How Context Debt shows up

Where this came from

The Context Debt Map already tags each of its twelve fragments with one or more of six categories: Missing, Contradictory, Outdated, Unowned, Difficult to retrieve, Unsafe to automate. This page takes five of those and gives each a proper name, a mechanism, and the evidence that was already sitting in the demo. The sixth, Unsafe to automate, isn't a sixth pattern. It doesn't behave like one in the data, and that's addressed at the end rather than forced into a tidy list of six.

Each pattern below corresponds to a specific layer of The Context Stack. These are the same seven layers, viewed through the perspective of failure instead of successful operation.

The five patterns

Authority Collapse

Two sources disagree, and nothing states which one governs. The expense policy says claims over £200 need director sign-off. The finance intranet page, unlinked from anywhere staff would normally look, says £150. Two HR staff give candidates different probation lengths, three months and six, each with full confidence, each having made verbal offers using their own number. A model doesn't hesitate the way a person does when handed two contradictory numbers. It picks one.

Stack layer: Precedence. The rules aren't missing. What's missing is which rule wins.

Confidently Wrong

Information that was true when it was written, presented with no caveat and no owner checking whether it still is. The 2019 staff handbook's remote-work policy hasn't been accurate since 2021. An IT helpdesk macro, last edited in 2022, still sends staff to a supplier the business dropped eighteen months ago. A live chatbot instruction points employees asking about parental leave straight to that same outdated handbook section, today, with full confidence.

Stack layer: History. The fact was correct once. Nobody owns whether it still is.

Nobody's Job

People often assume that someone else is responsible for confirming this step. Three separate teams each believe another team is in charge of verifying that right-to-work documents have been received. If a model is asked "has this been done?", it has no reliable source to check. This is not because the information is missing, but because no single location is accountable for confirming whether the status is current.

Stack layer: Exceptions, in the broader sense the Context Debt article uses: knowledge that lives in an assumption about a person or team, not in a system.

No Record

A decision or current state exists in reality but was never recorded in a searchable system. An office manager schedules bookings to avoid two meeting rooms with broken AV, even though the booking system shows them as available. A manager uses email to approve a temporary four-day work week for an employee, but the change is never updated in the HR database. A supplier is stopped after a dispute over a delivery, yet no written record explains the reason for the change.

Stack layer: Facts or History, depending on whether it's a current state (the rooms) or a past event (the supplier decision) that was never recorded.

One Room Only

Information that exists, is accurate, and is owned, but is reachable in exactly one place. A fire warden rota and evacuation point is pinned to a wall in the Leeds office only. Asked about fire procedure anywhere else, or by anyone not standing in front of that specific wall, a model has nothing to draw on. This is the one pattern where the content itself isn't the problem. The problem is that "written down" and "retrievable when needed" aren't the same thing.

Stack layer: Any layer. This pattern cuts across the Stack rather than sitting in one place; the failure is in reach, not in content.

Not a sixth pattern: Unsafe to automate

It would be tidy to organise this into six patterns instead of five. However, the data does not support that structure. Among the twelve fragments, "Unsafe to automate" appears four times. In every instance, it is paired with one of the other five categories rather than standing alone. It occurs twice with Outdated, once with Difficult to retrieve, and once with Contradictory. This category functions as a severity flag applied to a pattern instead of acting as an independent mechanism. It is useful to identify for that reason: it signals which of the five patterns is too dangerous to use with an AI system before it is fixed. It is not a sixth way that context fails.

What's not here yet

These five patterns come from one internal audit demonstration based on a single fictional company. This is initial evidence rather than a complete library. Developing this resource properly means applying the same tagging process to material from real engagements, rather than creating more examples from Halloway Partners. The next step in research is to test these five patterns, or any new ones that emerge from real audits, against non-fictional material.