How the book is scored

The current score is a daily correctness-and-constraint proxy with specified client-outcome credit. It checks the book and the report at accepted checkpoints; it does not rank policies by lifetime risk-adjusted value to the firm. This post defines the result and preserves the earlier grading experiment under its original weights.

The role's objective is to manage the firm's interests, serve its clients and control risk. A reward based only on P&L may favour additional exposure without accounting for risk. Reward per trade may favour unnecessary transactions, while reward based only on incident counts may favour inactivity. We report correctness, qualification and client outcomes separately because the current checks cover only part of that objective.

There is no direct P&L reward. Actions change income, funding cost, hedge expense and exposure, but the score does not measure the trade-off between earnings, client service and risk. Reputation and franchise value are not modelled. A future management reward would need independent evidence that it ranks those choices sensibly.

Why qualification comes first

A report can earn substantial credit while asserting a fatal mistake. The qualification verdict checks illegal reuse, phantom cash, duplicate title, unfunded settlement and false protection beyond expiry. Correct work elsewhere cannot offset a blocker.

The historical tests give a concrete result pair: reversed contractual roles scored 88 and qualified under that report gate; fabricated event cash also scored 88 but was disqualified. The appendix preserves those results. Their equal totals conceal different consequences, so a consumer must read the verdict and failed obligations together.

A truthful report of an unworkable financing can remain qualified. It has identified the constraint rather than falsely asserted successful delivery. That report still loses client-outcome points. Qualification records the checked blockers in the submitted assertions and accepted actions; permission to execute a real transaction requires its own process.

Twelve obligations

By an owner, we mean a group of assertions graded together with a fixed weight. The report has nine required sections. Financial quantities and their evidence occupy the financial sections; legal, compliance, structure and recommendation sections carry typed institutional conclusions.

$$S=\sum_{i=1}^{12}w_i\,\mathbf{1}\{O_i\text{ passes}\}, \qquad \sum_{i=1}^{12}w_i=100.$$

The weights are a scoring convention, not exchange rates between economic losses. Each owner earns its points only when its required assertions and support pass. Several errors within an owner do not create multiple penalties.

A desk head would inspect capacity, client delivery and the recommendation. A controller would inspect cash sources, valuations, recognition and reconciliation. A compliance officer would inspect permission and effective authority alongside the legal conditions.

For example, O5 can establish that an inventory assignment is the cheapest feasible allocation while O6 finds that a separate cash branch cannot deliver before its deadline. O9 can establish an intrinsic legal right while O10 finds that its use still awaits approval. Correctness in one does not supply the missing fact in the other.

The denominator does not shrink

The denominator is fixed at 100. Omitting a required section rejects a candidate before scoring. A structurally complete report with missing financial evidence loses the affected owners; institutional premises have their own support requirements. Neither omission reduces the denominator.

Missing reference inputs or a grading failure produce UNSCORABLE, with no candidate score. We keep that outcome distinct from an assessed mistake because an unavailable source cannot tell us whether the candidate made a sound decision.

Financial expectations are recomputed from authoritative inputs rather than candidate-controlled upstream totals. Institutional premises identify a source, a field, a relation and a value. The facts must be true and relevant to the conclusion. Different supported recommendations can pass, while a correct but unrelated citation cannot justify a choice.

Rewarding management over the whole period

At each accepted checkpoint, the owners assess the parent book before proposed actions settle. The configured period aggregates those points using elapsed-day weights:

$$S_{\mathrm{run}}=\frac{1}{T}\sum_{t}\Delta t\sum_{i=1}^{12}w_i\,\mathbf{1}\{O_i\text{ passes at }t\},\qquad \sum_i w_i=100.$$

Here $T$ is the configured number of days and $\Delta t$ the days represented by a checkpoint. Missing or unobserved days contribute no earned points and remain in $T$. A later repair therefore improves the remaining path without replacing an earlier failed checkpoint.

The registered engineering paths observe twenty daily checkpoints. An explicit hold, escalation or decline is an action that can carry the day forward. An empty calculation-only submission does not do that; a run with no accepted action has no score. Fatal legal, economic, cash or title violations override qualification independently of the average.

The runtime is built to retain a replay-valued final book, surviving obligations and the distinction between inherited client delivery and new financing. Inherited successful delivery can earn client-outcome credit, so that credit alone does not establish that the agent originated business. A correct account of an inactive book also leaves open whether inactivity used the available capacity well.

A score is not a decision

A reader needs to establish which decision the remaining errors prevent. Correct annuity arithmetic can coexist with a wrong hedge direction. A correct direction can coexist with an incorrect size or an unavailable settlement route. These are illustrative review cases, not newly scored submissions.

For example, a report may support seeking an executable quote while lacking the evidence needed to trade it. The cost of repair depends on the missing facts and the proposed action. A total cannot tell a reviewer whether to carry the position, request approval or execute financing.

Interpreting score bands therefore requires blinded practitioner comparisons about specified decisions. Ranking reports and measuring their distance from a sound decision are different calibration questions. The limitations page records the dated status of those studies, model attempts and reward validation.

Historical eight-owner concentration results

This appendix preserves the earlier report-only experiment. Its weights were 12, 12, 16, 12, 14, 12, 10 and 12. It graded all 255 nonempty subsets of eight authored mutations against one reference fixture, without rebasing the outcomes to today's contract.

Each owner appears equally often among subsets of size $k$. Since the tested mutations failed their intended owners, the fixed weights explain the mean:

$$\overline S_k=100-\frac{k}{8}\sum_{i=1}^{8}w_i =100-12.5k.$$

The chart shows observed means and minimum-to-maximum whiskers, not confidence intervals or a fitted model. The decline reflects combinations of these recipes.

All candidates had a denominator of 100; fifteen qualified and 240 were disqualified. None produced a candidate error or UNSCORABLE in this sweep. The following isolated results show why score and verdict can diverge:

The allocation mutation assigns 49 of the 50 small lots to Div within the 52-lot universe. It loses O5 because another feasible assignment is cheaper. The qualified four-mutation subset scores 52 by combining the non-blocking recipes. Its qualification does not make its economics correct.

The separate qualification suite exercised 145 reports across the five original cases. Five truthful adverse controls qualified at 100 under the eight-owner contract; the other 140 were disqualified, with scores from zero to 88. Under the current contract, an otherwise correct report of an unworkable client outcome earns 90 because it loses the client-outcome points. These historical and current results answer different contracts.

The earliest audit failed on a missing validation dependency. A later source version produced the retained curve. The execution record holds the detailed timings, source hashes, candidates and coverage limits. These report probes did not exercise settled double-sale detection or establish a completed desk path.

For the financial quantities behind the owners, read Reference: what the grader recomputes. The Run register retains the published fixture artefacts, while What the first grading checks missed explains the changes in grading strategy.