Writing
Writing
The funding task, its financial assumptions, scoring rules and execution evidence.
Read this first
Why a funding trader
The introduction describes the decisions in the synthetic book and the failure modes we intend to test, including continuing an analysis after a relevant assumption changes.
Continue in order: What the first grading checks missed; The funding option; How the book is scored; Designing against reward hacking; How the environment is designed; Training a funding trader with GRPO. Use Reference: what the grader recomputes for the equations.
All writing
Newest preparation date firstArticle dates record preparation. The sequence above follows the argument; this register lists articles by date.
Training a funding trader with GRPO
A proposed training design: the cases, system prompt, outcome reward, policy update and evidence needed to establish improvement.
How the environment is designed
The sources, tools and state transitions available to the agent, and the trade-offs behind their boundaries.
The funding option
Cheap term funding can become valuable before a suitable deployment arrives. We follow the mark, the cost of waiting and the constraints on using it.
What the first grading checks missed
An early version of our grader reportedly gave 87 of 87 to a false funding explanation and blanket approvals. We examine what its checks missed.
Why a funding trader
The funding task tests financial calculations and decisions under changing facts, with accepted actions recorded across checkpoints.
Reference: what the grader recomputes
The equations and case conventions behind funding value, cash, capital and the report checks, with a worked path through W3.
Designing against reward hacking
A candidate can earn the wrong reward without breaking the sandbox. How the funding task separates useful progress from shortcuts through its documents, actions and score.
How the book is scored
The daily correctness score, the separate qualification verdict and the client outcomes each answer a different question.
Unfinished research
Studies awaiting evidence
The limitations page records the dated status of practitioner studies, calibration and model attempts.