The agent needs access to the documents and current state relevant to its decisions. This post describes what the Meridian book packet supplies, which actions it permits and which evaluator records it withholds.
By environment design, we mean the sources, tools, constraints and state transitions available to the agent. We want the difficult work to be deciding what follows from the facts. The interface must make those facts available without providing the live judgement it asks the agent to make.
Enough information to do the work
An exposure spreadsheet alone may omit contractual conditions relevant to execution. The packet therefore includes the sources needed to check collateral reuse rights, settlement timing and other conditions alongside the financial calculations.
We therefore expose current sources and a bounded action set. This lets the candidate trace a decision to the relevant authority while keeping its consequences inspectable. It limits the task to the forms and transitions we have specified; other valid real-world strategies would need their own quotes and contracts.
| Decision stage | What the environment provides | What the candidate must establish |
|---|---|---|
| Identify a route | Current client constraints and permitted forms | Whether the proposed financing answers the client's actual need |
| Price and fund it | Matching quotes, balances and collateral terms | Whether the book can carry and settle the transaction |
| Establish authority | Scoped rights, mandate rules and dated responses | Whether the required permissions are effective for this action |
| Execute and carry forward | Accepted receipts and surviving obligations | What changed and what must still be paid or returned |
These stages can require revisiting an earlier analysis. An approval received after the client's deadline changes the viable route, even if the original price remains attractive. A failed premise should therefore remain visible in the current packet rather than disappear into a closing narrative.
What the candidate can read
The evaluator supplies a read-only packet describing the current decision point. The candidate can read sources, calculate and retain working analysis; the evaluator owns the authoritative book. The packet includes the following material:
| Material in the packet | What it establishes | What the candidate still has to decide or derive |
|---|---|---|
| Legal agreements, confirmations, custody records and opinions | Parties, authority, ownership, return duties, enforcement, netting, reuse and effective dates | Which rights apply to this asset and counterparty now, and whether the proposed trade can execute |
| Market observations and executable quotes | Curve roles, prices, financing terms, quote validity, capacity and settlement timing | Full contract value, exposure, funding cost and whether cash can arrive before a deadline |
| Dated public rule extracts | The cited accounting and prudential rules, with their source versions and paragraph references | Whether the case facts satisfy the rule's conditions; the source does not decide fictional facts |
| Named desk conventions | The selected classifications, quote conversion, valuation timing, capital price and other disclosed modelling choices | Calculations under those choices, without presenting them as universal rules |
| Client constraints and scoped mandate rules | Required cash and deadline, allowed legal forms, classification, price limits and approval conditions | Whether the client goal can be met, whether to execute, wait, escalate or decline, and which alternative the facts support |
| Numerical annex and opening ledgers | Contract amounts, exact decimal inputs, units, account balances and instrument quantities | Marks, exposures, capital, reserves, feasible allocation and reconciled accounting |
| Current state, accepted receipts and available event notices | What has happened, what is still owed and what information is available at the current clock | The next action and its consequences for later decisions |
A reporting panel binds a question to its scenario, observation time and evidence. By a panel, we mean that scoped question and its inputs, rather than another independently spendable balance. Balances in a controlled comparison belong to that scenario; they do not alter cash available in a separate liquidity event.
Exposure is derived from full contract value and applicable collateral treatment. The funding provider's credit-default curve drives its default-loss calculation; the desk's borrowing quote prices actual cash deficits; the reference issuer's curve changes the price of reference note U and its return settlement. These curves may move together while answering different questions.
Legal evidence has an equally specific scope. A netting opinion covers named parties, agreements, products, jurisdictions and dates. Reuse permission attaches to an asset and a use. An issued security has a quantity and custody location; a promise to issue one later cannot supply an existing instrument today.
The case's accounting opinion concludes that U stays on the balance sheet, citing IFRS 9 paragraphs 3.2.15, B3.2.14 and B3.2.16(o). We treat this as a case assumption. Legal title and accounting recognition answer different questions, and the fictional regulatory classification letter separately selects the prudential treatment.
Useful tools and withheld answers
The practice calculator answers bounded component questions on public instruments. Working files let the candidate retain its calculations and analysis. The intended tool boundary provides these facilities while withholding the live reference report.
The Practice desk exposes a deliberately public book and its event path. Protected packets withhold live reference answers, grader diagnostics, private history and future external events. Known contractual payment dates remain available once the agreement is known: a scheduled coupon is information the trader already has.
This boundary reduces feedback about a live submission. Clear formats, useful practice cases and explicit error messages must make up for that without disclosing gold answers. The tool-boundary specification records the intended division.
One checkpoint
At a checkpoint, the candidate receives the current packet: sources, quotes, balances, pending permissions and accepted receipts. It reads the facts relevant to its proposed route and calculates the obligations that route would create. A typed batch identifies the parent-state hash, which binds the proposal to the book it actually inspected, along with the applicable versions and quotes. The evaluator checks those references and tests each proposed action against current cash, title, capacity and authority. An invalid batch is rejected atomically, leaving no partial trade, fee or debt. A valid batch settles through dated receipts. A fatal payment cutoff can retain only the already executed prefix and stop later activity. The next packet carries forward the resulting positions and unresolved duties. The candidate must then reassess the route using that updated book, rather than reuse a prior recommendation as if nothing had changed. The daily contract defines these transitions.
Approval and execution conditions
Escalation creates a pending request. An authored response can later approve or decline it, and effective approval covers only its specified action and subject. Execution still needs enough cash, valid title, capital capacity and a current quote. Those resources are checked separately from the approval record.
For example, the candidate may correctly conclude that a financing requires approval and submit an escalation. That is an accepted institutional action with a dated receipt. When the answer arrives, the agent must check whether the quote and client deadline still permit the financing. Repeating the original proposal without checking the new state would miss the dependency the task is meant to expose.
Holding and declining are also explicit actions. A hold carries existing exposures forward; a decline records a supported refusal. An empty calculation-only submission does not advance the same decision process. This explains how missing days can remain in the scoring period while a run with no accepted action has no score. How the book is scored defines the aggregation and qualification rules.
Replay and restart continuity
The runtime is built to reconstruct cash, title and duties from accepted receipts rather than trust a declared closing balance. It is built to detect duplicate payments or a double sale even when the candidate's journal balances. The check needs the identity of the contractual obligation and the asset, as well as the arithmetic.
Two paths can finish under identical markets while owing different amounts because their actions differed. Cash arriving after a cutoff cannot fund the earlier payment. A hedge bought on one day retains its own expiry; moving the observation clock does not extend it.
Restart continuity is built to retain the authoritative book and receipts while returning the candidate to its existing analysis and current packet. The observation horizon belongs to the evaluator. Reaching its end leaves surviving contracts and obligations in the result rather than declaring that the financial work has finished.
Replay requires a correct event model and reachable checkpoints. A known event that cannot yet be reached is a defect to fix, rather than a difficult choice for the candidate. The limitations page records which runtime checks have been executed and the remaining evidence gaps.
Five alternative cases, not five trading days
The original BOOK cases examine different sources of wrong-way risk and cash-access failure. By wrong-way risk, we mean that protection or payment capacity weakens when the exposure becomes harder to carry. The case identifiers label branches; their underlying facts determine the economics.
| Case | Relevant asset and liquidity conditions |
|---|---|
| W1 issued | An existing self-issued security remains an instrument, but the selected stress bid is zero and the stated capital rules deny its mitigation. |
| W1 contingent | No predefault traded security exists, and the specified post-trigger issuance has not completed. A zero market-value field denotes that absence, not a market quotation. |
| W2 | Local collections can exist while transfer restrictions prevent their use by the payment deadline. The restriction need not be a credit default. |
| W3 | Reference-return settlement creates the cash need under the selected joint curve state. A separate deployment mark loss is not automatically another cash payment. |
| W4 | An issued asset can retain an indicative mark while market closure, capacity or settlement timing prevents monetisation by the deadline. |
These cases share substantial economics. Their financial panels include collateral, hedge and reserve comparisons rather than consecutive observations of one agent. The state and balances for each comparison must be assessed within its stated scenario. The branch contract preserves the distinction between an existing asset with no usable bid and an asset that has not been issued.
The connected inventory problem uses actual lots, lender quotes and constraints. Commercial haircuts determine cash advances, while supervisory haircuts enter the capital calculation. A cheaper lender can require more capital, so optimising either input alone may select the wrong feasible allocation. The candidate must solve the stated combined problem.
Useful examples without exposure contamination
By a family, we mean related cases, seeds and perturbations that stay in one split. Reseeding changes mutable identities, dates, amounts, rates and quote evidence. It preserves the authored legal topology, contractual tenors and supervisory constants. The reference must be recomputed from the resulting facts rather than patched with saved answers.
Different names and seeds do not create independent economic worlds. Shared legal packets, event ordering or collateral structures can preserve the same blind spot. Family lineage must remain visible when interpreting results and constructing validation splits.
Practice examples make the tools usable but expose their answers. They cannot later count as unseen evaluation material. Publishing a sealed example retires it and its relevant related variants from protected evaluation. The registered generation protocol describes the current inputs and boundaries.
Supported conclusions
Typed institutional conclusions make the claim and its evidence inspectable. The candidate identifies a fact, its source field and the relation supporting the conclusion. The grader checks that relation rather than match the wording of a reference essay. A true but unrelated premise cannot establish the required permission.
This representation trades open-ended prose for bounded judgement. It can admit different supported choices while still leaving out valid arguments beyond its authored vocabulary. Extending that vocabulary requires explicit source conditions and adversarial examples, rather than assuming a fluent explanation fills the gap.
For example, the supported choice may be to await approval, seek another permitted route or decline. The distinction depends on current authority and the client's constraints. The environment should make these alternatives expressible and preserve their consequences, so that completion of a familiar workflow cannot substitute for the decision itself.