Glioma is one of the most lethal diagnoses in medicine, and one of the most intensively studied. The distance between those two facts is usually put down to biology. Our conviction is that a large part of it is infrastructural: the systems a research team needs in order to ask its question precisely, revise it honestly, compare it fairly and hand it on are missing, and nobody in the field owns the job of building them.
A composite, not a case report: the shape of a longitudinal glioma study as it is usually run, with the moments where a missing piece of infrastructure quietly changed the answer. Nobody in it made a mistake a reviewer would catch. Each moment is tagged with the mechanism at work; the four mechanisms are named in the next chapter.
“Does progression differ between IDH-mutant subgroups in our longitudinal cohort?” It is said in a meeting, paraphrased in a grant, and never written down in a form a tool can read.
The sentence is recorded verbatim and turned into a scientific task specification: target quantity, population, observational unit, pairing, baseline, permitted transforms, output meaning. Version 1 is accepted by the person who asked.
An analyst pulls the cohort. The export counts scans, because that is how the registry stores follow-up. The word “denominator” does not appear anywhere.
The export runs against the specification. Unit is per patient; the export cannot be produced any other way without a versioned change to the accepted state.
The imaging pipeline computes tumour volumes per patient and calls subgroups from the pathology report using one assay. The notebook will later use another. Nobody knows yet.
The pipeline reads the same subgroup definition the export used. A second assay can be added only as a declared transform, and the specification records that it was.
Kaplan–Meier curves by subgroup, complete cases only. The comparison looks clean. It is a comparison between three slightly different cohorts.
Every figure carries the specification version and the identifiers of the artifacts it depends on. The curves are the comparison that was asked for, and can be shown to be.
Forty diagnosis dates are corrected upstream. Nobody can say which tables and figures depend on them. The team re-runs what it remembers.
The dependency graph names every affected artifact. Those re-run overnight; everything else is certified unchanged; late outputs are marked against the version they served.
A new prognostic model is compared with the old one. The new model got six weeks of engineering; the baseline got a default configuration. The new model wins.
Both arms are held to the same protocol and the same engineering budget, pre-registered, on a held-out cohort. If the new model does not win, that result is published with the same receipts.
The denominator is not stated. Reviewers cannot see the drift because there is nothing to compare against. The paper is accepted.
The specification, its versions and the certificate of what each figure depends on ship with the manuscript. A reviewer can check the comparison, not just the prose.
The postdoc leaves. The methods are in one repository, the decisions in one chat, the accepted definitions in one person’s memory. The next team starts from an empty folder.
Contract, procedures, reference cases and components are published at their established scope. The next team clones a working investigation and spends its first month on its own question.
A glioma team holds the observations, the methods and the expertise. The thread runs between them, and it is lost in four recognisable ways, each tagged where it appears in the investigation above. Each is an engineering failure with a scientific cost, and each is repaired by infrastructure rather than by effort.
The comparison in the manuscript is not the one anyone intended, and it still produces a p-value. Because the denominator was never written down, reviewers have nothing to compare against, and on publication the drift becomes permanent.
A machine-readable definition of cohort, unit and endpoint, versioned and enforced in every tool that touches the data.
Hundreds of tables and figures depend on the corrected records and nobody can say which, so the team re-runs everything or nothing: a quarter of work, or a result that is quietly wrong. In a disease with an eight-month median, either arrives too late.
Dependency tracked from raw record to figure, so a correction re-runs exactly what it touches and certifies what it does not.
The field learns nothing from the comparison, and the negative result that would have been informative is never published. The trial that could settle the question instead would need more patients than the disease provides in a decade, so the computational comparison is often the only one there will be.
Equal-engineering baselines, pre-registered protocols, held-out cohorts, and negatives reported with the same receipts as positives.
Capability that should compound across labs restarts in each one. The field pays for the same pipeline as many times as there are teams.
Contract, procedures, reference cases and components published at their established scope, so the next team inherits a working system.
Notice what the four have in common. In each case the science was sound and the people were competent. What was missing was a small common centre that keeps identifiers, versions, evidence and accepted state, owned by no vendor and no single lab, while every method keeps its own assumptions. That centre is the thing we exist to build.
These are not values in the decorative sense. Each one rules something out, and each has cost us a shortcut we would otherwise have taken.
The systems that decide whether a glioma study is precise, revisable and fair are too important to be owned by a vendor, rented from a platform, or rebuilt privately in every lab. They should be public, inspectable, and inherited by the next team by default.
What it rules out. A hosted application, a proprietary format, a licence that prevents a competitor from running the same code.
Selection by knowledge of the scientific operation, not by title, institutional prestige or enthusiasm for automation. A requested dashboard or agent is a proposed means; the question we ask first is which comparison it makes possible.
What it rules out. Building for an imagined average user. Letting a tool’s convenience decide which question gets asked.
Identifiers, versions, evidence and accepted state are shared. Methods, models and interpretations are not forced into one framework, because a framework that flattens assumptions destroys exactly the information a comparison needs.
What it rules out. A monolithic platform. Any design that requires two methods to agree on more than the contract between them.
A valid analysis, a narrower interpretation, an uncertainty bound, or the observation needed next all count. In a field with this little evidence, an honest null saves the next team a year, and an inflated positive costs a patient.
What it rules out. Publishing only what worked. Comparing against a baseline that was given less care than the method.
A supported contract, a tested procedure, a reference case and a usable component are released for exactly what they have been shown to do. Where only design is complete, we say only design is complete.
What it rules out. Claims of clinical validity from engineering progress. Roadmaps written as promises.
Three choices define the shape of AURORA. None of them is the only reasonable choice, so each deserves its reasons stated.
We run no wet lab and treat no one. What we can change is the software between an observation and a claim: record linkage, segmentation, response assessment, survival modelling, benchmarking, provenance. That layer is rebuilt in every lab, and it is the layer where a definition can be made binding, a dependency traceable, a baseline equal.
It is also the layer where an autonomous research agent will either earn trust or forfeit it. We would rather set the bar for honesty in that layer than inherit one set by fluency.
Glioma combines a median survival measured in months with a molecular classification in which the subgroup decides the curve. If precise definitions, fast revision and fair comparison matter anywhere, they matter here, and their absence is visible fastest here.
The contract, the methods and the agents are built to carry into the rest of neuro-oncology. But a programme that cannot show its worth in glioma has not shown it anywhere.
Apache-2.0 for every repository, and deployment inside the institution that holds the observations. Nothing is hosted, nothing is pooled, and the receipts an agent leaves can be verified by someone who never runs the agent.
Open also means refusing money that would change the question: no grants, no subsidised computing, no fundraising as a pillar. The contribution is capability, and it stays with the field.
Most attempts to fix a field’s tooling fail in one of two directions. A platform that standardises everything flattens the assumptions that make two methods worth comparing. A loose collection of tools standardises nothing, and the drift described above is the result. The centre we build is deliberately narrow: the handful of things that must be identical for a comparison to mean anything, and nothing else.
Everything under “Shared” is versioned and enforced. Everything under “Kept local” stays with the method that owns it, and is never forced to agree with anyone. The line between the two is the most important design decision in AURORA, and the one we expect to be argued with.
A research question put to AURORA can end in one of four ways, and each is a result the team can build on. The fifth ending, a confident claim the evidence does not support, is the one the whole design exists to make difficult.
The comparison that was asked for, on the cohort that was specified, with every dependency accounted for.
It requires. A binding specification, equal baselines, and a certificate of what the result depends on.
The evidence supports a smaller claim than the one hoped for. The smaller claim is stated and published as the result.
It requires. The discipline to publish at the scope the work has earned, and a venue for negative and partial findings.
The question cannot be settled with these observations, but the range of what is consistent with them can be stated precisely.
It requires. Honest statistics, pre-specified analysis plans, and no pressure to convert a bound into a verdict.
The analysis identifies exactly which measurement, cohort or follow-up would decide the question, and stops there.
It requires. A planner that can tell an absent observation from a failed command, and a culture that counts this as progress.
It is the default ending of research done under time pressure with drifting definitions and unequal baselines, and it is the most expensive outcome in a field where a result is acted on before it is replicated. Every element of the centre, from the binding specification to the certificate every automated step must leave, exists to make this ending harder to reach than the honest four.
Each of the four repairs, and each of the three bets, makes a prediction about what a glioma team will be able to do with this infrastructure that it cannot do without it. If a prediction fails, the claim behind it was wrong, and we would rather learn that in year one than defend it in year five.
So the plan commits to reporting these tests as they are run, in public, including the ones we fail. Where only design is complete, the report says only design is complete.
If, with a machine-readable specification in place, two AURORA tools can still disagree silently about a cohort, the definitions pillar has failed and the specification is decoration.
If a correction to a nine-month investigation is not faster to propagate, and not more completely accounted for, than in the same team’s current practice, the revision pillar has failed.
If a new method can still win a comparison run under our protocol because it received more engineering than its baseline, and the receipts do not show that it did, fair comparison has not been built; a form has been filled in.
If a team that had no part in building a module cannot deploy it from the published contract, procedures and reference cases alone, we have published documentation, not capability.
If the research agent ever produces a claim whose supporting evidence cannot be reconstructed from its certificate alone, without the agent, the honesty bar has not been met and the agent should not be trusted with the next task.
If a second, independently implemented peer cannot substitute for the first without changing the scientific contract, agent-to-agent infrastructure has not been built; a single implementation has.
If the contract cannot be carried to a second CNS tumour with only its local definitions changed, glioma first was a boundary rather than a foundation, and we should say so.
The gap described on this page is old. What is new is that each of its repairs has recently become feasible, and that one of them will be built badly if it is not built carefully.
Since the WHO 2021 classification, a glioma’s subgroup is set by markers such as IDH status and 1p/19q codeletion, and the subgroup decides the survival curve. A cohort defined loosely is now a cohort defined wrongly. Machine-readable definitions stopped being a nicety.
Software can now plan analyses, write code and draft findings fluently. Fluency is not honesty. If the bar for an autonomous research agent in glioma is not set now, by people who understand the investigation, it will be set by default, by whatever ships first.
The pipelines described here run on hardware a hospital or registry already owns. Deployment inside the perimeter, with federation that never pools raw data, is a practical choice rather than an aspiration, which removes the last good reason to centralise observations.