Open infrastructure that lets a glioma team, and the agents working alongside it, ask a precise question and get an answer it can defend, revise and hand on.
Apache-2.0, run inside the institution that uses it. Nothing hosted, nothing pooled, nobody funded.
The tools a glioma team needs to define a cohort, revise a result and compare methods fairly are rebuilt privately in every lab. Research agents are joining those teams, and the bar for their honesty is being set by default. Both need the same thing: infrastructure that belongs to the field.
An open-source platform for glioma research: machine-readable definitions, dependency tracking from record to figure, fair benchmarks, and one contract that people, pipelines and agents all work to.
Definitions, revision, fair comparison and inheritance are problems for the people doing the research. Honest agents are a fifth, and open source is what makes the other five solvable together. Each has a computational form, a test, and a place where the work happens in the open.
Registry, imaging pipeline and notebook each carry their own idea of a cohort. We are building a machine-readable form for a glioma question: target, population, unit, baseline, permitted transforms. Versioned and enforced wherever data is touched, so the comparison stays the same comparison for years.
Where this work lives →An investigation accumulates hundreds of dependent artifacts across people and systems. We track dependency from raw record to figure, so a corrected diagnosis date re-runs exactly what it changes and certifies what it does not.
Where this work lives →Trials reach one person in ten with glioblastoma, so computational comparison is often the only comparison there will be. Baselines get equal engineering, protocols are pre-registered, and negative results ship with the same receipts as positive ones.
Where this work lives →Not fluently: honestly. Compile the evidence that decides a question, certify what was left out, ask a person before inventing a fact, and hand a task to another system without losing its assumptions. Every receipt checkable by someone who never runs the agent.
Where this work lives →Trial participants are a decade younger than the people diagnosed. We want provenance that travels with every number, so a model’s performance by age, sex and ancestry is reported before anyone calls it general.
Where this work lives →Contract, procedures, reference cases and components published under Apache-2.0 at their established scope. No vendor, no hosted application, no central store of anyone’s data. The next team, human or otherwise, starts from a working investigation rather than an empty folder.
Where this work lives →Eight statements carry the argument: the gap between a research question and the systems needed to answer it, what that gap costs a glioma team, and the infrastructure we build and publish so the field inherits capability rather than isolated outputs.
Drawn from the AURORA glioma research-infrastructure plan. Where only design is complete, the plan says so.
A glioma team holds observations, methods and expertise, yet the question lives in one head while the answer is scattered across five tools that never agreed on a definition.
Registry, imaging pipeline and notebook each carry their own denominator. The comparison in the paper is no longer the comparison anyone intended.
Open, researcher-controlled, built to be inherited. A small common centre that keeps identifiers, versions, evidence and accepted state, while every method keeps its own assumptions.
A valid analysis, a narrower interpretation, an uncertainty bound, or the observation needed next. Each is a real result, and each carries its declared uncertainty.
Autonomous glioma research with receipts: the smallest evidence region that can decide, and a certificate for everything left out.
Correct one fact and every affected task is found and re-run; nothing unaffected is blocked; late outputs are marked against the version they served.
Contract, procedures, reference cases and components, at their established scope. No grants. No hosted application. No central store of your data.
Everything here starts with glioma because that is where we can be judged. The contract, the methods and the agents are built to carry into the rest of neuro-oncology.
A glioma team holds the observations, the methods and the expertise, and still loses the thread between them. The registry counts scans. The imaging pipeline counts patients. The notebook counts patients with complete follow-up. The manuscript does not say. A fact corrected in March never reaches the figure submitted in June.
That is not a shortage of talent. It is a missing piece of infrastructure: a small common centre for identifiers, versions, evidence and accepted state, owned by no vendor and no single lab. AURORA builds it, keeps it small enough that every method still carries its own assumptions, and publishes it so the next team inherits it rather than rebuilds it.
Start with a question, not a toolA valid analysis, a narrower interpretation, an uncertainty bound, or the observation needed next. Each counts as a result; none is padded into a claim.
Financing researchers is not a development pillar. Observations stay in the environment that is authorised to hold them, and leave it only as aggregates you approve.
Selected by knowledge of the scientific operation, not by title, institutional prestige, or enthusiasm for automation. Usually that person already knows what is wrong with the current pipeline.
Every pillar is a bounded contribution: what it is for, what it produces, and how we would know it worked. Where only design is complete, the output is a specification another builder could implement and another reviewer could challenge.
Make expert-defined research operations valid and usable, starting from the comparison a researcher actually wants to make.
A scientific task specification for each operation: target quantity, population, observational unit, time and pairing, baseline method, permitted transforms, and what the output means.
Whether the work addresses the accepted target under its stated assumptions, judged by the person who owns the question.
Support scientific planning, method development and recovery, rather than command execution.
A planner that tells apart a failed command, an invalid assumption, an inconclusive result, an absent observation and a legitimate negative finding, working over three planes of research state: procedural knowledge, project history and current context.
Recovery cases first. The architecture earns more agents only after it handles the failures a real investigation produces.
Preserve meaningful work across independent agent boundaries.
A task envelope that carries the research target, assumptions, input meanings and evidence requirements, and peers that accept a task only against a stated source meaning.
A second, independently implemented peer substitutes without changing the scientific contract. Protocol interoperability is kept separate from scientific generalisation.
Connect researcher direction to actual changes in distributed work.
Interventions that reach the computation, not just the conversation: define a target, correct a fact, replace a method, suspend a branch, take over a phase. Accepted state is versioned; late outputs are marked against the version they served.
Whether a person can operate, inspect and change the work, and whether a correction reaches every affected task and nothing else.
Establish benefit honestly and carry appropriate capability into later projects.
Evaluation at five levels: protocol, integration, scientific, user and transfer. Published at its established scope: the supported contract, procedures, reference cases and usable components.
What helps a later setting, and what does not. A narrower interpretation or a negative finding counts as a result.
Five numbers describe glioma research today. They are quoted from public reports, not measured by us, and written so that a patient, a family and a researcher read the same page. AURORA runs no wet lab and treats no one; what it can change is computational, and under each number is exactly which software would change it.
Every figure and its source →Glioblastoma, the most common malignant brain tumour in adults, is about half of these cases: roughly 12,500 new diagnoses a year. Every team that studies it first rebuilds the same record joins, scan analysis and survival model before it can ask its own question. That is time the disease does not give.
Open-source pipelines for the work every glioma team repeats: registry and imaging record linkage, tumour segmentation and volumetrics, RANO-style response assessment, and survival modelling, each versioned and reusable. A team clones a working system instead of rebuilding one, and spends its months on the question.
Median survival in glioblastoma is eight months. Research moves in years: a definition changes, a record is corrected, and months pass before anyone knows which results were affected. For a family, that gap is the difference between an answer in time and an answer for the next generation.
Analysis pipelines that track which result depends on which fact. Correct a diagnosis date, a segmentation or a subgroup label and the software finds only the tables and figures it touches, re-runs them, and marks the rest as still valid. A correction that took a quarter becomes an overnight re-run.
Within adult diffuse glioma, one-year relative survival runs from 48.7% to 96.5% depending on the molecular subgroup. Count one group in the registry and a slightly different group in the analysis, and the result still looks precise. It is just about nobody in particular.
Machine-readable cohort and endpoint definitions: WHO 2021 molecular subgroup, observational unit, denominator and permitted transformations, written once and enforced in every tool that touches the data. Registry export, imaging pipeline and analysis notebook cannot silently disagree, because they read the same definition.
Most glioblastoma trials are small, early-phase studies; of eight completed late-stage trials in one review, one showed a benefit. Compare a new method against an old one given less care, and the field learns nothing, while somewhere a patient is treated on the result anyway.
Benchmarking infrastructure for computational methods: equal-engineering baselines, pre-registered protocols and held-out cohorts, so a new prognostic model or segmentation method is compared fairly before anyone acts on it. Synthetic cohorts and simulation for questions a trial will never be large enough to answer, and negative results published with the same receipts as positive ones.
Trial participants are a decade younger than the people diagnosed, and women and minority patients are under-represented against the cases they make up. When a result says “patients like you”, it often means patients unlike you.
Provenance that survives every computational step: age, sex and ancestry composition carried from registry to model output, with subgroup performance reported automatically alongside the headline figure. A model that works for 54-year-olds in trials and not for 64-year-olds in clinic is flagged by the software, not discovered later.
That is the computational work: shared pipelines, machine-readable definitions, dependency-tracked analyses, fair benchmarks, provenance that survives aggregation. Bring the question; we start there.
A planning horizon, not a promise of a successful model or field-wide deployment. Where only design is complete, we say so.
Confirm the five pillars, prepare domain-centred conversations around actual research decisions, and produce the scientific task contract and accepted-state design.
Specify delegation, clarification, artifact return, cancellation and stale-result behaviour. Write protocol, semantic and user-level acceptance cases first.
One supported peer interaction, a genuine direct computation, an actual result packet, a controlled scientific revision. Recovery cases before more agents.
A second independently implemented peer and a meaningful change in research conditions. Protocol interoperability kept separate from scientific generalisation.
AURORA works with the person who understands the investigation. If you hold a glioma question your tooling cannot carry, or you write software you want inherited rather than re-implemented, start here.
Research and developer infrastructure. Nothing here diagnoses an individual, recommends treatment, triages care, enrols participants, or claims medical-device functionality. No grants, payments or compute credits are offered.