Try to kill your hypothesis before reality does.
An adversarial research pre-mortem for assumptions, confounders, competing explanations, falsification tests, and evidence boundaries.
FALSIFY by Kalaris Labs
Most AI research tools help you search, summarize, write, or find supporting evidence.
FALSIFY does something different.
You give it a hypothesis and it asks:
What would have to be true for this idea to be wrong?
It turns that hypothesis into a structured research pre-mortem before time, money, and experimental effort are committed.
The goal is not to be negative for the sake of it. The goal is to make weak assumptions, confounders, competing explanations, and unclear decision rules visible early enough to do something about them.
FALSIFY can break a research idea into:
- Claim decomposition - smaller claims that can actually be tested
- Hidden assumptions - things the hypothesis quietly depends on
- Competing explanations - alternative mechanisms that could explain the same observation
- Confounders - variables that may distort the apparent relationship
- Missing controls - controls needed before results become interpretable
- Evidence gaps - what is still unknown
- Falsification tests - tests designed to weaken the hypothesis, not merely confirm it
- Most dangerous failure mode - the failure mode most likely to invalidate the reasoning
- Minimum next experiment - the smallest useful next step
- Evidence boundaries - what should increase confidence, weaken confidence, or force revision
The output is not another chat transcript.
It becomes a structured research artifact.
HYPOTHESIS
|
v
DECOMPOSE
|
v
ATTACK
|
v
TEST
|
v
DECIDE
In practice:
Research idea
|
v
Atomic claims
|
v
Assumptions + confounders
|
v
Competing explanations
|
v
Falsification tests
|
v
Minimum next experiment
|
v
Decision boundary
Each investigation can be turned into a Falsification Passport containing the key reasoning from the analysis, including:
- hypothesis
- research status
- executive verdict
- critical unknown
- strongest alternative explanation
- most dangerous failure mode
- minimum next experiment
- evidence boundaries
- key limitations
Investigations can also be exported as:
- Markdown
- JSON
- printable report / PDF via browser print
A higher ratio of near-infrared reflectance to visible green reflectance can act as a low-cost proxy for algal bloom intensity in freshwater.
Instead of simply supporting the claim, FALSIFY may challenge questions such as:
- Is the apparent NIR signal actually being driven by turbidity?
- Does illumination geometry change the ratio?
- Is camera exposure introducing bias?
- Does NIR add useful information beyond visible bands?
- What should count as ground truth?
- What observation should make the researcher revise the hypothesis?
The goal is not pessimism.
The goal is a better experiment.
FALSIFY is not designed as:
- a generic chatbot
- a PDF summarizer
- a citation generator
- a literature dump
- a replacement for peer review
- a replacement for experimentation
- an AI system that agrees with every premise
It is a reasoning layer for the stage before experimentation.
Search tools help you find information.
Writing tools help you explain an idea.
FALSIFY helps you decide whether the idea deserves the next experiment.
FALSIFY can classify an investigation into one of four analysis states:
| Status | Meaning |
|---|---|
UNDER_SPECIFIED |
The hypothesis is not yet precise enough to falsify properly |
TESTABLE |
The claim is sufficiently defined for useful testing |
WEAKLY_SUPPORTED |
Some evidence exists, but important uncertainty remains |
EVIDENCE_BACKED |
The supplied evidence survives the current pre-mortem |
These labels describe the state of the analysis, not scientific truth.
FALSIFY uses OpenRouter for model inference.
The product is designed to work with OpenRouter's free-model routing so the analysis can run without requiring a paid model path.
The system expects structured output and renders that analysis into the Research War Room and Falsification Passport.
FALSIFY keeps investigations in the browser for a lightweight, account-free workflow.
No account is required for the core experience.
The application can preserve saved investigations locally and reopen them later.
FALSIFY is a reasoning aid, not a scientific authority.
It does not replace:
- domain experts
- literature review
- peer review
- ethics review
- laboratory validation
- statistical analysis
- real-world experimentation
The system is designed to avoid inventing citations, measurements, experimental results, or unsupported numerical confidence.
A strong FALSIFY output should distinguish between:
- evidence provided by the user
- model inference
- assumptions
- uncertainty
- proposed tests
FALSIFY can be useful for:
- students
- independent researchers
- research fellows
- labs
- early R&D teams
- science competitions
- experimental design
- technical due diligence
- research planning
Anywhere somebody is about to say:
"I think this is true."
FALSIFY asks:
"What would prove you wrong?"
FALSIFY was built as a rapid prototype for the NativeBuilder: Build Without Limits hackathon.
The build constraint shaped the product around a deliberately narrow goal:
One hypothesis in. One useful research decision surface out.
The project uses Native.builder for the application build and OpenRouter for free-model inference.
FALSIFY is a Kalaris Labs experiment around research infrastructure.
The broader direction is to make scientific reasoning more:
- structured
- inspectable
- reproducible
- adversarial
- useful to both humans and AI agents
FALSIFY is one small piece of that direction.
- literature-aware falsification
- evidence provenance
- claim graphs
- experiment comparison
- ATTACK / STEELMAN dual analysis
- collaborative investigations
- research version history
- domain-specific reasoning modes
- structured peer-review workflows
- integration with broader research agents
The goal is not to add features for the sake of it.
The goal is to make the reasoning before experimentation better.
Research tools should survive criticism.
If you find flawed reasoning patterns, weak assumptions, validation issues, UI problems, or useful new falsification workflows, open an issue or submit a pull request.
Appropriately enough, criticism is welcome here.