Scored improvement loops for long tasks
Set a goal with a checkable result, a compute budget and a time limit. Your agent proposes, Aftovelti scores every attempt with visible and held-out checks, archives it, and keeps the loop going until the goal is met or a dial runs out. You get the best attempt and a signed certificate.
Sign in
How a run works
- Pick a scorer pack (or bring your own checks) and set the four dials.
- We reserve the budget plus the fee (25% of the compute used, $2.00 minimum per run) from your balance.
- The loop runs: brief, attempt, score, archive, next brief, within the budget and the deadline.
- When it ends you see the held-out score, the gaming gap, the lessons, the best attempt and its certificate. Unused budget is returned.
Scorer packs
| Pack | What it scores | Problems | Typical cost |
|---|---|---|---|
| Trading strategy | Long-short equity strategy scored net of 10 bp costs on Sortino, with a look-ahead test, held-out years, a held-out universe and bootstrap stability. | longshort_synthetic, longshort_v1 | $2.20 to $3.50 |
| Chip block | Digital blocks: cocotb tests on Icarus, Yosys cells and depth, held-out traffic and a second gate mapping. | div32_seq | $3.00 to $6.00 |
| CAD part | CadQuery parts: geometry and envelope checks, mass, and a gmsh + scikit-fem stiffness and strength check. | bracket_v1 | $5.00 to $15.00 |
| Code performance | Code performance against a benchmark: pinned tests and bench, held-out inputs and a held-out benchmark. | near_duplicates | $1.90 to $2.60 |
| Three-statement model | Three-statement model from SEC XBRL facts: historical tie-out, accounting identities, live forecast formulas; hidden checks perturb assumptions and tighten tolerances. | your spec | $1.10 to $1.50 |
| Custom checks (your own spec) | Generic check runner: commands, metrics with targets, file and JSON assertions, hidden checks, pinned harness files. | your spec | $1.00 to $10.00 |