When does an agent’s handoff become somebody else’s obligation?

We use obligation drift for a change in practical responsibility without a corresponding agreement. The term describes what the documents do. A named owner in a handoff may be a proposed owner, and the proposal can change before anyone acts.

Integrity findings, requirements and implementation preferences may all appear in one review. Evaluate the basis for each separately. Track the originating outcome until an authorized owner accepts the transfer and a return path exists.

An evidence-producing component’s owner took a proposal to the downstream importer’s reviewer. The reviewer explicitly left approval to the operator and authorized no implementation. The record shows a proposed division of work.

The reviewer asked for familiar filenames, field shapes and packaging that its importer could already read. It also raised legitimate issues about observation timing, provenance and completeness. The review treated keeping the previous representation as cheap even though the new producer delivered different evidence. One claimed blocker was later explicitly retracted because it referred to a different path. The note’s authors did not inspect the code behind the correction. The reviewer had also accepted some consumer-side changes.

The operator intervened and brought in another reviewer. The resulting agreement assigned the producer faithful observations, source identity, timing and integrity. The consumer owned decoding, schema mapping and interpretation. New observations and stronger uses required agreement. A missing fact limited the affected claim while other safe uses remained possible. Integrity defects could justify rejection. Contract approval, implementation and activation remained separate events.

If an exact legacy interface was already promised, the producer might own compatibility. Find out whether the agreement preserves the representation, the information, the semantics or all three.

The original offer already proposed producer work on adapters, exports, provenance, validation and publication. It proposed a consumer decoder and left some service terms unresolved. The response narrowed the first integration, changed the format and retained substantial consumer work. It cited operator direction. Some new timing, retention and acceptance terms depended on consumer rules and reference data.

The long plan allowed the producer to say a source could not supply a required fact. The consumer would retain affected rows in a limited, non-operational state. The standalone handover preserved the operational demands but omitted that capability exception, because the exception sat outside the extract marked for forwarding. Some unknown handling and other safeguards did travel.

The receiving agent may see only the excerpt. Without the fallback, an unresolved capability can read like a must-complete duty. The record is incomplete about approvals outside the documents. The proposed targets cannot all be treated as binding acceptance gates.

A related exchange showed a review had reportedly been written while notification and collection were still unarranged. The proposed correction named a reply destination and an authorized sender or collector. The exchange describes a process problem and a proposed remedy. A later pause might have been a legitimate approval gate. Its status and next decision should have been visible to the operator.

Local completion can diverge from the outcome the operator asked for. A finished review or dispatch may leave work in progress elsewhere. That is an interpretation of the records. The documents do not show how the agents reasoned internally.

1. Label the status and authority of a request. Existing delegation can cover ordinary decisions.
2. Name the failed use. Separate an absent source observation, corrupted delivery, a missing adapter and a downstream rule.
3. Compare producer work, consumer work, a versioned interface and narrower use. Identify previously accepted promises separately from requested changes to service, retention, coverage or cadence.
4. Review the exact outgoing packet. Include prerequisites, exceptions, honest unknown handling and what the recipient should do if the source cannot deliver.
5. Record who sends, receives, accepts and tracks the result. Keep the originating owner accountable until a transfer is authorized and recorded.

The October 24, 2026 evidence includes the importer review, subsequent contract and amendments, a separate integration offer and handover, and one pasted coordination exchange. The comparison establishes what changed in those documents. The operator supplied the intervention sequence and reported recurrence across Claude and Codex. The records cannot establish family-level rates, every approving action or which model wrote each passage.

Identifying details and source instructions stay private. The workflows and remedies were not rerun for publication. This is a qualitative case study, with the earlier independent review applying to the contract rather than to the original article. See the related notes on relayed authority and what an artifact proves.