M6-L4.3M6-L4.4Extension: Beyond the Internship Desk
In one sentence: The desk’s later phases add what this version deliberately left out, and its core pattern (a deterministic backend that owns every number, agents that interpret and dissent, publishing checks and scoring against reality) carries over to other high-stakes decision-support problems.
What the PRD defers
Section titled “What the PRD defers”| Later phase | What it would answer |
|---|---|
| Forward trading beyond the internship | Whether the forward results hold over a longer period than three weeks |
| A rules-only comparison portfolio | How much the agents add over the rules alone, which this phase cannot measure |
| Scheduled weekly jobs and automatic usage tracking | Whether the desk can run without a person starting each cycle, and what it costs |
| More stocks and more signals | Whether the method scales beyond the fixed pool |
| Licensed transcripts and analyst estimates | Whether richer evidence improves the specialists’ challenges |
| Stronger isolation for the Risk agent | Whether dissent stays independent under a more determined adversary |
Live trading is explicitly a separate, regulated project and not part of this program.
The rules-only comparison deserves attention. Until it exists, the honest statement is: the published calls are the rules’ calls, made more cautious by the agents, and nobody yet knows whether that caution helps. A good FDE writes that sentence into the results report instead of letting a reader assume the agents added value.
The pattern in other domains
Section titled “The pattern in other domains”Strip out the stocks and the pattern is general. It fits any problem where a decision rests on numbers, the numbers change over time, and the decision can later be judged against an outcome:
| Domain | Deterministic backend owns | Agents interpret | Independent dissent | Scored against |
|---|---|---|---|---|
| Credit review | Financial ratios, policy thresholds | Business context, management quality | A credit-risk reviewer who does not see the analyst’s case | Later defaults and delinquencies |
| Insurance underwriting | Loss history, pricing model outputs | Risk narrative, unusual exposures | A second underwriter on a blind file | Claims experience |
| Procurement | Bid prices, delivery records, contract terms | Supplier fit, qualitative risk | A reviewer arguing for the runner-up | Delivery performance |
| Sales forecasting | Pipeline stages, historical conversion | Deal-specific signals | A sceptic reviewing each committed deal | Actual closed revenue |
In every row the same four questions from this lesson apply: what did we know at the time, where did each number come from, who argued the other side, and how did it turn out?
Typical interview questions
Section titled “Typical interview questions”How would you prove that the agents add value over the rules?
Run a rules-only portfolio alongside the published one, under identical costs and timing, and compare them over the same forward period. Also measure the Risk agent’s catch rate on planted counter-evidence and its false-alarm rate on clean cases. Until then, report the agents’ effect as unknown.
A client wants this desk to place real trades. What do you tell them?
That live trading is a different, regulated project with its own requirements for brokerage integration, controls, compliance and supervision. This desk has no order path by design, and adding one is not an extension of this system but a new one.
Related
Section titled “Related”Syllabus rows: M6-L4.3, M6-L4.4