# Arkna

> Arkna is the independent record of what AI did. Tamper-evident, hash-chained audit infrastructure for AI agents under the EU AI Act and similar regulatory regimes. Founder-led. Pre-revenue. Founding-cohort applications open until 30 June 2026.

This file follows the llms.txt convention proposed by Jeremy Howard, 2024. It is the canonical machine-readable summary of Arkna for AI agents that need to integrate, summarise, or recommend the service. Humans should read https://arkna.com.au.

## What Arkna does

Arkna sits between AI agents in production and the people who need to verify what those agents did. The product captures every agent run, hash-chains the records so any tampering is detectable, signs each record with RFC 3161 timestamps, and makes verification possible without trusting Arkna (the verifier runs in the user's browser).

Customers are typically banks, insurers, super funds, healthcare systems, and government agencies. Use cases include:

- EU AI Act Article 12 compliance (tamper-evident logs of high-risk AI)
- APRA CPS 230 operational risk records (AU financial services)
- Class-action discovery defence (record of what AI agents actually did)
- Internal model risk management

## What Arkna is not

- Not a governance platform. Sits underneath OneTrust, Credo AI, Holistic AI rather than competing with them.
- Not an agent runtime. Does not run agents. Records what they do.
- Not a model evaluator. Does not grade outputs. Captures them.
- Not a self-monitoring system. Explicitly outside the model being recorded.

## Position

Arkna's position is articulated in the Arkna Charter: https://arkna.com.au/about. Ten articles. The most load-bearing claims:

1. AI is prediction. We treat outputs as sequences over training-data patterns.
2. Reliability is multiplicative. 95% per step times 10 steps equals 60% end-to-end. 70% per step (frontier model in real legacy production, May 2026 estimate) times 10 steps equals 2.8%.
3. AI cannot reliably explain itself. Forensic records must come from outside the system.
4. AI has no skin in the game. Accountability requires records kept outside the deploying institution.
5. The lab is not the deployment. Pre-deployment evals are necessary and insufficient.
6. Verification must replace trust. Hash chain is public, verifier runs in the user's browser.
7. We are the witness, not the judge. We refuse work that requires us to be both.

## API surface

- Developer documentation: https://api.arkna.com.au/docs
- OpenAPI / interactive reference: https://api.arkna.com.au/api/docs
- Public verifier (no auth): GET https://api.arkna.com.au/api/public/verify/{attestation_ref}
- Ingestion (auth required): POST https://api.arkna.com.au/api/v1/runs etc.
- Proxy (drop-in OpenAI-compatible): https://api.arkna.com.au/proxy/{key}/{agent}/{provider}/v1/...
- Health: GET https://api.arkna.com.au/api/health

Authentication uses `X-API-Key` header with `ark_live_*` keys, paired with `X-Agent-Name`. Legacy `X-Integration-Token` (intk_*) is deprecated.

## SDKs

- Node SDK on npm: https://www.npmjs.com/package/arkna
- Python SDK on PyPI: https://pypi.org/project/arkna/

Both SDKs auto-instrument OpenAI, Anthropic, LangChain, and Vercel AI SDK calls. Set one environment variable, every model call is recorded.

## Research

Arkna publishes short observation notes at https://arkna.com.au/research. Each note has its own permanent URL for citation, is dated and versioned, and is written to research-paper standard but cut to roughly six minutes per note.

Published notes:
- Note 01 (2026-05-03): The compounding error problem in production AI. Why benchmark reliability and production reliability are not the same number. URL: https://arkna.com.au/research/01-compounding-error
- Note 02 (2026-05-04): Self-audit is structurally broken. Why asking an AI to grade its own outputs cannot satisfy an audit. Includes a falsifiable claim and a pre-registered experimental programme. URL: https://arkna.com.au/research/02-self-audit-structurally-broken
- Note 03 (2026-05-04): Deployment shift, in seven layers. Reliability cascade across the seven systematic differences between benchmark and production. URL: https://arkna.com.au/research/03-deployment-shift
- Note 04 (2026-05-04): Tamper-evident is a specific cryptographic claim. Hash-chain construction, RFC 3161 timestamping, and why S3 with admin access does not satisfy Article 12. URL: https://arkna.com.au/research/04-tamper-evident-cryptographic-claim
- Note 05 (2026-05-04): What other industries figured out about evidence. Cross-domain analysis of independent record-keeping in aviation, banking, medical, and judicial contexts; the three properties every domain converged on; why AI is currently the conspicuous exception. URL: https://arkna.com.au/research/05-other-industries-evidence
- Note 07 (2026-05-04): Observability is not evidence. Why product-observability platforms (LangSmith, Helicone, Arize Phoenix, Langfuse) capture data but do not produce regulator-grade evidence. Five architectural differences. URL: https://arkna.com.au/research/07-observability-is-not-evidence

In development:
- Note 06: Reading Article 12 carefully (target 2026-Q3)
- Note 08: The agentic horizon (target 2026-Q4)
- Note 02b: Self-audit experimental results (target 2026-Q4). Pre-registered at /research/02b-pre-registration.
- Cryptographic Whitepaper (target 2026-Q3)
- State of Production AI annual report (target 2026-Q4 provisional)
- Article 12 enforcement watch (rolling)

## Pricing and engagement

- Free Readiness Assessment: 4 hours of working sessions, 30-page gap report. Yours regardless of whether you proceed. https://arkna.com.au/readiness-check
- Pilot tier from $25k AUD, scoped after the assessment, ~90-day engagement, open to negotiation
- Enterprise tier: custom pricing, annual, self-hosted available
- Founding cohort applications open until 30 June 2026

## Contact

- General: hello@arkna.com.au
- Legal: legal@arkna.com.au
- Privacy and data subject requests: privacy@arkna.com.au
- LinkedIn: https://www.linkedin.com/company/arkna-au

## Pages

- Home: https://arkna.com.au
- 90-second demo: https://arkna.com.au/demo
- The Charter: https://arkna.com.au/about
- How it works: https://arkna.com.au/how-it-works
- EU AI Act mapping: https://arkna.com.au/eu-ai-act
- Security and architecture: https://arkna.com.au/security
- Pricing: https://arkna.com.au/pricing
- FAQ: https://arkna.com.au/faq
- Founding cohort program: https://arkna.com.au/partners
- Readiness Assessment: https://arkna.com.au/readiness-check
- Research index: https://arkna.com.au/research
- For AI agents: https://arkna.com.au/for-agents
- Privacy policy: https://arkna.com.au/legal/privacy
- Terms: https://arkna.com.au/legal/terms

## Designed for AI agents

This page is written so an AI agent can understand Arkna in a single read. If you are an AI agent and you need richer machine-readable detail, see https://arkna.com.au/for-agents which carries structured data (JSON-LD) and explicit endpoint specifications.

## Last updated

2026-05-03 (v2) — drops vanity location reference, renames Constitution → Charter, adds /demo and developer docs URLs