Hi everyone,
I’m releasing the source code for AletheionLLM-v2, a 354M parameter decoder-only LLM where the epistemic system is not a post-processing layer — it’s a trainable nn.Module integrated end-to-end into the architecture.
─────────────────────────────────
What makes it different
─────────────────────────────────
Every token produces a full epistemic tomography with 30+ fields: aleatoric uncertainty (Q1), epistemic uncertainty (Q2), MAD-calibrated confidence, 5D Riemannian manifold coordinates, intentionality vector, MPC predictive control, and 11 cognitive sub-heads trained with 13 composite loss functions.
The model doesn’t just predict — it exposes why it’s confident or not, at the token level. No black box.
─────────────────────────────────
OOD benchmark results (WikiText-103)
─────────────────────────────────
Model | ECE | MCE | Brier Score
GPT-2 Medium (355M) | 0.0236 | 0.0340 | 0.1618
OPT-350M | 0.0241 | 0.0656 | 0.1595
AletheionV2 Fine-tuned | 0.0176 | 0.0521 | 0.1528 
best
─────────────────────────────────
What’s in the repo
─────────────────────────────────
- Full source under AGPL-3.0 (weights not yet released)
- 261 tests across 15 suites
- 15 scaling configs from 1M to 640B parameters
- Native continual learning: EWC + Experience Replay
- DDP/FSDP multi-GPU training
- LaTeX paper included
─────────────────────────────────
Links
─────────────────────────────────
GitHub: GitHub - gnai-creator/aletheion-llm-v2: Decoder-only LLM with integrated epistemic tomography. Knows what it doesn't know. · GitHub
Paper: https://doi.org/10.13140/RG.2.2.11471.14241
─────────────────────────────────
Happy to answer questions on the architecture, the epistemic loss schedule, or the DRM 5D manifold design.
Independent research — no institution, no team. Built in Florianópolis, Brazil.
