Embedded Files
Latest Model - Nodri PRTG 0.1.1
How the model works
A controlled-language parser converts a prompt into a typed observation. Arithmetic uses learned 2 × 2 operator tensors; text nodes carry 16-dimensional fibers and learned relations connect them through transport tensors. Inference selects candidates within fixed depth, branch, node, and edge budgets. The answer includes the selected computation path. Relations have fast and slow tensor state, with effective transport equal to their sum. Supported updates can consolidate state; contradictions can reopen plasticity. Explicit teaching and feedback update the model, while snapshots make the saved numerical state available for inspection.
Two views of one saved state
The accompanying spatial figure shows 2,007 nodes and 1,604 operational relations across 403 graph components, captured at event 25. Colors distinguish node kinds.
The first view projects the 16-dimensional node fibers into three PCA coordinates. Those components retain about 21.3% of fiber variance, so substantial information remains outside the display. The second view starts from PCA and applies 120 bounded relaxation steps using the implemented scalar routing cost.
Both are diagnostic projections. Display proximity does not establish semantic similarity, and the picture does not demonstrate a recovered intrinsic manifold, curvature field, or all-pairs geodesic geometry. This later saved snapshot is separate from the historical benchmark state discussed on the Experimental Results page.
Inspecting a learned relation
The tensor closeup shows the exact 16 × 16 effective transport matrix for the first operational relation in the saved manifest. This example has numerical rank 1, an internal confidence score of 0.75, and a normalized squared transport residual of approximately 1.00 × 10⁻⁸. The confidence score is derived from support; it is not a calibrated probability of correctness.
The same saved model shown as fiber PCA and a relational-cost relaxation. Source: immutable snapshot e00000025-31ee232da476.
An actual effective transport tensor, with the source and target labels shown alongside its measured properties. The example was selected by manifest order.
A Rotating View
The numbered panels from top left to bottom right show the graphs orbit. A white ring follows the same recorded node through each view, representing the node count, relation structure, graph components, and limits of the spatial projection.
The prototype makes a bounded learn, route, update, and inspect loop concrete and measurable. Its current evidence concerns controlled synthetic tasks.
Read the Experimental Results page for the reference benchmark, tensor ablation, simpler baselines, and the limits of those comparisons.
Page updated
Google Sites
Report abuse