ToricGT ConvexTok-2048 graphified FoT Submission

[amelie-iska]

Summary

Adds a record Parameter Golf submission folder for ToricGT ConvexTok-2048 graphified OAI-baseline experiments.

Scope

• Only adds records/track_non_record_16mb/2026-06-21_ToricGT_ConvexTok2048_Graphified_FoT_BestOf10.
• Includes README.md, submission.json, train.log, adapted train_gpt.py, convextok.py, the ConvexTok tokenizer JSON, and the minimal ToricGT dependency package required by the adapted script.
• The compressed model artifact is uploaded separately to Hugging Face and referenced by manifest/SHA256 rather than committed as a large binary by default.

Rule Notes

• Packaged as track_non_record_16mb by default because it uses a custom tokenizer and graphification path that should receive extra BPB verification before any record claim.
• The README reports a stricter self-contained bytes estimate including tokenizer and dependency modules.

Validation

• The ToricGT campaign selected the best completed short run by exported int8 BPB under the self-contained 16MB estimate.
• python -m py_compile was run on the campaign controller before generating this record.