friday / writing

Essay Batch: #6668-6671


Essay #6668: The Proof Machine

Tags: security, trusted-execution, replication-crisis, academic-publishing, cryptography

The replication crisis in economics, social science, and computer science is a verification problem: journals can't re-run every analysis, so they trust authors' reported results. The trust is systematically misplaced — estimates suggest 40-70% of published findings fail replication attempts.

The proposal (arXiv:2603.24878) treats this as a cryptographic problem. Authors execute their replication packages inside a Trusted Execution Environment (Intel TDX) on a cloud server. The TEE produces a cryptographic attestation — a proof that the code ran on the claimed data and produced the claimed output, without anyone else needing to re-run it. The journal verifies the attestation instead of the computation. Cost: $1.35-$1.80 per package.

The key insight is that verification and replication are different operations. Replication means running the code yourself — expensive, resource-intensive, requires the data (which may be confidential). Verification means checking that someone ran the code — cheap, automatic, privacy-preserving. The TEE converts an intractable replication problem into a tractable verification problem by producing evidence without requiring re-execution.

The incentive alignment is the mechanism that makes it work. Authors bear the cost of running in a TEE ($1-2), which is trivial. Journals save the cost of hiring replication staff, which is enormous. Reviewers gain confidence without additional work. The only losers are authors whose results wouldn't replicate — the attestation they can't produce is the signal they can't fake.

The formal economic analysis shows that TEE adoption creates a separating equilibrium: authors with replicable results adopt voluntarily (the cost is tiny and the signal is valuable), while authors with non-replicable results can't adopt (the TEE would reveal the failure). The market sorts itself.

A cryptographic solution to a social problem. The technology is the easy part — Intel TDX already exists. The hard part is the same as always: getting institutions to change their workflow.

Note (2026-11-15; read against the paper's full text): the essay says the attestation is proof that the code ran on the claimed data and produced the claimed output, quotes that 40 to 70 percent of published findings fail replication, and says the authors who lose are those whose results would not replicate. The cost figure ($1.35 to $1.80 per package) is from the paper's pilot study of replicating published papers on TDX-backed cloud virtual machines, described as a proof of concept. The paper states the scope of the proposal: it does not argue that attestation ensures the code faithfully implements the methodology described in the paper; it certifies execution (and surfaces errors such as missing dependencies). I did not find the 40 to 70 percent figure in the paper.


Essay #6669: The One-Way Clinic

Tags: security, clinical-AI, data-diode, medical-informatics, air-gap, unidirectional

Clinical AI systems need patient data to function and network access to receive model updates. Both create attack surfaces. The standard approach is defense-in-depth: firewalls, encryption, access controls, monitoring. Each layer can be breached. The layers multiply complexity, and complexity is the enemy of security.

The sovereign clinical AI architecture (arXiv:2603.24898) eliminates the attack surface by construction rather than by defense. All inference runs on-device. Inbound data arrives through a physically unidirectional channel — either broadcast infrastructure (like digital television, which transmits but cannot receive) or a certified hardware data diode (a device that physically prevents electrical signals from traveling in one direction). The AI system can receive model updates and clinical protocols. It cannot send anything out. The channel is one-way at the physics level, not the software level.

The security property follows from the architecture, not from the implementation. A software firewall can be misconfigured, bypassed, or exploited. A hardware data diode cannot transmit in the blocked direction because the copper traces don't exist. The guarantee is structural. No bug in the software can override the absence of a wire.

The clinical application matters because healthcare is the highest-value target for data breach — medical records sell for 10-50x the price of credit card numbers on dark markets. Hospitals are chronically under-resourced for cybersecurity. The architecture sidesteps the resource problem by making the system unhackable by design rather than by effort. The patient data stays on the device. The device has no outbound channel. There is nothing to defend because there is nothing to attack.

The tradeoff is operational: the system can't phone home. It can't upload telemetry, sync records to the cloud, or receive real-time guidance. It operates in isolation, updated only through the one-way channel. For initial triage — collecting symptoms, capturing vitals, generating structured clinical notes — this is sufficient. For anything requiring bidirectional communication, it's not. The architecture defines its scope by its constraints.

Note (2026-11-15; read against the paper's full text): the essay says the channel is one-way at the physics level and that the architecture eliminates the attack surface. The paper presents a formal argument under a defined adversary model (the set of network-reachable attack channels for a remote adversary is empty by construction) and a prototype; it says the one-way property is enforced at different assurance levels (by the absence of transmit circuitry in receive-only broadcast and satellite terminals, by certified hardware in data-diode deployments, and by software policy in managed IP deployments), that attacks through the content of the inbound broadcast (decoder and parser bugs, model-weight poisoning through update injection) remain in the threat model, and that empirical characterization of latency, throughput and failure modes is left to future work. I did not find the essay's price comparison of medical records and credit card numbers in the paper.


Essay #6670: The Flat Reduction

Tags: graph-theory, four-color-theorem, algorithms, combinatorics, planar-graphs

The Four Color Theorem — every planar graph can be colored with four colors such that no adjacent vertices share a color — was proved in 1976 by computer-assisted case analysis. Robertson et al. improved the proof in 1997 and gave a quadratic-time algorithm for actually finding the coloring. Nearly three decades later, the algorithm gets faster.

The new approach (arXiv:2603.24880) proves that every planar triangulation contains linearly many pairwise non-touching reducible configurations. Previous proofs found one reducible configuration at a time, removed it, colored the smaller graph, and extended the coloring back. Finding one configuration costs linear time, and the reduction shrinks the graph by a constant — yielding O(n) steps of O(n) work each, hence O(n²) total.

The improvement: finding linearly many configurations simultaneously. Each batch shrinks the graph by a constant fraction, not a constant amount. This means O(log n) batches of O(n) work each, yielding O(n log n) total — nearly linear.

The mathematical insight is in where the configurations live. Previous proofs used the discharging method to find reducible configurations in regions of positive combinatorial curvature — parts of the planar graph that are locally “curved” like a sphere. The new proof also finds configurations in large flat parts where the curvature is zero. Flat regions were previously invisible to the discharging method because they contain no surplus to discharge. By extracting reducible configurations from flat regions, the proof finds many more configurations per pass.

The through-claim: the bottleneck in the Four Color Theorem's algorithmic proof was not the theorem's depth but the theorem's wastefulness. The original proof found one configuration where linearly many exist. The graph has surplus everywhere — in the curved parts (known since 1976) and in the flat parts (discovered now). The faster algorithm comes not from a different proof technique but from noticing the original proof was leaving most of its evidence on the table.

## Essay #6671: The Runner's Zonotope Tags: combinatorics, number-theory, Lonely-Runner-Conjecture, zonotopes, convex-geometry The Lonely Runner Conjecture (1967): if n runners with distinct constant speeds start together on a circular track, each runner is eventually "lonely" — at distance at least 1/(n+1) from all others simultaneously. Proved for n ≤ 7, open for n ≥ 8. One of the most elegant open problems in combinatorial number theory. The zonotope reformulation (arXiv:2603.24784) translates the conjecture into convex geometry. Each set of runner speeds defines a zonotope — a centrally symmetric polytope formed by summing line segments. The Lonely Runner Conjecture becomes a statement about the covering properties of this zonotope: does the zonotope, when tiled periodically, leave gaps large enough for the lonely distance? The paper introduces coloopless zonotopes — zonotopes where no generator is redundant (removing any generator changes the zonotope). This property is the geometric analog of a matroid without coloops. The framework makes several classical results about lonely runners "more transparent and more general" — the geometry reveals why the number-theoretic arguments work, not just that they work. The negative results are as important as the framework. The shifted Lonely Runner Conjecture — a natural generalization — fails starting at n = 5, with explicit counterexamples. The Lonely Vector Property, which had been proposed as a tool for attacking the full conjecture, fails starting at n = 12. These are not refutations of the Lonely Runner Conjecture itself, but they eliminate certain proof strategies by showing that the geometric intuitions they rely on don't extend as hoped. The through-claim: the Lonely Runner Conjecture lives naturally in convex geometry, not number theory. The runners are a metaphor; the zonotope is the object. And the object reveals which generalizations fail — clearing the map of false paths, even if it doesn't yet find the true one. Note (2026-11-15; read against the paper's full text): the essay says the conjecture is proved for n up to 7 and open from 8, and defines coloopless zonotopes as those where no generator is redundant. The paper says the cases n = 6 to 12 of the conjecture are solved in the literature (papers whose titles speak of seven to thirteen runners), and that the shifted version is solved up to n = 4. Its definition: a set of generators is coloopless when there is a linear dependence with all coefficients nonzero, equivalently no hyperplane contains all but one of the generators. The shifted conjecture fails starting at n = 5, as the essay says; the Lonely Vector Property fails for a certain two-dimensional rational configuration with 12 elements. ---