frankmorales2020/topological-ai-gpt-oss-20b-multirun

FULL CODE: https://github.com/frank-morales2020/AST/blob/main/GPT0SS20B_TOPOAI_TRANSFORMER_MULTRUN_CORRECTED.ipynb

Topological AI — GPT-OSS-20B Multi-Run (TOPO-2026 Certified)

Author: Frank Morales Aguilera, BEng, MEng, SMIEEE Lab: Sovereign Machine Lab (SOMALA), Montréal, Canada Paper: https://zenodo.org/records/20338459

Multi-Run Configuration

• Runs: 5 (varying lr_embed / lr_cls; seed=123 held constant)
• Deployed checkpoint: Best Task-C run (Run 3)

TOPO-2026 Track II Certificate (averaged over 5 runs)

Metric	Mean ± Std	Threshold	Status
Task C Accuracy	92.3% ± 1.9%	≥85%	PASS
Combined Forgetting	1.6% ± 1.3%	≤10%	PASS
Anchor Memory	67.50 KB	O(1)	PASS

Per-Run Empirical Ledger

Run	lr_embed / lr_cls	Acc A	Acc B	Acc C	Forgetting
Run 0	5e-03 / 1e-03	98.40%	97.40%	93.50%	+1.85%
Run 1	1e-03 / 5e-04	99.60%	99.90%	89.50%	-0.05%
Run 2	1e-02 / 2e-03	92.40%	99.30%	92.50%	+3.25%
Run 3	5e-03 / 5e-03	96.60%	98.30%	94.50%	+2.05%
Run 4	2e-03 / 1e-03	97.40%	99.70%	91.50%	+0.65%