The pipeline must preserve source evidence while making inconsistencies explicit. Option D records the amount stated on the invoice separately from the total derived from extracted line items. A mismatch then becomes a machine-detectable validation condition rather than an invisible extraction defect.
This approach is superior because it does not silently overwrite source data or ask another model to guess which value is correct. Anthropic's evaluation guidance recommends automated, code- based grading whenever the criterion can be expressed deterministically. Arithmetic reconciliation is precisely such a criterion.
In production, the summation should preferably be calculated by application code using normalized decimal values, even though the option describes the model populating calculated_total. The essential design principle remains the same: preserve stated_total, compute an independent total, compare them, and route discrepancies for adjudication.

Improving error handling is broad and ambiguous, requiring analysis of failure paths, design decisions, and review before implementation. A simple function rename is deterministic and can usually be executed directly.

Persistent behavioral guidance belongs in the system prompt. Defining tone centrally ensures consistent communication across conversations regardless of varying user inputs or retrieved knowledge.

The current context is showing degradation and losing project-specific details. A focused summary preserves the relevant VulkanPipeline and FrameGraph findings, while a fresh sub- agent provides clean context for exploring physics and relating it to rendering.

Grounding responses with trusted enterprise documents helps Claude generate answers supported by authoritative sources. Retrieval significantly reduces unsupported statements while improving factual consistency across knowledge-intensive applications.