unbrowseIn December 2024 I published a whitepaper with a note to myself in the middle of it.
I published the note too.
"need to remove rewards upon ingestion - people can ingest creative rubbish"
I'm Lewis Tham. These days I run Unbrowse and I'm first author of "Internal APIs Are All You Need". This is the story of the reward I designed before that, why my own margin note was right, and what I'd pay for instead.
alBERT paid you for being different
The whitepaper was called alBERT, and I posted it on X on 1 December 2024. A decentralised answer engine: contributors index web pages, people searching pay to retrieve them, and everyone settles in a token.
The reward was called Proof-of-Creativity. My own line for it: "humans now take over the robot's job."
It worked like this:
alBERT reward, 2024 (as designed)
new content -> embed it
nearest match -> find the closest embedding already indexed
reward -> scored by dissimilarity from that match
far from everything = paid a lot
close to something = paid little
retrievers -> pay in the token, half to the contributor
The idea was that new information is valuable, so pay for new information. Find what traditional crawlers miss.
It sounds reasonable for about a paragraph.
Coherent nonsense is the most novel thing there is
The cheapest way to be far from everything in an index is to say something nobody would say.
Truth repeats. A thousand pages agree that water boils at 100 degrees at sea level, and they all embed close together. A false claim can be false in endless new directions. Every one of them is novel.
I saw part of this coming. The plan was to run NLI scorers over submissions and reject anything "irrational or gibberish."
A gibberish filter catches asdf qwer zxcv.
It does not catch a fluent, grammatical, confident article on the agricultural policy of the moon. Language models write those all day, at whatever volume you pay for. Each one reads fine, passes the filter, and sits far from anything already indexed.
So the highest-paid content in that network would have been the content nobody asked for and nobody should read. And I'd written in the abstract that the goal was to index all perspectives, "including controversial or potentially false information." The reward would have paid the most for exactly that.
The bug was when it paid, as much as what
Read the margin note again. "Remove rewards upon ingestion."
Two mistakes were stacked on top of each other.
It paid at the moment of submission. Before a single person had asked for the content, used it, or checked it. Whatever the network was worth later, the contributor had already been paid.
It paid for a number the contributor controlled. Distance from the index is something you can compute at home before you submit. Anything you can compute at home, you can optimise at home. Give people a score they can check privately and the score becomes the product.
Novelty was never the real target. It was a stand-in for "useful and new," and the useful half was the half nobody measured.
What we'd pay for at Unbrowse instead
Unbrowse maps the APIs websites already run behind their buttons, so agents can call them instead of driving a browser. Routes are the content. A network of people mapping routes would face the exact same temptation: pay per route submitted, and you get a mountain of routes.
Today Unbrowse pays nobody for routes. There are no contributor rewards and there is no token. But if outside people do learn routes for the network one day, the plan flips both mistakes:
alBERT (2024) Unbrowse plan (2026)
paid when on submission on verified use
paid for distance from index a request someone made
checked by a gibberish filter someone else's replay
breaks later keep the reward bonus clawed back
nobody asked for it paid the most paid only if used
In plain terms:
Bounties exist only for real misses. An agent asked for something and no route covered it. That miss becomes the job, weighted by how often it's asked for.
Escrow releases when the requester's own call works. The person who wanted the route is the one who checks it, by using it.
Someone else replays it. The recording is never trusted on its own. Checking is one HTTP call; learning the route is the expensive part. That asymmetry is the whole point, and I wrote it up in Trust the Replay, Not the Recorder.
Any bonus vests. A quarter up front, the rest over 90 days, clawed back if the route breaks or fails an audit. A route that passes once and dies in three weeks loses most of it.
Unrequested indexing is allowed, and paid only when used. Map whatever you like. If nobody ever calls it, it earns nothing. That's the creative rubbish clause, almost two years late.
Every way to farm it pays nothing
Try to farm it.
Fabricate routes for sites nobody asked about. No bounty, no calls, no pay.
Fake a route for a real miss. The requester's call fails. Escrow stays put.
Post a bounty and fill it yourself. A job you pay for carries no bonus from the network, so you've paid your own money to yourself, minus fees.
Ship a route that works once. The bonus is clawed back when it breaks.
None of these need a clever detector. They need the payment to sit downstream of someone else wanting the thing and getting it.
If you design rewards, steal this
Four rules I'd give 2024 me:
- Pay for demand, not supply. Someone other than the contributor must have wanted it.
- Pay after use, not on arrival. Submission is the cheapest moment to fake.
- Let someone else check. Never the person who gets paid.
- Make the money wait. Vesting turns "passed once" into "kept working."
And one smell test: if a contributor can calculate their reward before anyone else touches their work, they will mine the metric instead of doing the work.
Plenty of networks have learned some version of this the hard way. I just had the rare luxury of writing the warning into my own whitepaper and publishing it.
Where this actually is
Live today: learn a site's request once in Unbrowse's cloud browser, then replay it over HTTP. Outcomes count only when verified (an HTTP 200 is not a success). A public registry of scrubbed, read-only routes, each published only after a clean replay with no cookies passes on words it has never seen and fails on nonsense.
Plan, not built: bounties, escrow, the vested bonus and clawback, replay by outside checkers. There is no token. Unbrowse is free during the beta and pays no contributors, so nobody is being paid for anything yet, which at least means nobody is being paid for rubbish.
If you want to see the part that works today, it's one MCP server: unbrowse.ai/mcp. Setup is in the docs.
Don't pay for new. Pay for what someone asked for, once it works.