Embedded Files
Nodri 0.1.1 project records 1,797 exact matches from 1,797 scored test prompts at seed 17. The total comprises 1,397 math cases and 400 controlled-language cases. Fifty expected unknowns are included in the 400 text cases. The report records zero errors and zero mutating responses during scoring. These results support feasibility within the tested domain. These results are prospecive at best but may lead to measurable general-language ability and/or an established advantage over current architectures.
What was tested
The math cases comprise addition (266), subtraction (232), multiplication (87), exact division (84), and one-variable integer equations (728). All were scored by exact match. The model trains arithmetic operators; division and equation answers are inferred through bounded inverse search rather than supplied as training labels. The text test covers direct facts, two- and three-hop consequences, and unknown cases. Before scoring, evaluation supplies 400 facts across 50 test worlds: eight facts per world, followed by eight scored prompts per world. World labels are disjoint from the training worlds. This tests controlled reasoning with taught context, rather than recall of unseen facts. Evaluation-context teaching and non-mutating answer scoring are distinct phases.
Correct and total cases across all twelve benchmark categories. Bar length represents case count; all categories reached 100% in this recorded run.
Do learned relation tensors affect the result?
In a historical ablation with identical taught-world context, zeroing the relation tensors made all 400 text prompts abstain. Text exact match fell from 400/400 to 50/400 (12.5%). The 50 remaining correct cases were precisely the expected unknowns. This supports dependence on relation tensors within this implementation..
Comparison with simpler controls
A separate development comparison uses exposed seed 17 and 1,797 cases in each of the core and structural-transfer suites. The recorded results are:
Full PRTG (A0): core 1,797/1,797 (100%); transfer 1,797/1,797 (100%).
Random/frozen control (B0): core 160/1,797 (8.9%); transfer 153/1,797 (8.5%).
Frequency-only control (B1): core 250/1,797 (13.9%); transfer 250/1,797 (13.9%).
Static message-passing graph (B2): core 1,797/1,797 (100%); transfer 1,797/1,797 (100%).
Development controls on the left; the separate historical tensor ablation on the right. The two panels represent distinct experiments.
How much of the graph was active?
In the historical reference run, mean activation was 4.05 nodes and 2.27 edges per scored prompt. Observed maxima were 7 nodes and 6 edges, below configured caps of 32 nodes and 64 edges. These counts describe bounded routing, not wall time, memory consumption, or physical energy. They do not establish an efficiency advantage over another model.
Mean active node and edge counts for math, text, and the full reference test set.
Page updated
Google Sites
Report abuse