- In short
- Attribution loss in synthesis is when a final synthesis agent produces claims with no traceable source, and diagnosing it means recognising the cause is upstream context passing that dropped the metadata, not the synthesis agent itself. The fix is to require the earlier subagents to emit structured claim-source mappings so provenance survives the handoff.
What attribution loss in synthesis looks like
Attribution loss in synthesis is the failure where a multi-agent pipeline produces a confident final report whose statements point at nothing. The numbers are there, the assertions read well, but you cannot follow any claim back to the document, URL, or record it came from. The symptom always shows up at the end, in the synthesis agent's output, which is exactly why it is so often misdiagnosed. The place a problem becomes visible is rarely the place it was caused.
Diagnosing this correctly is an evaluate-level skill because it asks you to resist the obvious explanation. The synthesis agent is holding the unsourced claims, so blaming the synthesis agent feels natural. But a synthesis agent can only attribute what reaches it. If the claims arrived already stripped of their provenance, no amount of fixing the final step will restore links that were severed two hops earlier. Judging where the fault truly lives, rather than where it surfaces, is the whole point.
- Attribution loss in synthesis
- A condition where a synthesis agent emits claims with no traceable source because the context passed to it lost the metadata linking each claim to its origin. The defect is upstream, in how context was passed, not in the synthesis agent that merely exposes it.
Why the cause is upstream
To locate the fault, recall what a subagent actually receives. A subagent starts in a fresh context and inherits none of the parent conversation; the only channel into it is the prompt string it is spawned with. That means provenance is not something the pipeline has by default and might lose. It is something each handoff must deliberately carry forward. The moment one subagent returns a finding as flat prose, with the source no longer attached, the link is gone, and every step after that, including synthesis, is working with already-orphaned claims.
So the chain of custody is the thing to inspect. Walk backwards from the synthesis output and ask, at each handoff, did the source metadata travel with the content. The break point is wherever a structured finding became unstructured text. This is the direct consequence of structured context passing: get that right and attribution survives; get it wrong anywhere along the line and synthesis inherits the damage. The synthesis agent is downstream of the real decision.
How to evaluate the candidate fixes
The exam wants you to rank remedies by whether they address the cause. Adding a citation instruction to the synthesis prompt is the tempting non-fix: it tells the final agent to cite sources it does not have, which either produces nothing or invites fabrication. Enlarging the synthesis context window does nothing, because the missing data was never sent. The fix that actually works is structural and upstream: require each subagent to return structured claim-source mappings, pairing every claim with its origin, and pass those mappings into synthesis without flattening them.
Evaluating well also means recognising when attribution loss is not the real story. If the final report is missing whole topics rather than missing sources, you may instead be looking at narrow decomposition failure, where the work was carved up so the coverage gap, not the provenance, is the defect. Part of the diagnostic skill is telling these apart: unsourced-but-complete points to attribution loss; sourced-but-incomplete points elsewhere.
A diagnostic procedure
Treat it like a fault trace, not a guess. Confirm the symptom: are claims genuinely unsourced, as opposed to merely incomplete. Then walk the handoffs from synthesis backward, checking at each one whether content still carried its metadata. Identify the earliest point where provenance disappeared; that is the true defect site. Finally, validate the fix by re-running with structured claim-source mappings enforced at and before that point, and confirm the synthesis output can now cite each claim. The discipline is to fix the earliest break, not the last symptom.
Designing provenance in from the start
The most defensible cure is to make attribution the default rather than something you reconstruct after a failure. That means treating the claim-source mapping as a contract every agent in the pipeline must honour: a worker may only emit a claim if it also emits the source that backs it, and any stage that passes findings onward must carry those pairs forward intact. When provenance is a precondition for producing output at all, there is no later moment at which it can quietly go missing, because an unsourced claim is simply not a valid result.
Designing this way also changes how you read a healthy pipeline. Instead of hoping the sources survive, you can assert at each boundary that every claim still carries its origin, and treat a bare claim as an immediate red flag at the exact hop where it appears. That turns diagnosis from an archaeological dig into a boundary check. It is the same instinct that motivates structured claim-source mappings as a first-class output downstream: provenance is cheapest to keep when it is never allowed to drop, and ruinous to recover once it has.
Telling attribution loss apart from its neighbours
Evaluating well means not over-applying the diagnosis, because several distinct failures can look superficially similar in a final report. Attribution loss is specifically unsourced-but-complete: the right topics are covered, but the claims float free of their origins. If instead whole subjects are missing, the defect is more likely a decomposition problem, where the task was split so that some ground was never assigned, which is the territory of narrow decomposition failure. Conflating the two leads you to fix handoffs when the real gap is in how the work was divided.
A third neighbour is outright fabrication, where a claim has no source because no source ever existed, not because the metadata was dropped in transit. The tell is whether the underlying content is real and merely unlinked, which is attribution loss, or invented, which is a grounding problem the metadata discipline alone will not solve. Sorting these apart, unsourced-but-complete, sourced-but-incomplete, and unsupported-because-invented, is the discriminating judgement that earns full marks at this evaluate level, where naming the symptom is never enough on its own.
How this is tested on the exam
Task statement 1.3 poses this as a root-cause judgement under pressure. A scenario hands you a finished pipeline with unsourced output and several plausible culprits, and asks which to fix. The strongest distractor is always the synthesis agent itself, because that is where the symptom appears. The correct evaluation names the upstream context passing as the cause and structured claim-source mappings as the remedy. Because this is the highest Bloom level in the task statement, partial reasoning is penalised: identifying the symptom is not enough, you must justify why the synthesis agent is not the cause.
A second variant tests whether you can reject a fix that targets the wrong layer. It offers a remedy that hardens only the synthesis prompt, perhaps adding a stern instruction to always cite sources. You should be able to explain that this cannot work when the sources never arrived, and that it risks fabrication. Evaluating the layer a fix operates on, against the layer the defect lives in, is the assessed skill.
Worked example
A research pipeline of three retriever subagents plus one synthesis agent ships a report where every statistic is stated without a citation. A teammate proposes rewriting the synthesis agent to be stricter about sources.
You reproduce the issue and confirm the symptom precisely: the claims are accurate-looking but entirely unsourced, and coverage is fine, every expected topic is present. That rules out a decomposition gap and points squarely at provenance.
Now you walk the handoffs backward. The synthesis agent received three blocks of fluent prose, one per retriever, and faithfully summarised them. Inspecting the retriever outputs, you find the break: each retriever had the source URLs in hand but returned its findings as a narrative paragraph, dropping the URLs. The metadata died at the retriever handoff, two steps before synthesis ever ran. The synthesis agent never had a source to cite.
So you reject the teammate's proposal on principle: hardening the synthesis prompt asks the final agent to cite data it does not possess, which yields either blanks or invented citations. The defensible fix is upstream. You require each retriever to return structured claim-source mappings, every statistic paired with its URL and a short identifier, and you pass those mappings into synthesis unflattened. On the re-run, the report attributes each figure correctly. The lesson you can defend under questioning: the synthesis agent surfaced the loss, but upstream context passing caused it, and that is where the fix belongs.
Common misreadings to avoid
Misconception
If the final report has unsourced claims, the synthesis agent is the thing that is broken.
What's actually true
Misconception
Adding a firm instruction to the synthesis prompt to always cite sources will fix the missing citations.
What's actually true
Why this caps the task statement
Diagnosing attribution loss is the evaluate-level capstone of task statement 1.3 because it forces every earlier idea together: you must know how subagents are spawned, what they inherit, and how context is structured to judge where a real pipeline went wrong. It also reaches forward into Domain 5, where structured claim-source mappings and coverage annotations in synthesis make provenance a first-class output. Master the diagnosis here and you can reason about attribution across an entire multi-agent system, which is exactly the architect judgement the exam is measuring.
The single sentence to carry into the exam room is that a symptom and its cause sit at different ends of the pipeline. Attribution loss is loudest at the synthesis output and quietest at the handoff where it was actually born, and the discipline is to trust the trace over the appearance. Whenever you are handed an unsourced report, resist the pull to fix the agent you can see and instead follow the claims backward until they stop carrying their sources; that earliest break is the defect, and structured claim-source mappings applied there are the cure. An architect who can articulate not just what is wrong but why the obvious fix would fail, and where the real one belongs, is demonstrating the evaluate-level reasoning that this capstone knowledge point exists to measure.
A four-agent research pipeline outputs a report whose every claim lacks a source, though all expected topics are covered. Investigation shows each retriever had source URLs but returned plain prose without them. Which fix is correct and why?
People also ask
Why does my multi-agent output have claims with no sources?
Is the synthesis agent to blame for unsourced claims?
How do you fix attribution loss?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
Anthropic: How to Build Multi Agent Systems
Why watch: Covers Anthropic's separate citation/synthesis pass and how subagent findings must carry source metadata, illustrating where attribution loss originates in context passing.
More videos for this concept
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.