5.6 Information Provenance & Multi-Source Synthesis

The final lesson of the course closes the loop on a problem that runs through Domain 1 (subagent context passing) and Domain 5 (context management): when an agent combines information from many sources into one synthesized output, how do you keep track of WHERE each claim came from? Task Statement 5.6 is about provenance — preserving the source of every piece of information through synthesis — and handling the conflicts and uncertainty that arise when sources disagree.

The core danger is stated simply: attribution dies during summarization. Each summarization or synthesis step is a chance to lose the link between a claim and its source. A research subagent finds 'AI adoption rose 40% (State of AI 2024, p.12)'; the synthesis agent folds it into a paragraph as 'AI adoption is rising sharply' — and the citation is gone. Now the final report makes claims you can't trace or verify. It's like a game of telephone where the facts survive but the sources evaporate, leaving you unable to check anything. For any report meant to be trustworthy, that's fatal.

This lesson is about deliberately PRESERVING provenance through every synthesis step, and about three specific challenges: combining sources that conflict, handling time-sensitive data, and rendering different kinds of content appropriately. Let's start with the preservation mechanism.

The mechanism for preserving provenance is a structured claim-source mapping: every finding is recorded not as bare text but as a structure that bundles the claim WITH its origin. The exam specifies five fields: the claim itself, the source URL, the document name, the relevant excerpt (the actual passage supporting the claim), and the publication date.

The key requirement is that downstream agents must PRESERVE and MERGE these mappings rather than flattening them into prose. When the synthesis agent combines findings, it keeps each claim attached to its source structure — so the final report can cite every statement. This is the exact same principle as the context-passing rule from Lesson 1.3: pass findings as STRUCTURED data (content + metadata together) so attribution survives every hop. There, it kept citations alive between subagents; here, it keeps them alive through synthesis. Structured-not-prose is the through-line.

Synthesizing many sources inevitably surfaces CONFLICTS — two credible sources give different numbers for the same thing. The instinct is to RESOLVE the conflict: pick one value and present it as the answer. The exam says that's wrong. Arbitrarily selecting one value hides a real disagreement from the reader and may present a contested figure as settled fact.

The correct handling is to ANNOTATE both values with their attribution rather than choosing between them. Present 'Source A reports 12%, Source B reports 8%,' each with its source and date, and let the reader (or a human decision-maker) see the disagreement. At the document-analysis stage, include the conflicting values explicitly and flag the conflict; at synthesis, structure the report to DISTINGUISH well-established findings from contested ones, preserving each source's original characterization. Honesty about disagreement beats false precision — surfacing a conflict is more useful than silently picking a winner.

Two more synthesis challenges. First, TEMPORAL awareness — and this is why publication date is one of the five mapping fields. Two sources giving different figures might not be in CONFLICT at all; they might be describing different points in TIME, making the difference a TREND rather than a contradiction. 'Adoption was 8% in 2023 and 12% in 2024' isn't a disagreement — it's growth. Without the dates, a synthesis agent could misread that as two sources contradicting each other. So require publication/collection dates in structured outputs, so temporal differences are interpreted correctly instead of being flagged as conflicts.

Second, CONTENT-APPROPRIATE rendering. Different kinds of information are best presented in different formats, and forcing everything into one uniform format degrades readability. Financial data belongs in TABLES (so figures align and compare), news belongs in PROSE (narrative flows), technical findings belong in structured LISTS (scannable steps/specs). A synthesis agent that renders everything as, say, uniform paragraphs makes the financial comparison hard to read and the technical specs hard to scan. Match the rendering to the content type.

The 5.6 traps test provenance preservation, conflict handling, temporal interpretation, and rendering. They echo the structured-context-passing of Domain 1.3.

• •Letting attribution die. ✗ Synthesizing claims into prose without source mappings. ✓ Preserve structured claim-source mappings (claim + url + doc + excerpt + date) through synthesis.
• •Picking a winner on conflict. ✗ Choosing the 'most recent' or one credible source's value and presenting it as fact. ✓ Annotate both with attribution; distinguish established from contested.
• •Misreading trends as conflicts. ✗ Treating two different-dated figures as a contradiction. ✓ Use dates — it may be a trend, not a conflict.
• •Uniform rendering. ✗ Converting everything to one format. ✓ Tables for financial, prose for news, lists for technical.

You've reached the final exercise of the course. You now know how to preserve provenance, handle conflicts, reason about time, and render content appropriately. This exercise ties Domain 5 together — and connects back to the multi-agent research system from Domain 1.

That completes Domain 5 — and the entire CCA-F curriculum. Step back and see the whole arc: Domain 1 gave you the agentic foundation (loops, orchestration, hooks, sessions); Domain 2, tool and MCP design; Domain 3, configuring Claude Code; Domain 4, prompt engineering and structured output; and Domain 5, the reliability layer — context, escalation, error propagation, codebase exploration, human review, and now provenance — that makes everything else trustworthy in production. You're ready to architect with Claude.