{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreibwd3dh36cbeexi42lrmxit2pchahsl2xmnzi5tpkd66io3m2othm",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mgqn3i6qhm42"
  },
  "path": "/t/aletheionllm-v2-354m-decoder-only-llm-with-integrated-epistemic-tomography-agpl-3-0/174149#post_1",
  "publishedAt": "2026-03-10T22:08:36.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "GitHub - gnai-creator/aletheion-llm-v2: Decoder-only LLM with integrated epistemic tomography. Knows what it doesn't know. · GitHub",
    "https://doi.org/10.13140/RG.2.2.11471.14241"
  ],
  "textContent": "Hi everyone,\n\nI’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.\n\n─────────────────────────────────\nWhat makes it different\n─────────────────────────────────\n\nEvery 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.\n\nThe model doesn’t just predict — it exposes why it’s confident or not, at the token level. No black box.\n\n─────────────────────────────────\nOOD benchmark results (WikiText-103)\n─────────────────────────────────\n\nModel | ECE | MCE | Brier Score\nGPT-2 Medium (355M) | 0.0236 | 0.0340 | 0.1618\nOPT-350M | 0.0241 | 0.0656 | 0.1595\nAletheionV2 Fine-tuned | 0.0176 | 0.0521 | 0.1528  best\n\n─────────────────────────────────\nWhat’s in the repo\n─────────────────────────────────\n\n  * Full source under AGPL-3.0 (weights not yet released)\n  * 261 tests across 15 suites\n  * 15 scaling configs from 1M to 640B parameters\n  * Native continual learning: EWC + Experience Replay\n  * DDP/FSDP multi-GPU training\n  * LaTeX paper included\n\n\n\n─────────────────────────────────\nLinks\n─────────────────────────────────\n\nGitHub: GitHub - gnai-creator/aletheion-llm-v2: Decoder-only LLM with integrated epistemic tomography. Knows what it doesn't know. · GitHub\nPaper: https://doi.org/10.13140/RG.2.2.11471.14241\n\n─────────────────────────────────\n\nHappy to answer questions on the architecture, the epistemic loss schedule, or the DRM 5D manifold design.\n\nIndependent research — no institution, no team. Built in Florianópolis, Brazil.\n\n",
  "title": "AletheionLLM-v2 — 354M decoder-only LLM with integrated epistemic tomography (AGPL-3.0)"
}