Chenyang Jiang, Benyang Pan and Puqu Wu submitted Call-Level Revision Contracts for LLM-Authored Workflows to arXiv on 4 November 2026. The paper proposes a contract for repairing workflows written by a large language model (LLM), addressing when completed calls can be reused after orchestration code changes.[1]

The abstract says durable execution engines replay recorded history without comparing activity payloads. Its proposed contract gives each capability call a structural identity combining its callsite, fanout position and occurrence count. Committed call expressions become frozen while surrounding code remains editable. A static gate rejects changes to committed calls before a container starts; a runtime digest stops changed payloads at committed identities; an occurrence ledger identifies repeated paid payloads.[1]

The authors claim that, under three assumptions, an accepted revision replays every committed call it invokes, executes only new work and reports skipped identities. The saved abstract does not enumerate those assumptions. It also describes a limit: local identity cannot both survive relocation and avoid incorrect reuse, so relocating a call triggers a stop before spending.[1]

The abstract reports that 22 gate-passed media-workflow repairs replayed 258 committed identities without duplicates, and that a live six-way fanout made seven paid calls instead of twelve. It also describes a 12-task customer-support study using two models. These are author-reported results.[1]

[1]: Call-Level Revision Contracts for LLM-Authored Workflows, arXiv, submitted 4 November 2026.

Sources

  1. Call-Level Revision Contracts for LLM-Authored WorkflowsarXiv · 4 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

026199a6906ccf369dccc8924815e7d1ac1ec4ba86bfaec98d56a9d8546632a0