Data lab for AI · quant finance & crypto
Verifiable quant tasks
for AI agents.
Expert-built RL environments, eval tasks and benchmarks for quantitative finance and crypto. Synthetic data we generate and own, an automatic grader in every package, and pitfalls taken from our own documented research. We publish what fails.
Ships asHarborTestGenDockerpytest
We also build trading bots and isolated signers on your servers →
- task → seeded data + planted pitfalls
- agent → code, a number and a report
- verifier → re-runs the code on a hidden seed
- reward → 1 only if every check passes
Diagram of how a task is graded, not a live run.
- 10tasks in our benchmark pilot, all auto-graded by code
- 0/10passed by a naive agent that ignores the pitfalls; our reference solution passes 10/10
- 950+pre-registered backtests in our own research
- 2.6%real fill rate we measured, vs 30–38% simulated. We publish what fails.
Services
What we build
The data lab comes first. We also build and verify trading software, delivered on your servers.
01 · Data lab
RL environments, evals and benchmarks
Verifiable tasks for quant finance and crypto: synthetic data, a reference solution and an automatic grader in every package.
02 · Build
Custom trading bots
Binance, Hyperliquid, Solana/Jupiter. Kill-switches, spend caps, live dashboards.
03 · Secure
Isolated signers
Keys never touch the strategy process. Allow-lists, caps, audit logs.
04 · Verify
Honest backtests
Real costs and funding, untouched holdouts, look-ahead tests, random baselines.
How grading works
Graded by code,
not by opinion.
- 01The number must match our reference solution within a tolerance.
- 02Error codes come from a closed list and are scored on precision and recall.
- 03The agent’s code is re-run on a hidden series with another seed.
- 04Cutting off the future must not change past signals. No LLM judge.
Real work
We build it. We measure it.
And we publish when it fails.
Our tasks come from our own research, including the results that did not hold up. That is where the pitfalls come from.
HonestQuant-Bench: 10-task pilot
Ten auto-graded tasks with pitfalls taken from our own failures. Reference solution 10/10, naive baseline 0/10. Not yet published.
Memecoin copy-trading: it fails
12,736 labelled signals: it loses ~29% per trade after costs. We saved you that mistake.
C7: trend bot, run with real money
Daily trend bot on Binance spot with a $100 cap and a −25% kill switch. Ran live from 4 to 13 October 2026, then retired.
Process
How a project runs
- 01
Scope call
30 minutes. What you need, the format, what must not go wrong.
- 02
Fixed proposal
Price, timeline and acceptance tests in writing.
- 03
Build and check
Every task passes its QC, every bot runs in paper mode first. Weekly updates.
- 04
Hand-over
Packages with QC records and licence terms, or a bot deployed on your server with a runbook.
Pricing
Fixed price. Clear scope.
Data-lab prices start here. Every engagement gets a fixed quote in writing after a scope call.
Pilot pack
from $7,500typically $7,500–12,500 · per pack
25 private, calibrated tasks in quant finance or crypto, in Harbor/TestGen format, with a calibration report. Non-exclusive licence.
Custom generator
from $15,000typically $15,000–40,000 · per project
A seeded generator plus 100+ tasks in a domain you choose, such as perps and funding, order-book execution or time-series ML.
Agent evaluation
from $2,000typically $2,000–5,000 · per run
We run held-out tasks from our benchmark against your agent and send you a failure report: which pitfalls it falls into, and where.
Expert hours
from $90typically $90–150 · per hour
Quant review, rubric writing or task authoring by the hour, in your format.
Exclusive licence: the same pack withdrawn from our catalogue, at about 4× the non-exclusive price.
FAQ
Questions
What exactly do you deliver to an AI team?
Task packages in Harbor/TestGen format: instructions, a pinned offline Docker environment, synthetic data, a reference solution and a code-only verifier, plus a calibration report and a QC record for every task.
Where does the data come from?
We generate it. Every dataset is synthetic, built by our own code from hand-written parameters. No exchange or venue market data, no scraped social media, no personal data and nothing made for another client.
How do you stop reward hacking?
The verifier re-runs the agent’s code on a hidden series with another seed, checks that cutting off the future does not change the past, and scores error codes on precision and recall so listing everything does not pay. A null agent and a naive agent must fail every task before we ship it.
Do you sell profitable strategies?
No. We build and verify software. Anyone promising fixed returns is selling you their exit liquidity.
Do you hold my keys or funds?
Never. Everything runs on your infrastructure; we work with keys that cannot withdraw.
What does a bot build cost?
Launch prices for our first 5 clients: Starter $790 (list $1,200), Pro $2,200 (list $3,200), Custom from $4,500 (list from $6,000). Pro includes 30 days of support; the optional Care retainer ($250/month, list $350) starts after hand-over. Every project gets a fixed quote in writing after the scope call.
Need quant tasks your agent cannot fake?
Tell us the domain, the format and the difficulty you need. Leo replies within 24 hours with questions, not a sales script.