| ai.viXra.org > Artificial Intelligence > 2609 |
|
 |
| ai.viXra.org > Artificial Intelligence > 2609 |
|
|
[4] ai.viXra.org:2609.0031 [pdf] submitted on 2026-09-12 18:40:58
Authors: Brent Hartshorn
Comments: 13 Pages.
Robotic vision policies trained on photorealistic synthetic imagery fail when the real world presentsa texture, a shadow or a reflection the renderer never produced. We argue the fault is the representation rather than the realism: a policy that consumes photographs must learn to ignore appearance, when it could instead be handed a representation from which appearance has already been removed. We describe a chained architecture in which a perception stage converts camera input into a non-photorealistic line drawing paired with a semantic map, and a control policy consumes only that abstraction. Decoupling perception from control allows each to be trained independently [4], and abstract intermediate representations such as sketches are known to preserve spatial grounding while discarding visual distraction [5].The contribution is an open simulator that makes the architecture testable end to end: it generatesthe aligned multi-modal corpus the interface requires — photorealistic, line art, object index and metric depth from one viewpoint — carries a physically grounded robot with an optional rigid-body contact backend, imports externally authored scenes and assets [3], and executes the real control firmware against the simulated plant at two distinct fidelities: a value-level seam fast enough for the training loop, and an offline register-level gate in which the same commands are driven through emulated I2 C peripherals, H-bridges and a serial lidar protocol [11]. The second catches a class of defect the first cannot represent, and we quantify the divergence between them. We report a perception result on 479 samples, a multi-label scene-description head trained from labels the renderer already knew, and a control policy cloned from a privileged expert that reaches its goal in 8/8 held-out scenes while seeing only the abstraction.Every number is reported beside the score of doing nothing, and we record the two hypotheses about thecontrol bottleneck that measurement destroyed, together with two measurement choices the register-level gate itself invalidated.
Category: Artificial Intelligence
[3] ai.viXra.org:2609.0019 [pdf] submitted on 2026-09-10 00:59:20
Authors: Brent Hartshorn
Comments: 13 Pages.
In September 2026 a computer-checked proof of Fermat’s Last Theorem and a Lean-formalizedproof of blowup in the forced Navier—Stokes equations were announced within days of each other,and both are believed because a kernel checked them. This paper applies that discipline at theother end of the scale. A Calculus of Constructions micro-kernel with no axioms, is given an imperative language—assignment, branches, bounded loops, and while with an explicit variant—compiled by symbolic execution into kernel terms. Contracts are decidable propositions, Holds b := Eq Bool b true, about the very expression a runtime check would evaluate. A while loop raises four obligations, one of them the soundness of its own lowering, and termination is proved for every starting state. Records and a state invariant give a model of an seL4-style OS whose routing layer carries contracts that hold for every input. A bridge to Crust, a self-contained C/C++/Rust toolchain, settles its source-level contracts in the kernel; building it found two of Crust’s passes reading one contract two ways. Finally a certificate changes generated code: a SIMD kernel’s scalar remainder is omitted on it, and a contract proven at every call site downgrades 80% of the memory-safety checks in a leaf initializer while the check it leaves still catches an overflow.
Category: Artificial Intelligence
[2] ai.viXra.org:2609.0005 [pdf] submitted on 2026-09-03 23:03:35
Authors: Prakash Vaithyanathan
Comments: 7 Pages. CC BY-NC-ND 4.0
When a very large AI model is trained, the training sometimes breaks on its own. The loss suddenly jumps up, the numbers inside the model become too large or too small to handle correctly, and the run has to be stopped and restarted from an older saved copy. This is called training instability, and on a large training run it can waste a very large amount of computer time and money. Earlier work in machine learning has shown that this kind of breakdown is closely tied to a single number: the sharpness of the loss surface, which is the largest eigenvalue of the Hessian matrix of the loss function. When this number grows too large compared to the learning rate, the training stops behaving smoothly. The problem is that for a model with a billion or more parameters, the full Hessian matrix cannot be built or stored by any realmachine: it would need far more numbers than there are atoms within reach. We show that this is not actually needed. A small hardware block can track the same leading eigenvalue using only the gradient vectors that the training hardware is already computing at every step, through a standard numerical method called power iteration, combined with the well-known fact that a Hessian-vector product can be estimated from the difference between two gradients without ever building the Hessian itself. We also show that watching a small, carefully chosen part of the model, rather than the whole thing, is often enough to catch the problem early. We describe a domain-blind hardware controller that watches this shrinking safety margin every single step and estimates how many steps remain before the run is expected to become unstable, so that a corrective action can be taken before the run actually fails rather than after.
Category: Artificial Intelligence
[1] ai.viXra.org:2609.0003 [pdf] submitted on 2026-09-02 18:16:51
Authors: Raymond Deiotte
Comments: 42 Pages.
Agent-facing knowledgebases compute confidence and provenance at answer time, as properties of the re- sponse. This paper argues that is too late, and that the argument is not a matter of engineering quality: information about whether a claim is true enters through the primary record, and every later stage is a rendering of that record, so no downstream scoring can add a bit that the record did not already carry. Confidence computed from the exhibit has an evidential ceiling fixed before it runs.We describe the store that follows from taking that seriously. The substrate is an uncertainty-typed property hypergraph: every unit carries the same five fields, a knowledge type, provenance graded on a four-valued lattice and carrying a validity horizon, the condition under which the entry held, a version with its conflict history, and a confidence separated into write-time and serving-time quantities, and relations are hyperedges because the relations that matter in agent memory are irreducibly n-ary and reifying them into binary edges destroys exactly the structure later stages need. On that substrate we specify three mechanisms. Consolidation is gated on whether a merged region still decomposes into the inferences it was meant to support, a structural and label-free test that needs no paired QA data and no per-environment training. Conflict is preserved through four lifecycle primitives, demotion, supersession, revocation and invalidation, of which three share a single inverse transition and revocation deliberately has none. Retrieval returns a coverage-guaranteed evidence subgraph with the guarantee conditional on the type of edge retrieved rather than marginal over a query distribution, and it answers with three outcomes rather than one score: supported, insufficient, refuted.The paper’s most portable result is a criterion for when two memory mechanisms can affect each other at all: only where one writes a property of the store that the other reads, at a granularity they share. The criterion is checkable from the specifications before anything is built, it discriminates rather than merely predicting failure, and it carries a stated falsifier.
Category: Artificial Intelligence