agent couldn't follow "as defined in Exhibit C, which incorporates Schedule 2.1(a)" - nested exhibit references three...

TL;DR: Build an explicit reference graph (which document incorporates or defines what) and resolve references transitively, up to a fixed depth, instead of asking the agent to chase them one hop at a time. The agent was resolving the first hop and stopping, so anything defined two or three levels down never got found.

agent couldn't follow "as defined in Exhibit C, which incorporates Schedule 2.1(a)"  -  nested exhibit references three levels deep

1. Confirm the agent stopped early. Ask it to show its work on the failing reference: which documents it opened and where it stopped. Expected: it opened the first exhibit and never followed the incorporation chain deeper.

1. Build the reference graph at ingest. For each document, extract its incorporation and definition references ("Exhibit C incorporates Schedule 2.1(a)", "as defined in Section 4.2") into a small graph: nodes are documents and sections, edges are "defines" or "incorporates". Expected: a queryable map of how the documents link together, built once per deal.

1. Resolve transitively with a depth limit. When the agent hits "as defined in Exhibit C", look up Exhibit C in the graph, follow its incorporation edge to Schedule 2.1(a), and keep going up to three or four hops. Expected: the final definition is retrieved with the full chain recorded (MSA -> Exhibit C -> Schedule 2.1(a)).

1. Detect cycles and dead ends. If the chain loops (A incorporates B incorporates A) or points at a document not in the bundle, stop and report the chain rather than looping or silently giving up. Expected: circular references produce a clear "circular reference detected" message with the loop shown.

1. Make the chain part of the answer. Every extracted definition should cite the full resolution path, not just the final document. Expected: answers read "defined in Schedule 2.1(a), via Exhibit C's incorporation clause" so a human can verify the chain.

Use this when

• the agent fails on references nested two or more levels deep
• exhibits incorporate schedules that incorporate other schedules
• the agent resolves a cross-reference against the wrong document in the chain
• definitions live several hops away from the clause that uses them

Not for this skill when

• the reference is a single hop and the agent still fails (check the section linker or the extraction step)
• the referenced document is missing from the bundle entirely (that is a document-collection gap; report it as missing)
• the agent follows the chain but misreads the final definition (check the answering logic, not the resolution)

Variant phrasings

• agent couldn't resolve a three-level exhibit incorporation chain
• nested cross-reference across exhibits defeated the agent
• "as defined in Exhibit C which incorporates Schedule 2.1(a)" never resolved
• agent stopped at the first hop of a multi-document definition chain

Why it happens

Agents resolve references the way they read: one step at a time, holding the intermediate result in working memory. Each hop adds a chance to lose the thread, misidentify the next document, or simply stop because the immediate question seems answered. An explicit graph removes the memory burden: the chain is computed, not remembered, and every hop is checkable.

Edge cases

• "The Agreement" inside an exhibit may mean the subcontract, not the MSA. The graph needs document-scoped resolution: each exhibit records which agreement its bare references point to.
• Incorporation by reference can pull in an entire external document ("the 2024 framework terms"). If it is not in the bundle, the chain ends with a named-but-missing document; report that explicitly.
• Amendments can rewire the graph (Amendment 3 moves arbitration to a new exhibit). Rebuild or patch the graph per version; a stale graph resolves against superseded structure.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstHHb9KAQzPvZi5ww4EbtHA