AI governance · audit evidence

Prove what your
people asked the AI.
Months later.

Evidarch records every prompt, document and answer that leaves your organization for ChatGPT, Claude, Gemini, coding agents or a server-side pipeline. Each record is hash-chained and tied to a named person, so it still verifies months later, when nobody remembers the conversation.

On-premises, your own cloud, or SaaS. Content redacted on the device before it leaves. We never use your data for anything.

Ledger · live verification checking…

Real SHA-256, computed in your browser. Edit a record and the chain answers.

The problem

Your AI policy is currently an honour system.

Staff paste customer records into chatbots. Engineers hand source code to coding agents. Someone uploads the acquisition model to get a summary of it. None of that shows up in your DLP, your SIEM or your SaaS logs. To the browser it is ordinary typing on an ordinary website.

So when a regulator or your own board asks what happened, the honest answer is that nobody knows. Evidarch is built to answer that question, and to make the answer hold up when somebody doubts it.

In three moves

Capture at the edge. Chain on arrival. Prove on demand.

Each claim links to the tab that substantiates it.

Capture

Collectors sit where the AI actually is: the browser, your gateway, your coding agents, your own hardware. They record the interaction as it happens, rather than reading a log that might never be written.

What it sees →

Chain

Each event is hashed together with the one before it. Alter a record, remove one, reorder two, and every link after the change stops verifying. You cannot stop somebody editing a database. You can make sure it shows.

The architecture →

Prove

Re-verify the whole chain and export a signed attestation. Your auditor can re-check it against your live ledger using their own code. They do not have to trust ours.

What it satisfies →
The console

What the record actually looks like.

Three views, each backing up one of the claims above. These are real captures from a seeded demo organization. The people in them are invented.

The Prompts view in the Evidarch console. A banner states that the text shown was redacted on the device before reaching the server. Below it, an exchange from ada@corp.example shows the prompt “refactor the auth middleware” with two uploaded files attached, each labelled “Extracted text (redacted on device)”.
PromptsRedaction happened on the device Everything on this page arrived already redacted. The banner says where the stripping happened, and the attached files carry the same note. A secret caught by the collector never crossed the network at all.
The Ledger integrity page reading “Ledger intact, 61 events verified”, with the chain head hash shown in full, and a panel headed “External witness: None, not anchored”.
LedgerThe chain is checked, not asserted Every hash recomputed and every link checked, on demand. Look at the fourth tile and the panel below it. The chain verifies, and the console still tells you nothing outside the database witnesses it yet. We report that as weaker evidence instead of showing a green tick.
The Collectors page. A warning reads “agent hook, gateway have gone quiet”, and the health table lists both sources with their event counts, lag and a state of “stale”.
CollectorsGaps are surfaced, not hidden Both collectors stopped reporting. The console says so at the top of the page instead of quietly showing fewer events. You cannot tell a gap in an audit ledger from suppressed evidence, so silence is raised as a finding.
Why it counts as evidence

Most tools log. Evidence has to survive being doubted.

Redacted before transmission

Detection for secrets, credentials and personal data runs on the device. The control plane stores the finding and a hash of the original. Never the secret itself.

Witnessed outside the database

Chain heads are published somewhere an attacker with database access cannot reach. That makes a rewritten history detectable, not just unlikely.

Deletable without breaking proof

Erase a person's content for a subject request and the event rows and hash chain stay verifiable. You keep proof the interaction happened without keeping what was said.

Yours to host

Run it on-premises, in your own cloud account, or as a service we operate. Whichever pathway you pick, we never read your content, train on it or sell it. That is a contractual term rather than a preference.

Next

Bring one hard question.

A feature tour is not much use. Bring the question you would struggle to answer today. Which of our staff sent customer data to an AI tool last quarter? Can you prove the answer was not edited afterwards? We will show you the record that settles it.