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.
A hypothesis is only as good as its weakest step — and a paragraph written by a language model does not tell you which step that is. It reads as one confident argument, whether every link in it was measured or half of them were simply plausible.
The short answer: a hypothesis worth spending lab time on should arrive as an explicit path — this measured effect, from this source, links to this mechanism, which links to this target — with every step open to inspection and the weakest step named. GitMir builds hypotheses that way, outside the language model, and checks them deterministically: the same evidence and the same check give the same verdict every time. That makes three things possible that a model on its own cannot do reliably — checking a hypothesis link by link, choosing the experiment that would settle it, and replaying the reasoning later.
This is a use case of the approach in scientific discovery outside the language model. Its companion piece covers the contradictions no paper states.
Why fluent hypotheses are expensive
Asking a model "what existing drugs might work for condition D?" returns a list with a short rationale for each. Some of the rationales are grounded in real measured effects; some are associations the model has absorbed; a few are confident inventions. All of them are written in the same tone.
A team then has two bad options. Check every claim in every rationale by hand — which is the work the tool was supposed to save — or trust the list and spend months of lab time finding out which entries were real. A large share of published research already fails to reproduce; a 2016 Nature survey found that more than 70% of researchers had tried and failed to reproduce another scientist's experiments. Adding a fluent but unverifiable source of hypotheses to that does not help.
A hypothesis as a path
GitMir represents a hypothesis as a chain of claims, each kept with its source and conditions:
- Intervention I has effect E — measured in two studies, in vivo, at doses comparable to clinical use.
- Effect E reduces process P — reported in one review and one primary study.
- Process P drives condition D — established in several studies across two models.
- Therefore I may help with D.
Laid out this way, the hypothesis is reviewable in minutes. A chemist or a clinician can open each step to its sources and see that the second step rests on thinner evidence than the others. The system says the same thing explicitly: the weakest step is step 2. That is where the first experiment should go.
Because the chain is built from claims the system keeps, it also knows what does not support the hypothesis: a study where E was measured and P did not move, a condition under which the effect reverses. Those appear next to the path, not buried in a reference list.
Choosing the next experiment on purpose
When several hypotheses compete, the useful question is not "which one is most likely" but "which experiment would tell them apart". With hypotheses as explicit paths, that becomes a question the system can answer:
- Hypothesis A and hypothesis B both explain the observed result, but they predict opposite outcomes for marker M under condition X.
- Measuring M under X therefore separates them; most other cheap experiments would be consistent with both.
The output is a ranked list of experiments by how much each would change what is known — with the reason for each written as the hypotheses it would confirm or rule out. A model on its own can suggest experiments; it cannot show that a given experiment discriminates between two specific explicit hypotheses, because it does not hold them as explicit hypotheses.
Deterministic checks and replay
Every check in this process is deterministic. Run the same evidence through the same check and the result is the same, which means:
- A reviewer can rerun it. The reasoning behind a decision to fund an experiment can be reproduced by someone who was not there.
- New evidence shows exactly what it changed. When a new study arrives, the system shows which steps it strengthens or weakens and which hypotheses move — instead of producing a fresh paragraph that differs from the last one for no visible reason.
- Negative results are kept. An experiment that refuted step 2 stays in the structure. The next hypothesis that relies on the same step inherits that knowledge rather than rediscovering it.
This is the property that makes a hypothesis engine usable inside a scientific process rather than beside it.
Use cases
- Drug repurposing. Paths from existing approved compounds to new indications, each link sourced, the weakest link named — a list a medicinal chemist can triage in an afternoon. This is part of the work we describe with the pharmaceutical industry on the GitMir Science page.
- Target prioritisation. Ranking candidate targets for a condition by the strength of the evidence chain behind each, not by how often they are mentioned.
- Mechanism hypotheses for an unexplained result. When an experiment produces something unexpected, assembling the candidate explanations as explicit paths and choosing the experiment that separates them.
- Materials design. Chains from synthesis conditions to structure to property, used to propose untried combinations and the measurement that would confirm them first.
- Internal R&D portfolios. Making explicit which projects rest on the same unreplicated result, so that one replication de-risks several programmes at once.
Status
Scientific discovery is a research direction GitMir Science is actively working on. The use cases describe what the approach is built for and where we are applying it, not discoveries already made; results will be published with method, scope and limitations. If your team decides where to spend lab time and needs the reasoning to be checkable, 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.What your AI actually costs →Your engineers, your loaded cost, your monthly AI bill — counted on your own numbers, not ours.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.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.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
Why not just ask a language model for hypotheses?
It produces plausible rationales in the same confident tone whether a step was measured or invented, so every claim has to be checked by hand. A hypothesis built as an explicit path, each step sourced and the weakest step named, can be reviewed in minutes.
What does "deterministic" mean for a hypothesis check?
The same evidence and the same check always give the same verdict. A reviewer can rerun the reasoning, and new evidence shows exactly which steps and hypotheses it moved.
How does the system choose the next experiment?
With competing hypotheses held explicitly, it finds the measurements on which they predict different outcomes and ranks experiments by how much each would change what is known.
What happens to negative results?
They stay in the structure. A refuted step is remembered, and any later hypothesis that depends on it inherits that knowledge instead of rediscovering it.
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.
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.
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.