Product launch · Live on every plan · 11 August 2026
Your AI content can now prove a human stood behind it
Every CrawlQ workspace can now turn on a governance layer for AI content: a named person approves every AI output before it publishes, the approval is locked into a tamper-evident record, and the record is anchored to a public transparency log entry that neither you nor we can alter — the precise shape of that claim is spelled out below. This is the long version of how it works, what it deliberately does not do, and why we built it the way we did.
One honest note before the details. No software feature makes a business “compliant”, and we will not pretend otherwise. What CrawlQ records is your evidence of human review — a record, not a certificate. The legal conclusions are yours and your counsel’s; the evidence is ours to provide.
The question your content could not answer
“Did a human check this?” Clients ask it. Platforms ask it. Increasingly, regulators ask it. Until now, most teams could answer only with a policy document: we have a review process. A policy with no per-content record has a practical failure mode — when the question lands about one specific piece of content, you cannot demonstrate the process happened for that piece.
The layer we shipped answers with a record instead: for this exact document, this named person read it and approved it on this date, and here is a proof trail that would visibly break if anyone — including us — tried to rewrite it.
How it works, layer by layer
1. The review gate
With human review enabled, an AI-generated asset cannot be downloaded or published until a named person approves it. An admin or editor opens the review queue, reads the full document — the same bytes that will publish, not a summary — and approves, rejects, or sends it back for revision. Their name and role go on the record. Your own uploads are never blocked; the gate applies to AI output.
2. The fingerprint that keeps everyone honest
Each approval records a SHA-256 fingerprint of the exact content the reviewer saw. This is the detail that separates a real review record from a rubber stamp: an approval never silently survives an edit. If anyone changes the document afterwards, the fingerprint no longer matches, the asset is held again, and the interface says so plainly — “a reviewer approved an earlier version”. The earlier decision is not deleted; it stays in the audit trail as evidence of what was actually reviewed, while the new version waits for its own approval.
Without this, a “reviewed” label would quietly cover content the reviewer never saw — which is worse than no label, because it manufactures false confidence.
3. The public anchor
Approvals are batched into a cryptographic summary that is published to an open transparency log — Sigstore’s Rekor, a Linux Foundation open-source project we use, not a partner — usually within about 30 minutes of approval. Once published, that entry cannot be altered: not by you, and not by us. The log itself is outside our control, which is precisely the point. The strongest evidence is the kind the vendor cannot quietly change.
Two precision points, because sloppy versions of this claim are common: the public log holds the batch summary, and CrawlQ keeps the proof that ties your specific approval to it. And while checking a published log entry needs no CrawlQ account and no CrawlQ involvement, the evidence export that produces your records originates in your own workspace — so the verification step is independent of us; the end-to-end journey is not. We say it exactly that way because an auditor will.
Why there is no backfill button
Your review record starts the day you turn it on — not a day earlier. Everything you published before then stays yours and fully downloadable, and it is labelled honestly: never reviewed, no sign-off record. We built it this way on purpose. A record you could backdate would be worthless to everyone, including you — the entire value of the trail is that nobody, on any side of it, can rewrite history. The practical consequence is simple: the sooner a workspace enables, the longer its provable track record runs, and nobody can catch up by editing the past.
The 30-day cooling-off window
Turning human review off is deliberate too. Within 30 days of enabling, an owner can disable it with a recorded reason. After that, the workspace’s review regime is permanent. This is CrawlQ’s own integrity design, not a legal requirement — a review regime that could be switched off the moment it becomes inconvenient would undermine the record it produces.
Where regulation fits — deliberately last
We built this as governance first; the regulatory fit is the by-product. But the timing is real: since 2 August 2026 — now in force for covered AI systems — “did a human review this?” is a question with legal weight in the EU, and the US FTC, UK, Australia and Singapore are moving in the same direction. The questions arrive on a schedule. Your evidence does not — unless you start it first. For the legal mechanics of the human-review route itself, see Article 50(4): what the human-review route actually requires.
What it deliberately does not do
- It does not certify you, your content, or your business as compliant with anything. A record is not a certificate.
- CrawlQ itself does not add watermarks or machine-marks to your content. Anthropic — the provider behind the model CrawlQ generates through — has announced model-level text watermarks and C2PA file metadata for Claude models. Models launched from August 2026 ship marked; the model CrawlQ currently serves is in the retrofit queue with no confirmed date, and C2PA file metadata may not survive every delivery pipeline. Provider marks say what model touched the content. What CrawlQ records is the layer above — who stood behind it. We do not claim provider marks as our feature, and we do not claim detection until the tooling to verify them actually ships.
- It does not backdate, backfill, or grandfather a sign-off onto content nobody reviewed.
The accountability stack
Provider marks tell you what model generated it. CrawlQ tells you who approved it. Those two statements are not competitors — they are layers of one stack, and they prove different things. AI providers are starting to mark their own output at the model level; the question they cannot answer is the one a client, a platform, or a regulator asks next: fine, a model made it — who signed off?
| Layer | What it proves | Who provides it | Status |
|---|---|---|---|
| Provider watermark / C2PA metadata | What model generated or processed the content | Anthropic & other model providers | Arriving upstream — announced August 2026; older models being retrofitted |
| Tamper-evident approval record | A named person approved this exact content, and the record has not been rewritten | CrawlQ | Live |
| Public anchor (Sigstore Rekor) | The record is independently verifiable and timestamped, on a log neither you nor we can alter | CrawlQ | Live |
| Jurisdiction evidence packs | Which regime applies to a piece of content and what evidence answers it | CrawlQ | In progress |
| Provider-provenance surfacing | Upstream provider marks, detected and displayed alongside the review record | CrawlQ | Planned — activates when provider detection tooling ships |
Read the stack bottom-up. The provider layer says a model touched the content — useful, and not ours; we will not claim it as a CrawlQ feature, and its detection tooling has not shipped yet. Every layer above it is CrawlQ’s: the named-human approval, the tamper-evident record, the public anchor. Provider announcements do the market education — provenance matters — and then hand the accountability question to the layer that can actually answer it.
Turning it on takes about 90 seconds
Human review is opt-in, per workspace, and off by default — nothing changes unless a workspace owner enables it. Owners find it under Workspace → Governance. Enabling is a deliberate three-step flow with a typed confirmation, because a permanent review regime deserves a considered yes. It is included on every plan.
The next time someone asks “did a human check this?”
Have a real answer, on the record. Included on every plan.