One of the earliest design decisions we made for Pilgrim was about how to handle the case where two sources in a project disagree. It is a case that comes up constantly in real research, and it is where most AI-assisted research tools either fail silently or actively make the problem worse.
The failure mode of AI synthesis that worries us most is not hallucination, though that is a real issue. It is false consensus: a synthesis that smooths over contradictions between sources by finding language that is technically compatible with both positions, presenting a unified view of what "the sources say" when the sources actually disagree. The researcher reads the synthesis and comes away with the impression that the evidence points in one direction. The evidence does not point in one direction. The synthesis hid the tension.
This is not a hypothetical risk. AI language models are trained to produce fluent, coherent text. When asked to synthesize contradictory positions, a model produces synthesis by default, which means finding the overlap and presenting it as the whole. The positions that did not overlap get dropped or qualified into irrelevance. The output reads as confident. The researcher has less accurate information than they started with.
Why Contradictions Are Valuable
Contradictions between sources are not a problem to resolve. They are information. When two credible sources disagree about a fact, a finding, or a causal claim, that disagreement tells you something: there may be methodological differences worth examining, the phenomenon may behave differently across contexts, the measurement approaches may not be tracking the same thing, or one of the sources may be wrong. All of these are worth knowing.
A researcher who knows their sources disagree on a key point can investigate the disagreement. They can look at how each study was conducted, what population it sampled, when it was done, what it was measuring. They can form a view about which finding is more applicable to their question, or they can accurately characterize the state of evidence as contested. All of this produces better work than a synthesis that concealed the disagreement.
A researcher who does not know their sources disagree, because a synthesis tool smoothed the contradiction, may confidently make a claim that is undermined by material in their own source base. This is an underappreciated risk of delegating synthesis to a tool without understanding what the tool does to contradictions.
The Architecture Problem in Source Comparison
Supporting contradiction detection requires a different architecture than standard summarization. To know that Source A and Source B contradict each other on a specific point, you need three things: the claim from Source A, the claim from Source B, and the ability to compare them at the claim level rather than the document level.
Most tools that process sources do so document by document. Each document is summarized in isolation. Comparing the summaries can reveal large-scale disagreements, but fine-grained contradictions, where two documents make compatible overall arguments but diverge on a specific finding, are invisible at the summary level. The contradiction is only visible if you can compare the specific claims each document makes about the same question.
This requires a claim-level representation of source content: extracting specific, resolvable claims from documents and storing them in a form that can be compared. It also requires knowing which claims to compare, which is determined by the questions the researcher is investigating. Without the research question as an anchor, you cannot know which claims across sources are addressing the same thing and which are talking past each other.
What We Built
Pilgrim's synthesis view is built around a question-and-claim structure rather than a source-and-summary structure. When you define a research question in a project, sources are processed not just for their overall content but for the specific claims they make that bear on that question. Those claims are stored with their source attribution and their polarity relative to the question: supporting, contradicting, qualifying, or addressing a different aspect.
When the system identifies that two sources are making incompatible claims about the same thing, it flags this as a tension rather than choosing one or silently reconciling them. The tension appears in the synthesis view alongside the relevant claims and their sources, so you can see both sides, understand what is being disputed, and decide how to handle it.
The decision about what to do with a contradiction is always left to the researcher. Pilgrim does not resolve contradictions. It surfaces them, provides the claims and sources that constitute the disagreement, and leaves the judgment to the person working on the project. The tool's job is to make contradictions visible, not to collapse them into a position it has no standing to take.
What This Does Not Solve
There are types of disagreement between sources that claim-level detection does not handle well. Sources can contradict each other through framing rather than through explicit incompatible claims: two analyses of the same data can tell very different stories by emphasizing different findings, without either making a false statement. Sources can also disagree through omission: Source A presents a clean finding and Source B, by including additional variables, implicitly complicates it without ever directly contradicting it.
These subtler forms of disagreement require the kind of deep reading and contextual judgment that only a human researcher can provide. A claim-level system flags explicit factual and empirical contradictions well. It is not a substitute for the researcher reading critically and noticing when the argumentative terrain of a source is designed to steer conclusions in a particular direction.
We are also not claiming that surfacing contradictions is the only thing a synthesis tool should do. Contradictions are one category of relational information in a source base. Corroboration, where multiple sources independently arrive at the same finding, is also important and worth surfacing. So is gap: the questions your source base addresses well versus the questions it leaves underspecified. Pilgrim surfaces all three, but contradiction handling is the one we feel most strongly about getting right, because the failure mode of hiding contradictions is more actively harmful than the failure mode of not surfacing corroboration.
A Design Principle
The principle underlying our approach to contradictions is that a research tool should make your source base more legible to you, not less. A tool that produces a synthesis the researcher has to take on trust, without being able to trace it back to specific claims in specific sources, reduces the researcher's understanding of their own material. A tool that shows you what the sources say, where they agree, where they disagree, and what questions are still open gives you a better picture of your source base than you had before, and leaves the argument in your hands.
Good research does not require that all your sources agree. It requires that you understand where they agree and where they do not, and that your argument accounts for both. The synthesis view in Pilgrim is designed to give you that understanding without prematurely resolving the tensions that are actually there.