Problem

paper_flow could not produce a Paper end-to-end against a live model — a defect the test suite never caught because it mocks the SDK (scripts/verify.sh stayed green). Running it for real surfaced three failure modes:

1. Strict JSON schema rejected Paper. Agent(output_type=Paper) fails at Runner.run with UserError: Strict JSON schema is enabled… — key-independent. Paper → TreeKnowledge.nodes: dict[UUID, TreeNode] maps to an open-ended additionalProperties object, which strict mode forbids.
2. The model cannot reliably emit UUIDs. With a non-strict wrap, it returned semantic ids ("paper_001", "introduction") for the many UUID-typed fields (id, root_node_id, node_id, the nodes keys, Citation.tree_id/node_id).
3. Even when instructed to emit UUIDs, it corrupts them. Prompt-patching only reduced the errors: the model hallucinated an invalid hex char (…-bdle-… for …-bd1e-…) and reused one UUID for two nodes (a duplicate JSON key that silently drops a node).

Asking an LLM to hand-generate dozens of unique, valid, cross-referenced UUIDs is structurally unreliable.

Fix

Stop the LLM from ever touching an identifier:

• New leaf module quantmind/flows/_paper_draft.py defines an id-less DraftPaper (it is the root node: carries title/summary/content/citations/children directly) with nested DraftNode/DraftCitation.
• assemble_paper(...) is a pure function that walks the draft, mints every uuid4(), wires parent_id/children_ids/root_node_id and each Citation.node_id/tree_id, and builds the real Paper.
• paper_flow sets the Agent's output_type to AgentOutputSchema(DraftPaper, strict_json_schema=False) and derives provenance (SourceRef/as_of/ExtractionRef(flow="paper_flow", model=cfg.model)) in code rather than trusting the LLM.

Design rationale: docs/design/en/2026-07-07-paper-flow-draft-assembly.md.

Verification

• scripts/verify.sh green: ruff, basedpyright, import-linter (apex contract intact), pytest — 247 tests, 89.79% coverage (floor 75%).
• Live end-to-end on a real PDF now returns a valid multi-node Paper with no extra_instructions, code-owned provenance, and no schema/UUID errors.
• A test pins the Agent's output_type to the non-strict DraftPaper schema so this regression cannot silently reappear behind mocks.

Notes

• DoiIdentifier still raises NotImplementedError (OA-PDF resolver is a separate follow-up).
• SourceRef.fetched_at/content_hash intentionally left None in this MVP; documented as follow-ups.

🤖 Generated with Claude Code