The Contradictions No Paper States
The most valuable finding in a field is often a contradiction no paper states. Why summaries hide it, and how holding every claim with its conditions brings it out.
The most valuable finding in a field is often a contradiction that no paper states. Two results that cannot both be true sit in different journals, years apart, written in different vocabularies — and because neither paper contradicts anyone, nobody notices.
The short answer: finding those contradictions — and the hidden links that are their mirror image — requires holding every claim in a field at once, with its conditions and its source, and checking them against each other. A language model reads a handful of retrieved passages and writes a fluent synthesis, which tends to smooth disagreements over rather than surface them. GitMir keeps each claim separately, with what it was measured under and where it came from, and checks them against each other deterministically. The output is a list of specific disagreements and specific unexplored connections, each traceable to the sources on both sides.
This article is a use case of the approach described in scientific discovery outside the language model: what the problem looks like, why summarisation hides it, and what a useful result looks like.
The problem: knowledge that exists only between papers
In 1986 the information scientist Don Swanson noticed that one body of literature reported that fish oil lowers blood viscosity and platelet aggregation, and a separate body of literature described Raynaud's syndrome as involving high blood viscosity and platelet aggregation. No paper connected the two. The connection — that fish oil might help Raynaud's — existed only between the literatures, and was later supported by clinical work. Swanson called this "undiscovered public knowledge", and the field it started is known as literature-based discovery.
The same structure produces the opposite case. Study A reports that a compound raises a marker under one assay; study B reports that it lowers the same marker under another. Each paper is internally consistent. The disagreement — which might be a measurement artefact, a dose effect or a genuinely different mechanism — exists only between them, and it matters to anyone building on either result.
Both cases share a property: they cannot be seen by reading papers one at a time, and they get harder to see as a field grows.
Why a synthesis hides them
The natural way to use a language model here is to retrieve the relevant papers and ask for a synthesis. That fails in a predictable way:
- Retrieval decides what is seen. The model sees the passages that ranked highest for the question. The paper on the other side of a contradiction often uses different terms and does not rank at all.
- Fluency averages. Asked to summarise disagreeing sources, a model produces a reasonable-sounding middle — "results have been mixed" — which is exactly the information that should have been made specific.
- Conditions get lost. Whether two results actually conflict depends on dose, species, assay and timing. In a paragraph those conditions are the first thing to go.
- Nothing is kept. Each question starts again. A contradiction noticed once is not there the next time someone asks.
The result reads well and hides the most useful thing in the field.
What the approach does instead
GitMir treats the literature as a set of claims to be kept and compared, not as text to be summarised:
- Each claim is held separately — the finding, the direction of the effect, the conditions it was measured under, and the exact source.
- Terms are aligned by meaning, so that a marker, a gene or a compound named three different ways in three subfields is recognised as the same thing — and two things that share a name but not a meaning are kept apart.
- Claims are compared systematically. Two claims about the same thing, under comparable conditions, in opposite directions, are a contradiction. Two chains of claims that meet at a shared intermediate are a candidate connection.
- Every result carries both sides. A contradiction is shown with the sources and conditions on each side; a connection is shown as the path of claims that produce it.
- The check is deterministic. Run again on the same sources, it returns the same list. Run after new papers are added, it shows exactly what changed.
The language model is used where it is strong — reading the unstructured text of a paper and proposing the claims in it — and the structure does the comparison it cannot.
What a useful result looks like
The output is not an essay. It is a short list a researcher can act on:
- Contradiction. "Compound C raises marker M (source 1, in vitro, high dose) — compound C lowers marker M (source 2, in vivo, low dose). Conditions differ in dose and model; no source tests both." The action is obvious: the experiment that would resolve it.
- Unexplored connection. "Effect E of intervention I (sources 3, 4) — E is reported as a driver of condition D (sources 5, 6) — no source links I to D." The action: check whether the path holds before anyone spends on it.
- Weak foundation. "Twelve later studies rely on result R; R comes from one source and was never replicated." The action: replicate R before building more on it.
Each item can be opened to its sources. Each can be checked by a person in minutes, because the reasoning is a short path rather than a page of prose. More on that in hypotheses you can check step by step.
Where this use case applies
- Drug discovery and repurposing — connections between mechanisms, targets and existing compounds across disjoint literatures.
- Toxicology and safety — conflicting findings about the same compound that matter before a decision, not after.
- Systematic reviews — a structured map of where studies agree, where they disagree and under which conditions, instead of a narrative "mixed results".
- Materials and chemistry — reported properties of the same material that disagree because synthesis conditions differ.
- Any R&D organisation with its own archive — internal reports and experiment logs that contradict each other across teams and years. This is the same problem at a smaller scale, and often the most valuable place to start.
Status
Scientific discovery is a research direction GitMir Science is actively working on. This article describes the use case and how the approach addresses it; it is not a report of findings already made. When results exist they will be published with method, scope and limitations. If your team has a literature or an archive where this would matter, write to GitMir Science.
More on this, with the numbers behind it
What business logic is, and where it lives →Features, rules, customer paths and how they connect — why nothing a team already owns holds them in one place, and what that costs.Ask a published model your own question →A real product, read in full. Ask it what a change reaches and check the answer against the source it came from.GitMir Intelligence →Ask what a change reaches, what a rule holds and which customer path meets it — with the part of the product each answer came from.Put a number on your own workflow
Bring us the recurring workflow where AI still needs expensive people to supervise, review and correct. We baseline what it costs and put the target in writing before we build. Or connect a repository and see it on your own code first — 10,000 credits free, no card.
Build My Business Case → Start free →Frequently asked questions
What is literature-based discovery?
Finding knowledge that exists only between publications — a connection or a contradiction no single paper states. The classic example is Don Swanson's 1986 link between fish oil and Raynaud's syndrome, found by connecting two literatures that never cited each other.
Why can't I just ask a language model to find contradictions in the literature?
It sees only the passages retrieval returns, tends to average disagreements into "mixed results", loses the conditions that decide whether two results really conflict, and keeps nothing between questions. Contradictions need every claim held with its conditions and compared systematically.
How do you know two claims really contradict each other?
They must be about the same thing under comparable conditions and point in opposite directions. Every reported contradiction shows the sources and conditions on both sides, so a researcher can check it directly.
Does this work on a company's internal reports, not just published papers?
Yes. Internal reports, lab notebooks and experiment logs have the same problem — results that disagree across teams and years — and are often the most valuable place to start.
Related articles
Scientific Discovery Outside the Language Model
A language model can write a convincing paragraph about a compound it has never reasoned about. Discovery needs reasoning that persists, can be checked and improves with every experiment.
Hypotheses You Can Check Step by Step
A hypothesis is only as good as its weakest step. Built as an explicit path, every link sourced, it can be reviewed in minutes, and the next experiment chosen on purpose.
Autonomous R&D That Remembers What It Learned
A research loop that starts every cycle from a blank state is automation, not research. What makes R&D autonomous, why agents alone fall short, and where it applies.