Experts who drop under 90% on gold tasks get auto-paused.
Same standard we hold the tasks to. Grade the graders.
10
Domain experts who've done the work + graders that can't be gamed.
That's the quadrant labs actually want to buy from.
5
Generating and triaging 20 RL tasks cost us about $0.10.
Making sure they're right is where the work is.
8
ConcreteTakeoff-Bench v0.2 (synthetic quantity sheets):
Claude Haiku 5.5, text only: 38%
With the sheet images: 100%
What you show the model matters as much as which model you pick.
2
37
Applied Will retweeted
We're looking for P.E.s, estimators and construction PMs who want to get paid to build and verify AI training tasks.
Blind solves. Sealed reviews. Real domain work.
DM us.
We're looking for P.E.s, estimators and construction PMs who want to get paid to build and verify AI training tasks.
Blind solves. Sealed reviews. Real domain work.
DM us.
Why construction is our first RL domain:
→ answers you can check
→ mistakes that cost real money
→ labs can't verify it without estimators on staff
Tasks your agents haven't seen. Graders they can't game.
1
57
We planted 3 broken tasks in our pipeline. We caught all 3.
One had a wrong answer key that passed every automated check. Only a blind solve, done without seeing the key, caught it.
2
12
Every Applied Will task ships with a receipt:
✓ reference solution scores 1
✓ exploit attempts score 0
✓ two blind expert solves agree
✓ review sealed before the reveal
13
Construction is our first domain.
Quantity takeoffs have checkable answers and expensive mistakes. That makes them great RL tasks, and hard to verify without someone who has done the work.
1
10
Broken tasks teach models to cheat.
Three expert reviews of one 113-task public benchmark flagged between 23 and 56 broken tasks.
Verification is the product.
1
15
Raw data → task → grader → verified task.
Most data companies stop at step one. We start where they stop.
1
1
33
Every task we accept ships with proof:
→ reference solution scores 1
→ exploit attempts fail
→ two blind expert solves
→ sealed review, hashed before reveal
Grade the graders.
1
12
Why Applied Will exists:
The built world runs on numbers you can check. Takeoffs, load calcs, code requirements.
We turn that work into verified RL tasks, written and checked by people who've done it.
Tasks your agents haven't seen. Graders they can't game.
13






