TAPE
A multi-site editorial engine for controlled autonomous publishing.
Connect a website, understand what it already covers, research a topic, draft through configurable editorial personas, verify every factual claim against its sources, score the result, repair what is weak, and publish only when strict deterministic gates pass.
Private repository. Not public yet. Want a look inside? Drop us a line.
The engine is tested and the console works — the screenshots below come out of it — but there is no public instance and no paid connector: TAPE runs today against sites of our own.
An article that doesn't pass the gates isn't published. That decision belongs to the code, not to the model.
- Status
- In development
- Version
- v0.1.0 · no release yet
- Stack
- TypeScript · PostgreSQL · Redis
- Repository
- Private
Its default answer is no. It is built to prefer no publication over a weak one, updating an excellent page over creating a redundant one, and verified uncertainty over fabricated confidence.
Screenshots of what already works
Screenshots of the real console running an article end to end against its database, its worker and its API. The only substitution is the model, which here is a deterministic test one: that is what makes the same run repeatable.
TAPE
It verifies before it publishes
Every claim is checked against its sources and the result is scored. What does not clear the bar is repaired or does not go out.
It tells you what a run will cost
The composer prices a run before it is queued, stage by stage at the model each role routes to, with ceilings per run, per site and per month.
It honors robots.txt, for real
RFC 9309, longest match. And a site can reserve its rights against text and data mining (W3C TDMRep): that one is excluded from the evidence pool, with the publisher named.
Every article says an AI wrote it
By code and at the top of the page, whatever the model produced. Set against Article 50 of the EU AI Act, including the six things TAPE does not do.
Accessibility asserted, not assumed
We test what a screen reader is handed and what a keyboard can reach. Writing those tests caught four defects axe had passed on every run.
One failure doesn't cost the whole run
If a stage's model can't serve it, it is recovered on a second one, budgeted at its own rate and recorded in the log so the article's provenance says which model wrote it.
Ready to build something big?
Pick what you want to build, add a few extras, and watch your plan take shape in real time. No emails, no waiting. Go ahead — play.