There is an old way to test a database: run the standard benchmark, publish the number. There is a harder way: point the database at a problem nobody has solved, forbid yourself from working around it, and write down every place it fails you. In April 2026 we pointed 8DB at three-dimensional Navier-Stokes regularity. This article is what the database learned.1
The rule
The work runs under one discipline, recorded in the module that hosts it. Every attempt is composed from primitives the store already has: theorem and axiom instances, measure instances with uncertainty, sampled vector fields, differential operators, conservation invariants, proof-obligation emitters, a trust ledger. No turbulence intuition is transcribed into code that bypasses those primitives. If a maintainer finds themselves doing that, the discipline has broken, and the break is itself the highest-severity friction event, because it means the store was not ergonomic enough to keep attention on itself.
Every attempt produces a friction log: each gap classified as a missing primitive, an awkward composition, brittleness, slowness or emergence, with a severity, and each critical or high item spawns a work node. Both outcomes count as success. A full closure would be publishable. A precise reformulation of where the gap lives, as typed obligations a prover could discharge, is a refinement signal for the store. Only incoherent output that also fails to reproduce known results would be failure.
What the first attempt found
The first pass, on 27 April 2026, reproduced two calibration targets by composition, a viscous shock for the one-dimensional Burgers equation against its closed form,2 and an admissibility exponent from a preprint against the preprint's own value,3 then attacked the open lemma in that preprint's twelve-link chain. It did not close the lemma. It produced nine friction events, one critical and seven high, and seven work nodes. Some of them:
- A measure's uncertainty was a single scalar. The exponent's lower bound was rigorous and its upper bound depended on an unproven coefficient. The store could not say that. Asymmetric uncertainty became a work item.
- An identity between two exponents was a string. The store had no first-class atom for an exponent identity, so downstream tooling could not tell a derived constant from a typed one.
- A theorem template did not exist for a PDE statement. The open lemma has the shape "for all points in an admissible set there is a positive epsilon such that a bound holds", and the proof-obligation emitter had only an untyped inequality with free-text sides. Typed PDE templates and typed goal shapes became work items, and the gap recurred in the next obligation, which escalated it.
- The store's own novelty engine was not wired in. The attempt consumed pre-curated candidates instead of generating them through the substrate. That was the critical item.4
The store also queried itself: twenty-three atoms across three types and ten composition edges, with four axioms found load-bearing because theorems and measures transitively derived from them. That self-synthesis is now a standard output.
What later attempts found
A second stream wired the emergence and reflexivity machinery in and recorded zero friction events, which the log flags as suspicious in its own text: either the substrate was perfectly ergonomic on a first attempt or the recorder was not being called. A larger pass deployed everything at once against the same open lemma: a knowledge graph of thirty-three entities,5 a chain of twelve links with eleven marked proved by their author and one open,6 a simulator over seven canonical configurations, and an anisotropic embedding margin computed structurally. Its honest-limits section states that the substrate does not close the lemma; it produces a maximally precise statement of where the gap lives, and the next analytic step remains open. On 11 September the scaling exponents behind an impossibility argument in that work were re-derived from the dilation law as constant functions the tests recompute, and the correction history is in the module.
Then the frontier moved, and the programme's regularity target turned out to be working against the expected outcome. The response was not to abandon the harness but to change what it holds: from an attack on a lemma to a governed record of what the field has established, with our own earlier overreach as a first-class refuted entry. That pivot is the subject of the other articles in this series.
What changed in the engine
| Friction found by the problem | What shipped |
|---|---|
| Uncertainty as a single scalar | Asymmetric bounds on measures, rigorous side distinguished from conjectured side |
| Untyped proof-obligation goals | Typed goal shapes and PDE theorem templates in the Lean 4 emitter; a runner that drives external projects and audits axioms |
| Implications stored as plain edges | Sheaf-bearing atoms with restriction maps audited at write time; a hard violation is rejected at the type level, a soft one is logged |
| Trust assigned by hand | Tiers derived from typed verification events with a ceiling function; failures void the ladder |
| Open problems as nulls | Typed gap sections with cost, falsifier and trigger |
| Results as prose | A binary that exits non-zero if the counts or invariants drift, so the record is a test |
| Memory quoted without parallelism | Enforced commit clamps with machine-readable run summaries as events |
Each of these now serves work that has nothing to do with fluids. Asymmetric uncertainty is what a regulated clearing engine needs when one bound is a legal limit and the other an estimate. Write-time restriction maps are what a compliance graph needs so that an approval never outranks the control it depends on. Derived tiers are what an independent evaluator wants when they ask why a claim is rated the way it is.
Why do it at all
Because a database for consequential data will be asked to hold claims whose truth is contested, whose evidence changes, whose premises get retracted, and whose open questions have to be queryable rather than blank. Regulated finance has that shape. Defence evaluation has that shape. Intelligence analysis has that shape. An open problem in mathematics has it in its purest form, with the added discipline that the answers are eventually checkable by a kernel and the field will tell you when you are wrong. Six months of being told exactly that has made the store better, and the list above is the receipt.
What to bring us
- An open problem in your field. Not to solve it. To see what your knowledge looks like as typed atoms with evidence, and where the store fails you.
- Your friction. If you have run a research programme on infrastructure that was not built for it, the categories above are the ones we would like to compare.
- A prover. If you work with Lean or a comparable system, the obligation templates are yours to break.
Write to hello@8braid.com with the subject "Open problem as a harness". You will hear back from an engineer.
Sources and further reading
- Cavazzini, self-deposited preprint on Zenodo, DOI 10.5281/zenodo.19158797, the twelve-link chain referenced above (unreviewed)
- Buaria and Pumir, Journal of Fluid Mechanics (2024), and Buaria, Lawson and Wilczek, Science Advances (2024), the published depletion results the simulator was calibrated against
- Lean 4 and Mathlib
- 8DB: The Week Navier-Stokes Moved, and What a Database Company Did About It
- 8DB: What a Claim Ledger Must Catch: A Live Case From the Navier-Stokes Frontier
- 8DB: A Benchmark for Whether Your Agent's Data Tells the Truth
This article describes an engineering practice and makes no mathematical claim about the Navier-Stokes equations. References to the number of "proved" links in a preprint report the preprint author's own labels; the preprint is unreviewed and is held at hypothesis tier in our records. Counts describe our encodings on the dates given, not properties of the equations. 8Braid has no affiliation with any author referenced here.
Footnotes
-
The mathematics learned nothing from us. We never expected to solve the problem and we did not; no mathematical result is claimed, then or now. None of the limits recorded in these notes is false modesty; each is the difference between a harness and a claim. ↩
-
We have no Lean-verified fluid content and no PDE solver. The one-dimensional interval gate we pass is a scalar conservation-law check, not a fluid result. ↩
-
The preprint whose chain we composed is self-deposited and unreviewed, and it stays at hypothesis tier in every record. ↩
-
The novelty candidates the harness generated are candidates, evaluated against a small corpus of published papers, and none is claimed as a result. ↩
-
The knowledge-graph counts and margins in this article are properties of our encoding, not of the equations. Where we once said prediction, the record now says retrieval-validity signal. ↩
-
"Proved" is the preprint author's own label for eleven of the twelve links. The preprint is unreviewed and held at hypothesis tier. ↩
Continue the technical conversation
Where could this help your work?
Bring a research question, a database workload or an application you want to build. Let’s connect the ideas in this article to an evaluation that matters to your team.
Discuss this work