Thesis
Markets are the hardest open world. We make them learnable.
Language models now read regulations, filings and news faster than any analyst. But an agent is only as good as the world it learns in, and most market environments quietly break the one rule that matters: nothing from the future may reach the decision.
Aelor turns the market into a world an agent can live in. Every document, price and corporate action enters at the minute it became public; evaluation sets stay sealed; and the same contract carries over from historical replay to the live market.
Hindsight leaks
Revised documents, split-adjusted prices and survivor lists hand a backtest the future without anyone noticing.
Unrepeatable live demos
A live leaderboard shows one path of the market once. It cannot be replayed, audited or trained on.
No learning signal
Forecasts without a verifier are opinions. Reinforcement learning needs rewards that cannot be gamed.
Platform
Four layers, one contract
Each layer is useful on its own; together they close the loop from information to action to learning to capital.
-
01
World model
A point-in-time replay of the market's information state, joined to a live forward archive. Documents, versions and prices appear exactly when they became knowable, never earlier.
- Built
- Federal Register, White House and CPD versions with verified release times · minute bars and corporate actions 2022–2026 · deterministic, content-hashed snapshots
- Next
- Live forward archive (the forward period starts 2026-10-08; capture service in progress) · a reaction track with second-level timestamps for market-moving statements
-
02
Agentic environment
A typed protocol with tools, budgets, episode tokens and append-only ledgers, plus a sandbox that keeps the agent, the world and the verifier in separate trust zones.
- Built
- Protocol v1 · in-process, HTTP and Gym-style interfaces · leakage canaries · agent tokens separated from admin access (container sandbox validated)
- Next
- MCP endpoint for general agents · multi-agent desks
-
03
Learning: RL and test-time scaling
Verifier-scored reward vectors (outcome, format, cost, abstention) feed reinforcement learning, and test-time scaling chooses among an agent's own attempts without touching future data.
- Built
- Reward vectors · batch evaluation with cluster-robust statistics · veRL-aligned agent loop design
- Next
- RL fine-tuning runs on open models · best-of-N and self-consistency at inference
-
04
Trading
From research forecasts to paper execution to capital, with every step gated by sealed tests and a forward record.
- Built
- Sealed 2026 test split (at most two unseals, audit-logged) · forward split · execution-cost scenarios
- Next
- Paper-trading bridge · capital only after a forward record
Positioning
Not another backtest. Not another leaderboard.
| Dimension | Static backtests | Live leaderboards | Aelor |
|---|---|---|---|
| Information timing | Often end-of-day, often revised data | Real time, but a single path | Minute-level point-in-time replay plus the live market, one contract |
| Hindsight control | Survivor lists, adjusted prices | Not applicable | Frozen universe, as-of split adjustment, version-aware documents |
| Evaluation | Tuned on the same history | One window, no held-out set | Sealed test with at most two unseals, plus a forward record |
| Learning signal | None | A rank | Verifier reward vectors, ready for RL |
| Reproducibility | Ad hoc | Not reproducible | Content-hashed snapshots, deterministic tools, audit trail |
One episode
How an agent meets the world
- 1
A policy document becomes public
The clock stops at that minute. Everything later is out of reach.
- 2
The agent researches
It searches and reads documents, prices and entities through point-in-time tools, on a budget.
- 3
It commits to a forecast
A structured call per company: direction, conviction, evidence.
- 4
A sealed verifier scores it
Against what actually happened next, net of the market. Test labels stay locked.
- 5
The reward trains the next agent
The same reward vector drives evaluation, RL and test-time selection.
Evidence
Built, measured, sealed
- 400policy events, 2023 to 2026-09
- 1,161document versions with verified public times
- 194research tasks across 48 AI-infrastructure companies
- 18,654verifier labels; the 2026 test labels are sealed
- 0 / 336leaks found in leakage-canary probes
- 14 / 14automated acceptance checks passing
- 224automated tests
- 2maximum unseals of the 2026 test set
Generated from the system on 2026-10-16. Research preview; the human sign-offs of the first dataset are in progress.
Roadmap
From world to capital
Now
Point-in-time market world and contract · 400 events, 194 tasks · sealed 2026 test · LLM agents running end to end
Next
Live forward archive · RL fine-tuning on open models · test-time scaling · reaction track for market-moving statements
Later
Paper trading on live markets · capital after a forward record · the environment offered to labs and funds that train market agents
Founder
Bangji Yang
Founder & researcher
Reinforcement learning and language-model research, now building the market world for agents that learn.
About the founderTalk to us
We share the private demo, the protocol and the data contract with investors and research partners.
Request access