{
  "title": "Tilelli: a small routed byte-LM with verifiable claims, and a negative result on modular metacognition",
  "upload_type": "software",
  "version": "0.1.0",
  "language": "eng",
  "license": "Apache-2.0",
  "access_right": "open",
  "creators": [
    { "name": "Tilelli LLM Team" }
  ],
  "description": "<p>A 10.2 M-parameter byte-level language model with a 3-pathway heterogeneous routed block. Trained on a single GPU, runs on a laptop CPU. Every numerical claim in the README is bound to a reproduce script that exits non-zero if the bundled checkpoint fails to produce the documented number.</p><p>The deposit ships verified positive results (held-out &quot;I don't know&quot; gate 9/10 at 0% in-domain false-positive; false-inability probe 7/20) alongside three verified <em>negative</em> results, which are the scientific contribution: (1) the pre-registered router-entropy-as-metacognition claim is disproven (0/7 regimes); (2) abstain heads in small routed LMs are not modular &mdash; a head reaching AUROC 0.76 in situ drops to 0.54 when spliced onto a fresh base; (3) the router is fragile at this scale &mdash; cross-entropy backprop on an in-domain subset, with the metacognition loss set to zero, still shifts routing enough to break out-of-domain generation. The accompanying paper (PAPER.md) is included in the deposit. The preliminary single-seed architecture benchmark is reported honestly and is explicitly <em>not</em> a defensible win. See README.md and results/ for the full audited claim trail.</p>",
  "keywords": [
    "small language model",
    "mixture of experts",
    "routing",
    "calibration",
    "uncertainty",
    "abstention",
    "negative results",
    "reproducibility",
    "byte-level"
  ],
  "related_identifiers": [
    {
      "identifier": "https://github.com/TilelliLab/Tilelli-llm",
      "relation": "isSupplementTo",
      "resource_type": "software",
      "scheme": "url"
    },
    {
      "identifier": "https://tilelli.tech",
      "relation": "isDocumentedBy",
      "scheme": "url"
    }
  ],
  "notes": "Apache-2.0 for code, weights, the paper, and this deposit. Reference/research artifact, not a product. Negative results are first-class: the deposit's honesty is the point."
}