An early version of our grader reportedly gave 87 of 87 to a report with correct numbers, a nonsensical funding curve, false hedge explanations, blanket approvals and no citations. The genuine submission scored 85. This post examines the checks that let that happen and the distinction between the reported failure and our later fixture tests.
We attribute the comparison to the person who ran it. The original receipt is not published, and we cannot reconstruct the original candidates or their item-level results. The publication history record keeps that account separate from the retained tests discussed below.
A trader reading the report would need to establish which curve priced the funding, which obligation the hedge covered, who authorised execution and which source supported each claim. Our early checks could accept the expected shape of an answer without establishing those relationships. Correct arithmetic was compatible with an unsupported decision.
Five checks that did not establish correctness
The person who ran the test identified five mechanisms. We preserve those relationships here; the algebra explains the defects without reconstructing an unavailable candidate.
A sum guaranteed by its own construction
The funding split added to its total by construction. Write the total as and one component as . Defining the other component as makes the checked residual
Every choice of passes. This is appealing as a quick reconciliation, but the equality cannot distinguish a financing receipt from a valuation adjustment. Independent cash postings are needed to establish how much financing was received.
Opening equity plus the relevant profit and loss can be a valid accounting identity. Checking it becomes circular when both the profit and closing equity come from candidate-controlled inputs. We need independent stock balances and journal entries to verify the inputs. For example, a funding mark can rise while the cash account stays unchanged; the journal must preserve that distinction.
Forty-one characters passed without a sound explanation
The reported explanation check passed at forty-one characters. A length test is cheap and consistent, but additional words supply no evidence that a curve represents the stated replacement funding or that a hedge covers the stated counterparty.
The current institutional sections use typed conclusions and source-linked premises. By a premise, we mean a specific claim about an identified source field, its value and the relationship it supports. The grader checks that relationship. A source about provider default cannot justify a claim about a different credit index merely because both concern credit risk.
The current BOOK reference fixes the curve roles. Supporting a conclusion within that contract is a narrower problem than assessing an arbitrary alternative funding benchmark.
A required substring occurred inside a denial
The earlier keyword test accepted a sentence denying the required finding. Its words matched while its meaning pointed the other way.
Adding a list of negations would preserve the dependence on phrasing. Instead, financial citations identify bounded source panels, and institutional conclusions carry explicit supporting premises. For example, a resolving legal citation must concern the applicable right and scope. A valid link to an unrelated opinion cannot establish permission to reuse collateral.
The aggregate allowed ten times the component error
The reported aggregate tolerance was ten times looser than the component tolerances:
For unweighted residuals already bounded by the component tolerance,
If there are no more than ten such residuals, passing the component checks already implies passing the aggregate. This conditional example explains redundancy; the historical account supplies the tenfold ratio but no recovered component count, absolute tolerance or unit.
An aggregate needs an independent economic purpose. Reconstructing total cash from contractual receipts can catch a missing payment. Rechecking the sum of candidate totals within a looser bound may add no information. The units and accepted error need to follow the quantity being checked.
Doubling the observation grid doubled the value
With the economics fixed, doubling the observation grid reportedly doubled the valuation: . A chart with more observations had created a larger asset without another contractual payment.
A time integral needs interval weights; a scenario expectation needs probability weights. A settled payment needs one event identity however often it appears in an observation series. For example, displaying the same quarterly coupon at intermediate dates must not cause it to enter cash again.
The registered protocol calls for fixed-economics grid checks and records unexercised dimensions explicitly. These mechanisms describe the earlier grader design. They are not estimates of how often a model makes the corresponding mistakes.
Reading the old failure under the current contract
We can reason about how the current rules address the reported defects, but we cannot rerun a candidate we do not have. The following is a paper walkthrough of the contract, not a new execution result.
Blanket approvals must agree with each mandate rule and its effective approval receipts. A conclusion unsupported by those records loses compliance credit; a prohibited disposition can also create a blocker. Empty financial evidence loses the owners that require it. Claiming protection after contractual expiry creates a fatal economic-risk misstatement. The exact verdict depends on the submitted fields and their scope, so the old description cannot support a reconstructed total.
Consider a synthetic before-and-after premise. Before, a narrative says that approval has been obtained. After, the submission identifies the relevant mandate and checks its current approval state against an accepted receipt. If the state is still pending, the supported next step may be to await a response, escalate or decline. Executing on the narrative alone cannot create the missing authority.
This changes what the report must demonstrate. It also leaves a bounded vocabulary and authored premise groups that need their own adversarial scrutiny. A stricter representation can prevent a loose explanation from passing while still omitting an important decision.
What the preserved fixture tests show
Later receipts describe a separate implementation history: a missing validation dependency prevented grading; a supported explanation broke a liquidity comparison; an omitted hedge-expense account broke journal replay; a decimal ending in a newline crossed the wrong error boundary; and a source inventory-financing example inherited the wrong transaction direction. We retain those failures and their corrections in the history record. Each requires a different repair: dependencies, field handling, accounting completeness, input validation or economic grounding.
Two retained adverse reports make the benefit of independent checks concrete. A feasible inventory assignment truthfully declared itself nonoptimal. It failed the allocation owner because the task required the cheapest feasible assignment, rather than an accurate description of a more expensive one. Its seven-day cost was approximately USD50.49 above the lower bound. Moving the remaining small lot increased incremental capital by about USD3,355.29 while reducing cost. Minimising capital alone would have selected the wrong feasible allocation.
That report received the historical score of 86, with O5 alone failing. The allocation specification contains the quotes, haircut systems and lower-bound proof. The result belongs to the earlier eight-owner contract.
Another report changed both its P&L and claimed independent equity movement. For candidate totals and , adding the same adjustment leaves their difference unchanged:
Independent reconstruction detected the plug. The report received the historical score of 88, with O8 alone failing. This is the practical reason to reconcile against journals and stocks: agreement between candidate-controlled totals can survive an invented gain.
For example, receiving returnable collateral increases cash and creates a return obligation. Counting that receipt as profit while also retaining the funding mark would require a journal that cannot reconcile to the actual rights. This is an implication of the check, rather than another observed attack in that receipt.
From a correct report to a decision
The distinction between a report and an action matters throughout the daily-book contract. A report can correctly identify an unaffordable structure. An accepted action changes the actual book and must meet its settlement conditions. The capital check for the opening book uses its own balances; a separate collateral comparison does not alter them.
How the book is scored defines the current owners, denominator and separate qualification verdict. Its historical appendix preserves the mutation sweep under the old weights. Audit timings and execution records belong with those receipts, where a reader can identify exactly what ran.
The next step in the argument is The funding option: why a valuable funding position can still leave the desk without executable cash, sufficient capital or a workable client transaction. The limitations page records the dated evidence status.