When an AI System Changes What the Task Means

A user asks about one contract. The system retrieves a second contract because it was discussed more recently. The next answer cites the second contract correctly. The user corrects the mistake, but the correction is compressed into a status such as “ambiguity resolved.” Two turns later, the system returns to the wrong document.

The model has not merely forgotten a fact. The operative interpretation of the task has shifted. Hermeneutic drift is a change in what a system takes a task, document, entity, or user statement to mean as context is retrieved, summarized, remembered, and reinterpreted across steps.

The visible wording may stay stable. A phrase such as “the account,” “the contract,” or “the migration” can appear in every turn while the referent behind it moves. The result can be grammatical, well structured, and internally consistent because the system is reasoning coherently from the wrong operative state.

An agent does not act from model weights alone. It acts from assembled runtime state: instructions, retrieved material, memory, summaries, tool definitions, permissions, conversation, and information from prior steps. Each selection and transformation changes what later reasoning can see.

Retrieval mutation occurs when material a system receives no longer carries the meaning of the source it represents. The cause may be truncation, selective quotation, smoothing, reordering, or a dropped qualifier. The retrieved text can still look faithful.

Reduction drift occurs when a richer record becomes a summary, status, score, or memory item and one detail changes the balance of the next decision. A summary is not neutral plumbing when another agent will treat it as authoritative working state.

Recency can make the problem subtle. A newly retrieved passage can be relevant and still be wrong for the current referent. The failure is not that the system saw unrelated text; it is that the text became the answer to a different question.

A correction is useful only when the system preserves enough structure to apply it later. At minimum, a durable correction needs to retain what was interpreted, what that interpretation got wrong, and what should control instead. “User corrected the answer” is too weak: it does not identify the referent, the replacement, or the scope of the change.

The risk grows in multi-agent or long-running work. A later worker may never see the source message; it sees a handoff, a status line, or a compacted memory. If those layers preserve the topic but drop the decisive distinction, the next worker can inherit a confident but damaged premise.

The answer is not to avoid summaries, retrieval, or adaptation. The requirement is context integrity: relevant meanings, qualifiers, and constraints need to survive the transformations that make later work possible.

• Bind consequential references. Record which contract, entity, version, or request a conclusion concerns.
• Keep correction scope. Preserve the mistaken interpretation and its replacement, not only that a correction occurred.
• Return to primary material. Before a consequential action, reopen the source that controls the disputed referent.
• Expose reductions. Mark when a summary, status, or score stands in for a longer record.
• Separate observed state from interpretation. Let a later worker challenge an inherited reading without discarding the record.

These controls do not require a system to remember everything. They require it to retain the distinctions that decide which task it is actually carrying out.

A detected mismatch does not prove that a model is unreliable in general, that a user is inconsistent, or that the final answer is false. It may show a harmless change of scope, an intentionally superseded source, or a correction already applied elsewhere.

Likewise, preserving context integrity does not guarantee truth. A system can faithfully carry forward a mistaken source, a stale policy, or a bad instruction. Integrity preserves what the system had; it does not certify that what it had was correct.

The appropriate response is bounded: make a potential shift visible, inspect the controlling source when the decision matters, and report whether task binding was confirmed, changed, or unresolved. That is more useful than treating fluent continuity as proof that the same task survived intact.