Constellations of Borrowed Light

It took thirty seconds to read the scan that found my brain tumor. It also took decades. Only the thirty seconds happened in the room.

The decades belonged to people I will never meet. Doctors had been writing down the morning headaches of children since before my parents were born. Other children had carried the same signs before me, and their cases were what made the pattern a pattern. So by the time I was six, walking into doorframes and blaming the doorframes, the answer was old. It was sitting in textbooks.A 2026 systematic review of 17,236 children with central nervous system tumors traces the same presenting signs through decades of accumulated case literature. Each sign alone imitates common childhood conditions; assembled, they point one way.

And it still almost failed to arrive. The pattern crossed decades and thousands of children. The last few feet took four months. My parents spent them refusing to accept what they were told, with no name for what was wrong, only the certainty that the answers did not fit the child they knew. The twelfth visit happened because they would not leave without a scan. Once the image existed, the reading took thirty seconds.

To the system, each of those twelve visits was the first. Nothing inside it counted them, compared them, or carried one into the next.

The pattern reached me through people. Researchers argued over it in seminars and referee reports; doctors learned it and taught it forward. The papers survived. Much of the judgment around them did not. Science still runs on that relay.

AI labs now speak openly of superintelligence. Their systems write software and take on research problems across mathematics and science. Governments and companies are building the supercomputers, data centers, and power systems meant to push them further.The 2026 Stanford AI Index reports frontier models at or above human baselines on PhD-level science questions and competition mathematics, with $285.9 billion of US private AI investment in 2025. OpenAI and Meta now discuss superintelligence in public company principles and product visions.

Yet in that same world, the last handoff still fails. A national campaign cut the median delay for childhood brain tumors by more than half, but its slowest cases still took years. A later prospective study found that almost one child in ten with cancer passed through seven or more healthcare contacts before diagnosis. Teaching moved the middle of the distribution. The tail still waited.The 2025 UK Childhood Cancer Diagnosis study followed 1,957 children. The HeadSmart campaign cut the national median for childhood brain tumors from 14.4 to 6.7 weeks; its 90th percentile was 54 weeks, and its longest observed interval was 435 weeks.

Superintelligence cannot act on a case it never sees or inherit a judgment the surrounding system never records.

My twelve visits were not a failure of discovery; the pattern was already known. They showed the distance between knowledge that exists and knowledge a person can act on.

Starlight reaches us from different moments in the past. The light carries no route with it. Sailors drew lines and checked their charts against the sky before them.NASA explains that the stars visible at one moment reach us from different points in history. The stars in a familiar constellation may sit far apart and have no physical connection. Sailors used known stars with tables and instruments for celestial navigation. See NASA on cosmic distances, NASA on constellations, and the Smithsonian’s account of celestial navigation.

Science works by that same borrowed light. A researcher inherits observations made with instruments she never touched and definitions shaped by arguments she never saw unfold. A doctor draws on children whose symptoms entered the pattern long before the patient in front of her was born.

Papers and databases preserve much of that light, but the next reader still has to reconstruct the chart: which result holds, where it applies, which failure narrowed it, and which claim or task should change next. That chart often lives in one person’s head and disappears when the project changes hands.

Each handoff sheds context. A graduate student spends her first year rebuilding a chart that left with the student before her. A laboratory learns what reaches its weekly meeting, and an institution knows a fraction of what its laboratories know. By the time the field hears anything, it is a paper.

A public frontier is that missing handoff: the findings a community is prepared to build on, and the reasons it accepted them.

AI makes that absence more costly because machines can produce candidate work faster than any community can judge it. Communities do judge, in seminars and referee reports and hallway argument. Little of that judgment reaches the next reader in a form she can inherit.

Writing let an observation outlive the person who made it. It was the first repair to the relay: what a mind held could now survive the mind. But once enough texts survived, scholars faced a new problem. No reader could hold the whole collection in memory.

At the Library of Alexandria, in the third century BC, Callimachus made the holdings navigable through the Pinakes, a catalog organized into broad literary categories. Copyists and paid custodians kept the texts available, while the catalog gave readers a path through them. The collection remained useful only as long as people maintained both.The University of Chicago Library describes Alexandria’s bibliographic innovations, including an early systematic subject catalog, and notes that the research complex’s fate remains disputed. A Cambridge history discusses Callimachus and the category organization of the Pinakes. See also The Library of Alexandria.

Nearly nineteen centuries later, Robert Boyle called his circle an invisible college, held together across countries by letters. In 1665, Henry Oldenburg began printing what had once traveled one letter at a time. He made claims public and attributable across distance; generations of printers, editors, and referees kept them moving.

The Royal Society introduced collective editorial procedures in the eighteenth century and more systematic expert review in the nineteenth. Peer review is younger than the railroad.Boyle’s letters of 1646 and 1647 call his informal network “our invisible college”; historians treat that circle as a precursor to the Royal Society. The Society dates Philosophical Transactions to March 1665 and describes systematic expert review as a nineteenth-century reform.

Today the paper feels inseparable from science. It carries an argument, records an attributable priority claim, and lets another researcher inspect the authors’ account. But when evidence changes, authors may publish a correction as a separate notice while downloaded copies remain unchanged; a failed run may never enter the literature.

In the Phaedrus, Plato has Socrates compare writing to a painting: it appears alive but falls silent when questioned. A paper can carry a claim across centuries; it cannot revise the judgment around it.In Plato’s Phaedrus, Socrates says writing resembles painting because it seems alive but cannot answer a question or explain itself to the reader.

Vannevar Bush proposed a different kind of inheritance in 1945. His memex would let one person store a private record and build associative trails through it. Bush imagined a physician following an earlier case history when a patient presented a puzzling reaction. Those paths would endure, though each belonged to its maker.In “As We May Think”, Bush described the memex as an enlarged supplement to memory. His later examples include a physician following an earlier case and a chemist tracing related compounds.

Software engineers solved a narrower inheritance problem. A patch on a mailing list could be argued over, but its meaning still lived in a thread. The Web gave documents stable addresses; Git bundled a project’s history into something any contributor could clone. Tests made some disagreements cheap to settle.The Linux kernel’s contribution guide still accepts patches by email; Git added a portable history rather than abolishing mailing lists. Tim Berners-Lee’s 1989 proposal described links across existing systems without central control, and the official Git history describes distributed non-linear development.

A later developer could point to a shared object, inspect its history, and test some claims against explicit rules. A scientific community must still decide which result to carry forward, and science asks each researcher to reconstruct that judgment from papers and from whoever remains to ask.

A few fields built more structured records of their own. Structural biologists maintained experimental structures in the Protein Data Bank for decades before AlphaFold inherited it alongside sequence databases and other inputs.

The PDB gave the field a shared, curated, machine-readable record that supported work its founders could not foresee. Its durable unit is the experimental structure. Researchers still carry the changing claims built from those structures through papers and expert memory.AlphaFold’s published methods describe training on structures from the PDB alongside sequence and other resources. See Jumper and colleagues and the RCSB Protein Data Bank. The PDB offers structured inheritance within one domain.

Researchers can rent compute and reuse published algorithms, but no one rents a field’s accumulated judgment. Its members must decide which reviewed results to carry forward, and keep deciding. Without that record, a model inherits the archive’s volume and none of the field’s judgment.Peter Koo argues in “Before the Next PDB” (2026) that a field can hold the capital and several candidate datasets yet lack a mechanism to converge on which one to build before it commits, so the resource never gets made. The PDB was one enabling institution among sequence databases, algorithms, compute, benchmarks, expert practice, and governance. It is the scarce input in this example, not the sole cause.

As attempts become cheap, the bottleneck moves to the people who decide what the field should carry forward. Their decisions leave no durable mark the next researcher can find.

Researchers choose questions partly by cost: the weeks or months it takes to turn a hunch into something another person can examine.

A technician suspects the instrument rather than the sample. A clinician has four odd cases and a full clinic. A student wants to know whether a result holds outside the paper she read. Each weighs the idea against the months, and much of science depends on the people willing to pay.

AI first lowers the price of the attempt. Reconstructing someone else’s published result once cost a season of a working life. So did formalizing one branch of an argument to see whether it survives, or interrogating a scientific model without first learning to program. A person can begin any of them in an afternoon.OpenResearch reconstructs an arXiv paper into a runnable environment. AxiomProver took a natural-language conjecture to a kernel-checked Lean proof. Both lower the cost of an attempt. Neither settles what the attempt means.

Lower attempt costs matter most where research capacity is scarce: a hospital without a research arm, a field without a curated archive, or a student learning the boundary of her discipline. Established laboratories will move faster too. What does not get cheaper is deciding what the attempts mean.

Ted Chiang imagined the far end of that asymmetry in “Catching Crumbs from the Table”: metahumans make discoveries people can no longer follow, leaving humans to study their consequences.Ted Chiang published “Catching Crumbs from the Table” in Nature in 2000. The story later appeared as “The Evolution of Human Science.”

In June 2026, four AI systems returned thirty-nine proposed solutions to ten problems from active mathematical research. Some submissions were new and correct. Others were wrong, derivative, or so opaque that referees spent hours determining what they said. Thirty mathematicians had to judge them.The First Proof Second Batch report publishes the problems, human solutions, AI outputs, logs, referee reports, and methodology. Its organizers tested four systems and assigned each of thirty-nine outputs to at least two expert referees. A Harvard account describes the range of correct, difficult, uncited, and wrong submissions.

First Proof remained far short of Chiang’s scenario: its arguments stayed within human reach, yet thirty-nine proposed solutions required at least seventy-eight expert readings. At thirty-nine thousand, review becomes the limiting resource.

Code-generating models entered a field that had spent decades building repositories, diffs, tests, and review queues to absorb a stranger’s work. A failed test, rejected patch, or changed dependency leaves a visible trace.

In this narrow respect, a date-formatting library gives its next maintainer a clearer record of accepted changes and dependencies than the cancer literature gives its next reader. Science has no common mechanism for turning review into what the next researcher inherits.

First Proof preserved the referee reports, but as documents. An agent can summarize them, yet if the synthesis ends with the session, the next agent rebuilds it tomorrow.

Ten times the output becomes ten times the progress only if a field can absorb ten times the attempts, get each one to a qualified reviewer, and carry each verdict into the work that depends on it. Review and memory are that machinery, and both stay fixed while output grows.

Every check has a limit. A Lean kernel can tell whether a proof matches a formal statement. An instrument can report a measurement under stated conditions. Reviewers decide what either check means for a scientific claim.

OpenResearch can make the code from a paper run.The current OpenResearch paper surface says it deploys an agent to build a minimal reproduction, resolves setup problems until the run is error-free, and can then be directed toward larger experiments. Its public projects show runs and experiments in machine learning. The service preserves an emerging runtime and lineage record. Reproducing a paper’s scientific claim requires further judgment. A successful run shows that the code executed in that environment and produced an output. A reviewer must still decide whether the code matches the method and whether the output supports the paper’s claim.

Mathematics is the exception. It has a fast, exact proof checker, and even there, the kernel certifies validity against a formal statement while a person judges whether the statement captures the claim.

The physical world has no kernel. Instruments drift. Populations differ. A biomarker can move while a patient does not improve. A wet lab may need a year to test the next claim, and no machine will tell it what the result meant. The kind of check a field can perform, more than its size, determines which questions AI can settle now.

Where no kernel exists, reviewers settle claims through judgment. They need a record that preserves conditions and disagreement; provenance alone cannot supply a verdict.

Consider what that means for a single phrase. “Anti-amyloid works” sounds like a finding. It is a compression of many findings, and if you are deciding whether to take the drug, it tells you almost nothing you need to know. You need a narrower answer: will it preserve any part of daily life for someone at your stage of disease, and at what risk? Trials test different drugs in different patients. They measure disease markers and use separate scales for cognition and daily function while tracking different harms.

Lecanemab and donanemab slowed decline by small amounts on trial scales in people with early symptomatic Alzheimer’s disease. Those results do not settle every intervention, outcome, or disease stage gathered under the same label.A 2026 Cochrane review covered 17 trials and 20,342 participants. It found effects on cognition and function that were absent or below established thresholds for minimum clinical importance, alongside increased amyloid-related imaging abnormalities. The pivotal lecanemab trial appeared in NEJM, and the pivotal donanemab trial in JAMA.

A useful record keeps these claims apart: one finding for each intervention and population, with its endpoints, safety signals, and regulatory decision. A new result can support one finding, narrow another, and leave the rest unchanged.

Independence needs the same care. Ten papers can analyze the same group of patients; they are not ten confirmations. Two experiments can be independent and still share one assay, and with it one way of being wrong. If the record hides what the evidence has in common, the next reviewer counts it twice.

Failures need interpretation too. A failed synthesis may challenge a chemical hypothesis, reveal a boundary condition, or mean the freezer died over the weekend. A chemist who records the conditions gives the next reviewer evidence to examine, not a verdict.

In 2026, authors reported that AxiomProver translated a research conjecture into Lean and produced a proof that Lean’s kernel accepted.The authors of “Fel’s Conjecture on Syzygies of Numerical Semigroups” released the formal artifacts in a public repository. A checked artifact establishes validity against its formal statement, libraries, and environment. It does not by itself establish fidelity to the intended conjecture or the novelty of the argument.

Suppose a group of mathematicians accepts that result under public rules, and six research projects begin building on it.The Discovery Engine describes the proposal (informal statement, Lean theorem, proof term, environment, declared dependencies) and the reducer that replays accepted events into current state. A later referee finds that its Lean statement omits a case named in the English one.

The proof still checks. It no longer establishes the claim the group thought it had accepted.

Git preserves the proof and its environment. The paper explains the argument. The mathematicians must decide whether to replace the accepted statement with a narrower one and which projects need to recheck their work.

A writable frontier is that missing handoff, built: a community’s current starting point, holding the findings it is prepared to use, why it accepted them, and the work that depends on them.

When maintainers accept a correction, they change that starting point and keep the old version visible. They flag all six dependent projects for review. Anyone can see who made the change, under what authority, and why.

The correction belongs to the community that made it. Its record can travel, so another researcher can inherit the reasoning instead of reconstructing it.

A constellation is the chart drawn over those records: which findings a community treats as connected, and who drew each line.

Knowledge graphs, registries, and living reviews preserve parts of this record. The missing piece is a named community approving a current starting point and notifying the researchers whose work depends on it.The FAIR principles guide data stewardship. Nanopublications represent small assertions with provenance, the Open Research Knowledge Graph structures scholarly contributions, and Crossmark exposes updates. Each supplies part of this lineage. The proposed addition is community-governed state with declared downstream review.

The same handoff becomes harder in a laboratory, where a failed run can reflect the claim, the conditions, or the equipment.

For years, unsuccessful vanadium-selenite syntheses accumulated in laboratory notebooks. Raccuglia and colleagues turned those failures into training data. When they tested the model with new building blocks, its chosen conditions produced the target class 89 percent of the time, compared with 78 percent under traditional human selection.Raccuglia and colleagues reported the materials-synthesis program in Nature in 2016. The result concerns prospective crystallization success in this bounded program. The comparison was 89 percent for model-guided selection and 78 percent for traditional human strategies.

What made it possible matters more than the eleven-point gap: the failed runs had not been thrown away. The archive stayed inside the group that produced and interpreted it, which is why the model could learn from those failures and why no one outside could.

Under a frontier, the researchers would submit the notebooks with the conditions they tried, what they measured, and what they think those results show. Reviewers might accept a narrower claim. If another laboratory later failed to reproduce it, that laboratory would submit its run to the same record, and the maintainers might judge it too weak to change what the frontier holds.

No kernel settles these arguments. The record must keep both attempts open to inspection while the community decides. Formal proof and physical synthesis have almost nothing in common, yet both need the same thing: a starting point that holds its shape while people argue over what should change it.

A replication fails on Friday. On Monday, the original finding remains the community’s starting point while a reviewer checks the new run. The reviewer records the failure before deciding whether it should change what the community is prepared to build on. The run is activity; the reviewed judgment is state. Only accepted changes alter the frontier.

Someone has to be the reviewer on Monday. That person has experiments and deadlines of her own. If a community leaves review to spare time, researchers will submit proposals faster than maintainers can judge them. A frontier needs a charter that names who may accept changes, how others may challenge them, and how the community replaces maintainers. Acceptance updates one frontier’s record under that charter. It does not declare universal truth, and it does not license anyone to act.

No one speaks for all of science. One result can enter several frontiers, and their decisions may differ.

A laboratory may accept a synthesis on its instruments while a standards body waits for durability data and a manufacturer rejects the process as unscalable. Each community should publish what it will accept, who decides, and how critics can challenge a decision. When their frontiers diverge, the records should show whether they disagree about the question, the evidence, or the answer.

A rejected claim should remain visible with the evidence and reasoning that kept it out. A critic can then challenge the decision or build a competing frontier without rebuilding the disputed record.

Proofs fail to check, predictions miss, and replications do not hold. A community earns trust by keeping corrections visible and stating the limits of its evidence.

And no patient improves because a frontier accepted a claim. The best-kept record still needs a person to carry it the last few feet, to the child in the room.

A hospital can read what changed in a frontier. A clinician still has to notice, often inside a routine built before the correction arrived, and a team has to alter its practice. Adoption is another handoff, carried by people who practice together and correct one another.Atul Gawande concludes, “People talking to people is still how the world’s standards change.” In “Slow Ideas”, he contrasts the rapid spread of anesthesia with the slow adoption of antisepsis and safer childbirth practices. Mass communication can introduce a practice, he argues, while trusted people help others change their routines.

Some evidence cannot be public. Patient records, dual-use work, and embargoed results stay closed, and a published cryptographic fingerprint does not open them. The record can name who holds the evidence, the rule governing access, and the limits of outside review.

Not everything a field knows can enter its files at all. Some expertise stays in hands, and the honest record marks where its own coverage ends.

A form does more than store an observation. Its categories shape what a field can notice. When a form has no box for humidity, a researcher may bury the condition in prose or leave it out. Researchers decide which observations receive names and which failures remain private. Maintainers then decide which claims the frontier carries.

One laboratory may record “no crystal formed.” Another may record “target phase absent.” A shared grammar could let a search find both without forcing either laboratory to abandon its language. The same grammar could place an old negative result beside a new method, or a dormant technique beside assumptions developed after its authors abandoned it.

AI can search more literature and generate more candidate connections than any researcher can inspect. Researchers must decide which one deserves an experiment. To choose, they need to know where the field remains uncertain and which failures have already narrowed the question.

Scientists often talk about taste as a private virtue. In practice, taste directs time: which uncertainty deserves the next experiment or year of attention. A few researchers carry much of that judgment because they have read enough to recognize a consequential connection, or to tell you, gently, that your promising idea was tried in 1987. (If you are lucky, they remember where.)

A shared record cannot manufacture taste. It can expose the material behind it: partial work and failed assays, with their conditions intact.

People working decades apart can collaborate when each contribution changes a frontier others can inherit. Constellation lines are human conventions; different cultures have drawn different forms from the same stars. A scientific frontier gives researchers a shared chart while preserving who drew each line and which observations support it. Navigation begins when someone compares the chart with the experiment in front of her.

A proposal that cannot fail is a manifesto. This one should be tested where its failure would show.

Test it first on one theorem family or one executable machine-learning result, where each automated check has a clear limit. Then run the smallest experiment that could embarrass it.

I am building Vela to run that first test: an open protocol for keeping a community’s frontier, starting with small mathematical ones, where the checks are exact and the mistakes will be public. Its record is portable, so a community can leave and take its history. Whether an outsider can inherit and maintain a frontier no one explains to her is exactly what remains unproved.The Discovery Engine records what Vela already does and what this test must show: inheritance by outsiders, and extension beyond mathematics.

Here is what inheriting would look like. Give the record to a researcher who has never met its maintainers. She opens it, sees what the field accepts and why, and submits a result with the earlier findings it relies on. A maintainer reviews it under public rules. A second outsider then sees which projects that correction reopens and chooses her next task without asking anyone for a private explanation.

An agent should be able to start from the same place, and a separate implementation should rebuild the same record from its history.

When one laboratory reports a failed synthesis, the next changes its first attempt.

Compare the new record with an ordinary repository or living review. Measure how long an outsider takes to find the relevant prior attempts and whether seeing them changes what she tries first. Track the time required to review a correction and whether it reaches dependent work. Count proposals waiting without review. If the new system leaves those measures unchanged, its maintainers have added work without improving the handoff. If the original team must explain missing context, the handoff failed.

A field needs a public frontier: a starting point researchers can inherit and people paid to keep it current.

Researchers should not have to rent access to a field’s memory from the company that organized it. arXiv, the Web, and Git offer a useful precedent: public infrastructure that independent tools can extend. Findings, evidence, and review history must remain portable across implementations.

A permissive license does not create openness if one organization controls identities, reputation, vocabulary, or which version counts as authoritative. Researchers need visible governance and a credible right to leave.

Scientific history gives discoveries more attention than the labor that carries them forward. Institutions weaken open infrastructure when they treat that labor as volunteer cleanup. Review and correction should count as scientific work.

A referee who catches an unstated hypothesis changes what the field can build on; the technician who logged the humidity threshold may save another laboratory months of work. The referee disappears into anonymity; the technician receives a line in the acknowledgments. Both deserve a name in the record.

Maintainers must budget attention. Communities need published rules for queues, conflicts, appeals, and succession. Some proposals will sit unread, and the record should show it.

Scientific societies can choose a narrow frontier and publish the rules for changing it. Funders can pay the maintainers and independent reviewers long enough to see whether outsiders can inherit it.

An AI scientist working alone can increase output without improving shared memory. The maintainers of a frontier carry forward work science has already done, so corrections reach the projects that depend on them, failed experiments outlive the people who ran them, and the relay, at last, keeps records.

None of us made the light we work by. We received it from people we will never meet, many of them unnamed: teachers, technicians, referees, researchers whose failures narrowed the search, and children whose symptoms entered the pattern before mine. We hold it for a while, then work for people we will never meet, never knowing what they will do with it.

We do not own the light. We borrow it, add our lines to the chart, and pass it on, so that it arrives.