A team spends months teaching an assistant the rules of a business. Then the project rotates, the lead changes roles, or the platform team swaps the model. The system still answers, with full confidence, but the reasons behind its instructions are gone. Nobody can say which rule was a regulatory requirement and which was a preference.
Context Collapse makes the loss measurable. The Handoff Cliff names the moment context stops transferring. Context Half-Life names how fast decision relevant context becomes unrecoverable. The Context Custody Architecture captures context at each boundary, gates the handoff, and tests whether a successor can actually recover it.
Executives can see which AI systems depend on context that only one person or one session holds. Reassignments stop being silent risk events. Model swaps stop invalidating months of tuning without anyone noticing. The organization keeps what it paid to teach its systems.
Every enterprise AI system runs on two kinds of knowledge. One is written down in prompts, policies, and configuration. The other is held by the people and sessions that shaped it, and covers why a threshold was set, which customer complaint produced a guardrail, and which instruction exists because a regulator once asked. The first kind survives a reorganization. The second kind does not.
Context Collapse is the loss of the institutional context an AI system depends on, at a boundary where that context fails to transfer. The phrase has an earlier meaning in social media research, where it describes many audiences flattened into one. This brief uses it in a different and enterprise specific sense. Two further constructs describe where the loss happens and how fast.
The boundary event at which context stops transferring. A session ends, a project changes hands, an owner leaves, a prompt is copied into a new workflow, or a model version is replaced. Before the cliff the system and its steward share an understanding. After it the system carries instructions that nobody can explain. Every Handoff Cliff is a point where Context Collapse can begin.
Test. For your five most used AI systems, can the current owner explain why each instruction exists without asking the previous owner?The time after which half of the decision relevant context behind an AI system can no longer be recovered from any record or person still in the organization. It is measured per system, by asking successors to reconstruct a sample of past decisions. A short half-life means the system is being run on inherited instructions that nobody can defend.
Test. Pick one rule in your most important AI system. Who can say why it exists, and how long ago did they last touch it?Technical documentation and record keeping for high-risk AI systems are real obligations. What they leave to the implementing organization is the transfer protocol: the mechanism for moving intent, not just files, when a system changes hands or a model is replaced. That is where context is lost, and where this brief picks up.
Most assistants start each session without what the last one learned. Corrections given in conversation never reach the configuration, and long inputs can bury the details that matter.
Resolution. Promote every durable correction into the registerThe person who tuned the system holds the rationale. When they change roles, the instructions stay and the reasons go. Exit checklists cover laptops and badges, not intent.
Resolution. Make an intent handoff part of every role changeA working prompt is copied into a new tool and edited. The copy keeps the words and loses the origin. Within a few forks nobody knows which version holds the compliance wording.
Resolution. Give every prompt a source reference and a purpose lineA model is upgraded or replaced and the old instructions are carried over. Behavior tuned to the old model shifts, and no test exists for the context the old model had absorbed.
Resolution. Treat a model change as a Handoff CliffEach cause is ordinary. People change roles. Teams reuse what works. Platforms upgrade models for good reasons. Conversations are faster than documentation. Context Collapse is not caused by negligence. It is what happens when the cost of writing down intent is paid by one team and the cost of losing it is paid by another, later.
The Context Custody Architecture is a four component pattern that moves context from people and sessions into a record the organization owns. It does not ask teams to document everything. It asks them to capture the context that a successor would need in order to defend the system's behavior, at the moments when that context is most at risk.
The Handoff Gate triggers at each Handoff Cliff and holds the change until capture is complete. The Context Register stores the purpose, source, and owner of each instruction and data source. The Decision Ledger records why each material choice was made. The Recovery Test asks a successor to reconstruct a sample of decisions, and its pass rate is the measured Context Half-Life.
The Context Register needs three fields, not a template. For each instruction, data source, and exception, record the purpose in one sentence, the origin, and a named owner. A rule with no purpose line is flagged as inherited, and inherited rules are the first candidates for review.
The Recovery Test is the component that makes the rest honest. A successor is handed ten past decisions and asked to explain each from the record alone. Decisions they cannot explain count against the system, and the share they can explain is the Context Half-Life signal. It is a test of the record, not of the person.
Start with one system. Choose the AI system whose behavior the business would miss most, and ask its owner to explain ten of its rules. The rules they cannot explain are the first entries in the Context Register. The gate, the ledger, and the test follow. None of it needs new tooling.
Three roles carry the architecture. An AI System Owner keeps the Context Register current. A Platform Change Manager runs the Handoff Gate for model swaps and project moves. A People Operations Partner adds the intent handoff to role changes. Each maps onto an existing function, and the sponsor is usually the CIO or the Chief AI Officer.
The most common failure is to write the register and skip the test. A register nobody has tried to read looks complete until the day a successor needs it. The Recovery Test is the control that keeps the other three honest.
Before the next owner change or model swap on a critical AI system, the sponsor must be able to answer three questions. Where is the purpose of each instruction recorded. Who can explain the last five material decisions from the record alone. What will be replayed to show the new model or owner behaves as the old one did. If any answer is unknown, the handoff is a Handoff Cliff with no rail.
Context Collapse (in its enterprise AI sense), Handoff Cliff, Context Half-Life, and the Context Custody Architecture are original terms coined in this brief by Arjun Jaggi. Academic citation is welcome with attribution. Commercial use requires written permission.