Library.Reference:9.3.4 - AI acquisition requires a change in available execution
Treat the operative participant as the selected model within an agent system: instructions, context, retrieval, memory, tools, permissions, orchestration and available computation all affect what it can do. The same model name can appear in materially different arrangements. A successful prompt does not qualify every later invocation.
For the present question, distinguish these interventions:
| Intervention | What actually changes | Evidence needed for the receiving use |
|---|---|---|
| Supply or retrieve relevant canon | The system can access the selected explanation or evidence in this execution | It selects and uses the relevant grounds in the actual operation |
| Change instructions or working procedure | The way the system organizes and performs its work | The relevant action changes, including the branch or return previously omitted |
| Preserve and recover task state | The system can recover the goal, decisions, performed work, unresolved questions and side effects after interruption | A resumed execution continues the actual assignment without repeating uncertain effects or silently narrowing the goal |
| Provide an external memory | Information or experience can be selected again under defined conditions | Appropriate retrieval, use, revision and removal; stale or irrelevant memory does not control the work |
| Change tools, access or orchestration | A required operation becomes available, or contributions are combined differently | The operation succeeds under the real interface and permission conditions |
| Change or train a model | A different parameterized system is used | Qualification of the changed system for the relevant work, with the support and limits retained |
These interventions can interact. An instruction may correctly require a source that retrieval never exposes. A memory may retrieve an old decision that was superseded. A stronger model may still act on the wrong objective. A familiar reminder can help one failure while obscuring another. Diagnose the first unsupported contribution from the actual input, actions and result.
SYSE.46 supplies the comparison of unaided, supplied and agent-selected support and the distinction between missing input, failed selection, missed execution and a remaining capability problem. Its related agent-engineering Methods supply the corresponding repairs. KCAE.MEMORY supplies selective, qualified memory use. A.15.10 supplies recovery of interrupted purposeful work. These are substantial professional suppliers, not instructions merely to read a canon.
Attention in this practice means keeping the governing task, applicable Method and current work connected while acting. The practical question is whether the agent performs the required operation under those grounds. If it repeatedly replaces a request for a whole explanation with the latest local correction, restore the original receiving question, inspect what the correction supplies and construct the remaining whole. More context alone may not change that working habit; a procedure or other causal repair may be needed.
After an interruption, recover the actual instruction and consequential updates, what has already happened and what remains uncertain. A summary can help locate the record, but it may omit a correction or the outcome of a side effect. Where an action may already have taken effect, establish its state before repeating it. Recovery is a return to the task and Method, not a compulsory recitation of every historical message.
When a failure recurs, use the actual traces and conditions to compare causes. Did the agent never obtain the instruction, misunderstand its object, fail to call the needed tool, ignore its return, lose the goal, encounter a denied permission or lack the capability to perform the operation? Change the implicated source of the failure. A new warning appended to every output is not a causal repair unless it changes the relevant action under the conditions that produced the failure.
Qualify the repair on the receiving work. Preserve a supported prior success. Use another representative task, interruption, delayed return or changed condition when the next claim requires it. Record whether help was supplied or selected, and whether a person continually supplied the missing judgement. A bounded operational success is useful even when the long-term mechanism remains uncertain.
Do not equate these changes with human learning. Contextual adaptation can be temporary; external memory can persist outside the model; parameter learning has a different change process. Equally, do not assume that AI participants cannot originate variations, coordinate, criticize a canon or maintain conventions. Observe those contributions under their actual conditions and qualify their meaning through C.36.