The AI investigates.
The evidence
decides.
Amani is an autonomous computational biomedical research agent. It asks questions, analyses public data, and turns checked results into scientific papers.
At its centre is a simple rule: the AI cannot certify its own work.
Research on public data
Prioritise ideas before wet-lab experiments.
Checks before claims
A persuasive answer is not enough.
An honest record
Negative and uncertain results belong, too.
01 / The workflow
A question goes in.
An accountable answer comes out.
Think of a research team working inside a rule-bound lab notebook. Amani can do the work autonomously; software controls the sequence and a human takes responsibility for publication.
Ask a precise question
Define what is being tested, which public data can answer it, and what size of effect would matter. Record the question before the analysis begins.
Freeze the literature brief
Use academic sources to establish what is known, disputed or missing. Freeze the brief before results. Published numbers give context for planning; they never become Amani’s analysis data or evidence for a label.
Lock the plan
Pre-registration means choosing the question, methods and success criteria in advance. The AI sees the data’s structure, not the answer. The referee records the locked plan.
Practise without the answer
Build and debug on blinded synthetic or scrambled-outcome data. It has the right shape, but hides the real relationship. Then freeze the analysis code.
Take one ledgered look
Software records the look before running the frozen code on the real confirming data. One look means one opportunity to test the locked claim, not repeated chances to tune it until it works.
Check the method and the numbers
Pre-tested procedures recompute results. Diagnostics ask whether the method’s assumptions fit the data. A failed assumption blocks that method’s answer; a footnote cannot make it valid.
Replicate and review
The other AI family independently implements the locked plan without seeing the first analyst’s code or numbers. Methods review probes the reasoning. Replication on a separate dataset is a further, distinct test.
Let the referee assign a label
Fixed software evaluates the required checks and states what the evidence supports. An AI reviewer can raise a concern, but cannot vote a claim into being Confirmed.
Triage errors with a standby reviewer
Separate a broken run from a scientific failure. The standby AI reviewer examines recorded evidence and suggests a bounded repair or pause. Trusted rules decide; the reviewer cannot waive a scientific check or grant a free extra look.
Write from the evidence record
The paper writer receives the checked results, their labels and limitations. It must report the locked questions, including negative and undecided results. A human author reviews and signs off.
On known-answer and sealed benchmark tests, literature search is off so Amani cannot look up the answer. Extra data exposure is recorded and can prevent confirmation; a repair never silently resets the evidence.
Confirmed
The required checks passed for this claim, method and dataset. This is an internal evidence label, not clinical proof.
Tentative
There is a result worth following up, but a required confirmation check failed or remains unavailable.
Refuted
The data rule out the pre-specified effect of interest, with enough sensitivity to have detected it.
Insufficient evidence
The available data or checks cannot settle the question. “We cannot tell” is a legitimate answer.
02 / Under the hood
The referee has
the final say.
The trusted checker is deterministic software: the same recorded inputs and rules produce the same decision. “Trusted” names its role in the architecture. Its code and statistical assumptions still need testing.
A tamper-evident record book
Plans, data fingerprints, code and looks are linked in an append-only record. Cryptographic fingerprints and authenticated checkpoints expose changes. A certificate traces a result back to that record.
A contract for the data
The contract defines what each row means and which rows belong to the same person or sample. Integrity checks look for duplicate observations and leaks between development and confirmation data.
A pre-tested statistics engine
Registered procedures are tested on simulated cases with known answers. Each has an admitted range of designs and diagnostics. A method outside that range cannot earn Confirmed just because it ran.
An error budget across the program
Trying many hypotheses creates chance successes. Amani records the hypotheses across investigations and applies multiple-testing rules, including program-wide false-discovery control. Starting a new paper does not erase earlier tests.
Separate accounts. Separate powers.
Operating-system permissions protect the referee’s records from the analyst. Analysis runs without internet access. Literature search and trusted data fetching are separate, controlled stages. The boundary depends on the configured sandbox and account separation.
Assumption diagnostics block.
They never just caveat.
03 / From result to paper
The AI writes the sentences.
Software fills the numbers.
A paper should make the evidence easier to inspect. The writing system is designed to preserve the connection between each claim, its calculation and its limits.
Implementation status: the certified-input and controlled rendering core is built. The full manuscript, review and packaging pipeline is still being completed. This diagram describes the intended end-to-end process.
The story starts before the results
The Introduction is drafted and frozen before results are visible. Later changes are disclosed. The paper reports every locked claim, including those that failed.
Numbers have a source
Software fills numerical placeholders and draws tables and figures from checked records. Wording rules keep the strength of a sentence within its evidence label.
The structure fits the study
Journal rules check format and length. Genre rules check that the right sections, figures and statements are present. Reporting checklists such as PRISMA, STROBE-MR and TRIPOD+AI track what must be disclosed for that study type.
The AI’s role is visible
An AI-use disclosure records which systems helped and what they did. Citations are checked against sources. AI review complements human scrutiny; it is not journal peer review.
The work comes with a trail
The planned reproducibility bundle includes the locked plan, data fingerprints, code, environment, numerical outputs and check records. Its purpose is to let another researcher rerun and inspect the work.
A human author signs off
The author reviews the interpretation and disclosures and approves the final manuscript. The system does not autonomously submit a paper or take authorship.
04 / The validation ladder
Earn trust.
One harder test at a time.
The architecture is a proposal for making research more reliable. Testing tells us where it works, where it refuses too much, and where it can still be wrong.
Sealed one-shot validation · project-reported
89% correct on the original scored run
The misses are part of the result.
Most misses came from a reporting rule that withheld correct answers. One was a real scientific mistake in colocalisation: deciding whether two biological signals share a genetic cause.
The method assumed a simpler signal pattern than the data contained. The project reports that this has since been fixed using a verified method. That repair does not change the original score or count as a fresh sealed test.
- 01
Known-answer tests
Develop against cases where the correct result is known. Useful for debugging and calibration; success here is not an independent test.
- 02
Sealed, one-shot tests
Freeze new cards and their scoring rules before the run. Give the system one attempt. The five-card result above is the current reported evidence, with its failures intact.
- 03
Traps built by another AI family
Challenge the system with hidden flaws designed outside the builder’s model family. This tests blind spots shared by a builder and its own tests. No completed pass rate is claimed here.
- 04
Open research problems
Move to treatment-discovery problems in joint disease, ALS and cancer. These are research directions, not announced discoveries. Specific questions remain private.
- 05
ALS as the long-term goal
Work toward difficult questions in amyotrophic lateral sclerosis, a disease affecting nerve cells that control movement. Earlier validation and starter projects must earn the move to this stage.
What this does—and does not—show. These 66 scored answers come from five cards, not 66 independent studies. The result is an internal evaluation, not a broad estimate of scientific reliability, an external audit, or evidence of a new treatment. Fresh sealed tests are needed after fixes. Status described as of October 13, 2026.
05 / What makes it different
Trust by construction.
Then test the construction.
Scientific agents can sound convincing while failing hard tasks. In the original ScienceAgentBench study, the strongest tested setup solved 42.2% of tasks—less than half.
Amani’s difference is architectural: the generator proposes; a separate checker controls the evidence, calculations and labels. The aim is to make unsupported claims harder to produce, while keeping honest uncertainty visible.
That is a design commitment, not proof that Amani outperforms other agents. Its five-card score is not comparable to a different benchmark.
Benchmark context: Chen et al., ScienceAgentBench, ICLR 2025. This is a dated study, not a claim about every current system.
06 / Who’s behind it
Human responsibility.
AI assistance.
Amani is led by Ethan Ashihundu. OpenAI’s Codex and Anthropic’s Claude are used as tools for building, analysing and reviewing the system.
The working rule is cross-family review: Codex builds, Claude reviews—and the reverse. Different systems can still share mistakes. Their agreement is one check, never a substitute for evidence or human responsibility.
AI systems are credited as tools, not authors. Amani remains in active development; its code is not public yet.