The engagement is over when they don't need you.

The clearest way to see the boundary is to ask four questions of any engagement and notice who answers them. They are not questions about skill. They are questions about who is holding the outcome when something goes wrong at an inconvenient hour.

Read the table for the last column. Everything before it is a description of work. The last column is a description of ownership, and it is the only one that changes what you build.

Forward-deployed work usually opens with a solution-shaped sentence: add a copilot, automate the queue, build an agent. That sentence is evidence about what one stakeholder currently believes. It is not the mission, and building it directly is the most common way the engagement fails while everybody is being agreeable.

The work starts with reconstruction. Follow real cases from trigger to accepted outcome. Name who touches each one, which system is authoritative at each step, where a person waits, and where work is redone. Sample the ordinary path, the exception, the rejected case, and the one that crosses a team boundary — because judgment lives in the exceptions and the constraints live at the boundaries.

Then build the smallest slice that crosses enough of the real workflow to test the riskiest assumption. Not the smallest feature. The smallest thing that can prove you wrong. A classifier trained on an exported spreadsheet may be a fine model and still teach you nothing about whether the system can read the record it will actually have to read, under the identity it will actually have to use.

There is a specific failure that separates engineers who have deployed inside an enterprise from engineers who have demonstrated inside one, and it arrives in week six rather than week one. The logic is sound, the model behaves, the demo lands — and the whole thing was built on one engineer's broad credentials, reading data copied into an unmanaged context.

Production will not have those credentials. It will have a workload identity with least privilege, a separate read and write scope, short-lived tokens, and a security team that has questions. Redesigning identity and data lineage at that point is not a configuration change; it is often a re-architecture, and it happens exactly when the schedule has no room left.

So the first integrated slice runs under production-shaped identity, even though it is slower and less impressive. The trade is deliberate: you buy a worse demo and a deployment that survives.

If the client owner first appears at handoff, the system was built for dependency, whatever anyone intended. So ownership is named at the start and tested at the end, and the test is a performance rather than a document: the owner operates a normal case, diagnoses a failed one, changes one bounded behavior, reruns the evidence that change invalidated, and rolls the system back — while the engineer who built it sits silently and writes down every place they were tempted to help.

Each intervention is a defect in the package, not a shortfall in the owner. A rollback procedure that has been written down, reviewed, and never executed by anyone but its author is the exact shape of a procedure that fails during the first real incident. You find that out in a quiet room with a stopwatch, or you find it out at 3am with a customer on the line.

This is also the honest answer to the question executives actually ask, which is not whether the model is good. It is what happens to this system when the person who built it moves on. A discipline that cannot answer that has not finished the work — it has finished the interesting part of the work.

The role has a real cost, and it is worth stating plainly rather than selling around. Public accounts note that some engineers consider it undesirable: the travel is significant, and the pressure to solve someone else's problem inside a short window is constant. Both are true. Neither is a misunderstanding to be corrected.

What the complaint gets right is that the feedback loop is unforgiving. You are not shielded by a backlog, a roadmap, or a product manager's framing. The person whose job you changed is in the room, and they will tell you within a week whether you helped. What the complaint misses is that this is also the fastest engineering education available anywhere, because almost nothing you believe survives contact with a real operator narrating why they ignore the field you thought was important.

The role is not for everyone, and the industry's current enthusiasm for it will not make the travel shorter. It suits engineers who would rather be measured on whether the work was adopted than on whether it was delivered.

Everything above is why this platform does not certify reading. Comprehension is not the scarce thing, and it has not been for a while. What is scarce is the ability to convert an ambiguous, consequential workflow into a system that runs under someone else's constraints, prove how it behaves, bound what it may do, and hand it to a named owner who can operate it without you.

So the curriculum is organized around artifacts a reviewer can argue with, evidence that must survive an investigation months later, and a transfer that is observed rather than asserted. The rubric can conclude that a system should not be authorized. The ownership test can fail on a rollback that was never executed. Those outcomes are the instrument working, not the instrument being harsh.

The frontier will keep moving. The models will get better, the tools will churn, and the job titles will multiply. The obligation the role carries — that someone else can run this after you leave — is the part that does not go out of date.