Link GitHub research repository

README.md

@@ -23,6 +23,14 @@ the same **carry-aware TCN** (~10.7M params total across two shared weight files
 than memorising finite multiplication tables, and it verifiably generalises to primes never seen
 in training.
 
+## Source and reproducibility
+
+The complete training code, evaluation utilities, experiment notes, and reproducibility guide are
+available in the public GitHub repository
+[`Xllent-AI/modular_arithmetic`](https://github.com/Xllent-AI/modular_arithmetic). This Hugging Face
+repository remains the standalone submission artifact containing the model code, manifest, and
+trained weights.
+
 ## The idea
 
 Direct classification of the bilinear map `(a, b) -> a*b mod p` does not generalise across
@@ -137,11 +145,14 @@ position. Combined with gradient accumulation (effective batch ~26k) and the wor
 loss, this took tier 9 from **0.73 -> 0.99**, even across prime widths (held-out value-uniform
 validation 0.99; per-width 1015-1024 all ~0.99).
 
-The training scripts and the intermediate checkpoints they reference live in the companion
-research repo (not shipped in this model repo); the commands below document *how the weights were
-obtained* (the provenance the rules ask for) and are not runnable as-is from this repo alone.
-Intermediate warm-start files such as `weights_shared_64_512.pt` are prior in-flight cells, not
-redistributed here.
+The training scripts live under `exploration/` in the
+[GitHub research repository](https://github.com/Xllent-AI/modular_arithmetic); they are not shipped
+in this standalone model repository. The commands below document *how the weights were obtained*
+(the provenance the rules ask for) and are not runnable as-is from this repository alone.
+Intermediate warm-start files are absent from the current revision but remain available in the
+repository history; for example,
+[`weights_shared_64_512.pt`](https://huggingface.co/XllentAI/modular_arithmetic/blob/fff63d15412d615df043af59c90ff28f372173e8/weights_shared_64_512.pt)
+is the prior shared 64–512 cell.
 
 ```bash
 # Historical small-prime cells were first trained width-matched, then absorbed into the shared cell.