FAQ

The agent manages the funding decisions available in a synthetic book. It reads the current sources, establishes which routes are feasible and proposes actions on current quotes. Accepted decisions change the cash, rights and obligations it must carry into the next checkpoint.

For example, a financing can have an attractive spread while its required approval is pending. The agent needs to decide whether to wait, escalate, seek another permitted route or decline. More precise pricing alone cannot supply the missing authority. Why a funding trader explains the thesis and the Meridian transaction.

We want to examine decisions whose consequences remain in the next day's book. A static analysis task makes a useful calculation or report easy to inspect, but its final answer can leave settlement and later obligations outside the test. Our design makes accepted actions consequential so that the next decision starts from what the agent actually did.

This choice adds complexity. We must specify dates, rights, cash movements and available alternatives well enough for a practitioner to make progress. The Environment design page links to the decisions behind that boundary.

The registered engineering paths observe twenty daily checkpoints. That is an observation window within longer contracts, rather than twenty years of agent execution. A study fixes its observation and resource limits before it starts.

At each checkpoint, the agent reads the updated packet and submits a report with proposed actions. The evaluator checks the parent book and settles accepted actions through dated receipts. The next packet inherits outstanding obligations and newly available facts. Holding, escalating and declining are explicit decisions; an empty calculation is not a settled day.

The packet contains the current book, accepted history, contracts, opinions, mandate rules, client constraints, quotes and known payment dates. Calculation tools and writable analysis files support the work. Protected packets keep live reference answers and future external events outside that information boundary.

The agent must connect the sources. A collateral agreement can establish a right, while a custody record shows whether the asset is available and a quote establishes the price of using it. The packet table explains those relationships.

The agent would keep its conclusions tied to the current facts, meet the client's specified constraints where a feasible route exists and retain the costs and duties of each accepted action. It would change direction when a quote expires, an approval fails or a cash deadline becomes binding.

A supported refusal can be useful even though the client remains unfunded. The result must retain both facts. How the book is scored explains the difference between checked correctness, qualification and client-outcome credit.

On the Practice desk, you choose the actions. Fixture grading assesses a supplied report. Model execution asks an agent to read the packet and make its own proposals. These paths can share calculations while answering different questions.

Begin with practice to learn the interface, inspect a constructed report in the Book viewer, then use Run it for reproduction and model-worker instructions. The public practice schedule is inspectable, so that example cannot also count as unseen evaluation material.

Read the qualification verdict, failed obligations and supporting evidence together. A total can conceal errors with different consequences. For example, a correct funding mark cannot compensate for a payment that lacks cash at its deadline.

A score is not a decision explains why usefulness depends on the next action and the missing evidence. The dated status of model runs, practitioner studies and calibration is kept on the limitations page.

Yes, where the typed conclusion and its premises support an allowed alternative. A premise identifies a fact in a source and the relationship it supports. The grader checks those facts and their relevance rather than require the wording of a reference essay.

This is a bounded interface. Arbitrary financial prose does not acquire credit through length or vocabulary. A recommendation outside the available action set needs an explicit extension of the contracts and quotes before its execution can be assessed.

A malformed candidate is rejected without a score. Missing candidate evidence loses the relevant credit under the scoring contract. Missing reference material instead prevents grading, because the evaluator cannot establish the expected answer. An absent datum must remain missing rather than becoming zero.

Choose a fixture in the Book viewer and follow a discrepancy to its observation and source fields. The Run register links the full export and command trace. Each observation has its own clock and scenario; separate controlled comparisons cannot be combined into a trading history.

The viewer can load the shared compiled engine and recompute the export locally. This checks consistency with the published artefact. Financial assumptions and judgement about a decision require reading its sources as well.

Funding trader is a Dissei project. We publish the task, its design reasoning and the evidence limits together. The article byline names the publication; article dates record preparation, while execution dates belong to retained run records.

No. There is no bounty, no submission channel and no response commitment yet.