Keeping the source on every claim
When an answer is built from many places, the risk is that the source behind each claim quietly falls away as findings are tidied into one confident summary. Once it has gone, a checked figure and an offhand guess read exactly the same. The wider set of decisions this sits inside is Ways to make the output trustworthy.
Say it another way
Think of a recommendation reaching you two ways. The waiter who has known you for years and knows your taste says the fish is the one to have. A glossy advert on the television says the same about a meal. Both reach you as “this is good”, but they are not the same signal at all, and if you forget who each one came from you would weigh the advert as heavily as the friend who knows you. Keeping track of who said a thing, and when, is what lets you trust it the right amount.
Doing this with Claude Code
When Claude gathers an answer from several places, ask it to carry the source of each part through to the end rather than dropping it in the summary.
Keep each claim with its source. Have the model attach where a claim came from and when, and carry that through when it summarises, so a checked figure is never flattened into the same plain sentence as a guess.
Treat agreement as strength. When several sources say the same thing, keep all of them rather than reducing to one, because the agreement is part of why the claim can be trusted.
Show conflicts, do not pick. When credible sources disagree, have it surface both with their sources and route the choice to a person, rather than quietly settling on one the AI has no basis to choose.
Ask for dates, and lay it out by confidence. Require a date on each claim so a change over time is not read as a contradiction, and separate the well-supported findings from the contested ones so the layout itself shows what is solid.
Common questions
- What does it mean when several sources say the same thing?
Agreement is itself a signal. When more than one source says it, it is a stronger claim, and the fact that more than one says it has to be kept rather than tidied away, because it is part of why the claim can be trusted. - What should happen when credible sources disagree?
The AI should not pick one, because it does not have the context to resolve which to trust. Quietly choosing manufactures a false certainty and hides that there was ever a question at all. - So what does it do with a conflict instead?
Keep both values with their attribution and source, and send the conflict to a person. That attribution is exactly the context a human needs in order to resolve it. - Why does the source have to survive the summary?
When attribution is lost there is no way to confirm the value of a signal, so every signal becomes equal strength. A checked, authoritative figure and an offhand guess end up reading as the same plain sentence. - Why does each claim need a date on it?
Two figures that differ are not always in conflict. They may be the same number measured at different times, and only the date tells them apart, so a normal change over time is not read as a clash. - Where does the date have to live?
In the structured output, as a required field. Making it required is what makes the check on whether a claim is still current always possible, rather than something that works only when someone remembered. - What does a helper owe when it asserts something?
The attribution, the quality of the source, and when that claim was made. If it is doing research it has to provide all of its attributions, so they can be tested against what the other helpers found. - Should a helper resolve a conflict it runs into?
No. It finishes its analysis and carries the conflict up annotated, as data rather than as a failure. For that to work it has to know its job is to pass the conflict on rather than settle it. - So who does settle it?
The coordinating agent, because it holds the context from all of the helpers. It reconciles before synthesis, or escalates to a person when it cannot. - Does the layout of the answer matter?
Yes. Where a signal has more value it should be presented alongside other signals of the same quality, and the weaker ones grouped together, so the structure itself shows what is solid and what is thin. - Should everything go in the same format?
No. Figures belong in comparable tables where the items compared are in the same unit, narrative belongs in prose so the reasoning connecting it is kept, and technical points belong in lists.
Back to For developers. The craft around it is Working with AI.