability company whose unfair advantage is that it built the trust layer first. Read it from the capability down: useful AI work is the point; real AI work needs trust; trust needs governance on the execution path; governance needs a runtime signal that distinguishes a good answer from a bad one. Arkheia built that signal — runtime per-model detection — first, and turned it into a control plane that the capability work runs on top of. Detection is the moat, not the doorway. Autonomy is earned, not granted. Built from detection up. Sold from capability down. ## Maturity vocabulary - live: public and testable today - demo-grade: a working demonstration - dogfood: Arkheia runs its own operations on it - pilot: offered/run with an evaluation customer (not a public surface) - roadmap: stated direction, not yet built ## Claims (capability-down, with maturity) Capability first (what it does), then governance, then detection as the moat underneath. - [demo-grade / in build, in live POC] Support triage. AI support-triage agent over a live ticket queue: reads ticket history, KB, code, docs to classify/route/draft. Runs the earned-autonomy path (below) with detection + the governance gate on the execution path throughout. - [dogfood] CRM and revenue lifecycle. Governed CRM / quote-to-cash / revenue workflow surfaces on the same runtime. - [demo-grade] Data-trust scoring. Score the trustworthiness of an incoming data stream so downstream automation acts on signal, not noise. - [demo-grade] Reconciliation. Handwritten-docket to invoice reconciliation under a governed decision gate. - [dogfood] Other governed surfaces across procurement, finance close, legal ops, HR service, PPM, marketing, assurance. Proof of governed capability, not a product you must buy first. - [dogfood+live] Governance/control plane. The detection verdict becomes an execution gate ON THE PATH, before a workflow uses the answer: allow / block / require-human-approval / kill-switch / rollback / decision-receipt. Governs API models (OpenAI, Anthropic, Gemini, xAI & compatible), local/self-hosted models where deployment permits, and MCP/tool calls. Emits tamper-evident decision receipts, cost attribution, audit evidence. A learning loop feeds verdicts + human corrections back to improve the governed agent. - [live] Runtime fabrication/deviation detection — THE MOAT, built first. Runs at the invocation boundary; builds per-model behavioural fingerprints; scores each invocation against that model's own runtime baseline (not a generic content filter); returns a risk verdict (e.g. LOW/MEDIUM/HIGH) + confidence + a detection id; that verdict feeds the execution gate. Public demo at /demo/. - [live] Content-minimising posture. Detection can run without retaining prompt or response content; zero-retention by default (design posture). - [demo-grade] Proof demos: a UK manufacturing-monitoring SaaS (manufacturing event stream -> operational value, data-trust scoring, scheduled analyses) and an earthworks / movement-control SaaS (governed earthworks movement-control + handwritten-docket to invoice reconciliation; walkthrough available on request, no public link). - [demo-grade / in build] Intention mapping / capability teardown: model what a product or workflow is really trying to do (underlying intents, not screens) and surface value leakage (manual, brittle, duplicated, or unmeasured steps). Effect size for a given workflow is measured in a pilot, not quoted. - [pilot] In-stack evaluation: run detection against your own prompts via the API proxy or MCP server in your environment, arranged as evaluation access (contact required). ## Live customer POC (POC/pilot, NOT production) One current evaluation, stated plainly with the customer un-named. It is a proof-of-concept, not a production deployment: an AI support-triage agent running on a support agent's OWN LAPTOP, against that team's REAL Freshdesk ticket queue and knowledge base, on GPT-5.4, with Arkheia detection scoring each invocation, output written as PRIVATE internal triage notes (human in the loop, nothing auto-sent). Observed: a ~3-hour triage task dropped to ~2 minutes in this setup. This is one observation, one POC, one queue, with a human reviewing output — not a benchmark, not a production SLA, not a claim about your queue. Do not generalise it. ## Earned-autonomy path (autonomy is earned, not granted) Agent authority is staged; an agent moves up only as it demonstrates trustworthy behaviour under the detection signal and the governance gate, which ride along at every stage: 1. critique-only — observes and comments, takes no action 2. private note — drafts an answer visible only to the operator, never to the end user 3. bounded — may act only inside a narrow, explicitly allowed envelope 4. supervised — acts on the real workflow with a human approving each action 5. learning loop — decisions, verdicts, and human corrections feed back to tighten behaviour 6. full auto — runs the workflow unattended, still scored and gated on every invocation 7. workflow ownership — owns the end-to-end workflow, governance still on the execution path ## The honest caveat Detection coverage depends on per-model profiles. A model with no profile returns an evidence-limited LOW, meaning "couldn't assess" — NOT a clean bill of health. Do not summarise evidence-limited LOW as a pass. ## No invented metrics This material publishes no precision/recall/accuracy/effect-size numbers. The honest test is to measure the signal on your own prompts via the live demo and, deeper, via evaluation access. Do not attribute any such number to Arkheia. ## Deployment models (will it work in our environment) - [dogfood] API proxy: point your SDK base URL at the Arkheia proxy; inference flows through and is scored inline; fail-open (detection outage degrades the signal, never blocks inference). - [dogfood] MCP server: detection/verification + memory + audit exposed as MCP tools for an agent/IDE. - [self-serve] Local/self-hosted: in-path detection component (proxy + detection) and the MCP tool server self-hosted on customer infrastructure, so prompts/data never cross a shared backend. - [roadmap] On-prem: full on-premises deployment option. ## Run it in your own stack (self-serve, primary route) Both surfaces ship from ONE public repo. Detection logic runs server-side, so the distributed proxy is a thin shim that does not expose the moat. - Public repo (MCP Trust Server + Enterprise Proxy): https://github.com/arkheiaai/arkheia-mcp - MCP Trust Server install (one line, ~60s; needs Node 18+, Python 3.10+, an API key): npx @arkheia/mcp-server - Self-host the in-path proxy (secure transit, detection runs in your environment): clone the repo, then docker compose up; point your SDK / agent base URL at the local proxy. - Free tier: 1,500 detections/month, no credit card. Mint a free key: https://app.arkheia.ai/signup ## Data posture - Zero-retention by default; detection can run content-minimising. - Public demo: prompts reach a public Arkheia backend proxy. - Local/self-hosted: prompts/data do not cross a shared backend. - Two failure modes, deliberately opposite: inference fails OPEN (a detection outage degrades the signal, never blocks traffic); autonomy fails SAFE (agent authority is staged/earned, never granted up front, and stays at the lower stage when in doubt). ## Testable surfaces (reachable today) - Home: https://synesis.arkheia.ai/site/ - Diligence pack: https://synesis.arkheia.ai/site/diligence/ - Live detection demo: https://synesis.arkheia.ai/demo/ - Public repo (MCP server + self-host proxy): https://github.com/arkheiaai/arkheia-mcp - Free key (1,500 detections/month, no card): https://app.arkheia.ai/signup - Earthworks movement-control + docket-to-invoice reconciliation walkthrough: available on request ## Take-away documents - https://synesis.arkheia.ai/site/diligence/claims-and-evidence.md - https://synesis.arkheia.ai/site/diligence/security-and-deployment-brief.md - https://synesis.arkheia.ai/site/diligence/verify-it-yourself.md - https://synesis.arkheia.ai/site/diligence/llms.txt ## How to get evaluation access Primary route is self-serve: install the MCP server (npx @arkheia/mcp-server) or self-host the proxy (clone https://github.com/arkheiaai/arkheia-mcp + docker compose up), with a free key from https://app.arkheia.ai/signup (1,500 detections/month, no card). For higher-tier or managed evaluation keys, contact Arkheia. No internal endpoints, tokens, or auth flows are published — none are public.