Language Model

QuarkLM begins from a constrained world:

• a human-authored glossary
• a small grammar
• structured story facts
• an admitted-memory log
• generated lessons derived from those files

The model does not begin with internet-scale knowledge. It does not inherit a pretrained tokenizer. It does not use external embeddings. The training corpus is generated from files named in corpus/ledger.json.

Memory-Native Philosophy

Large language models typically compress broad public and licensed corpora into weights first, then use retrieval, fine-tuning, adapters, or prompting to steer that pretrained base toward a task. That path buys fluency and breadth with enormous compute, but the resulting knowledge boundary is difficult to audit: it is hard to prove exactly where a fact entered, why a behavior changed, or whether a later update erased an earlier one.

QuarkLM takes the smaller path on purpose. Knowledge enters as a lesson, becomes part of the admitted corpus, is served immediately by retrieval memory when the corpus can answer it, and only becomes a weight-update candidate after it has a verifiable source. The lifecycle is:

new lesson -> corpus -> retrieval memory -> training candidates -> guarded weight update -> evaluation -> accepted or rejected

In this design, retrieval memory is not a shortcut around learning. It is the first reliable expression of the model's closed world. Weight consolidation is the slower step that teaches the neural model language behavior, routing, paraphrase tolerance, compression, and generalization over that world. An update is accepted only when evaluation shows it improved the target behavior without violating the corpus boundary or regressing prior evidence.

That creates three different evidence states:

State	Meaning
Corpus-known	The lesson is admitted and can be used by curricula, probes, and memory.
Memory-served	Retrieval can answer from the admitted corpus with provenance.
Weight-consolidated	The transformer learned behavior from guarded candidates and passed promotion gates.

QuarkLM keeps those states separate because mixing them would blur the central claim. If retrieval answers correctly, the system knows the corpus contains the answer. It does not prove the neural weights learned the answer. If a guarded update improves one target but fails retention or branch diversity, the update is useful evidence but rejected for promotion.

Contrast With Conventional LLM Growth

Question	Conventional large model	QuarkLM
Where does broad knowledge enter?	During large-scale pretraining on broad corpora.	Through explicit corpus admission events.
What serves a new fact first?	Often the pretrained model, sometimes retrieval.	Retrieval memory over admitted corpus records.
When are weights updated?	During pretraining, fine-tuning, RL, adapters, or model editing.	Only after source-backed candidates pass guarded training checks.
What counts as improvement?	Better aggregate task, benchmark, or preference scores.	A promoted run that passes closed-world, retention, coverage, and docs gates.
What happens to failed updates?	They may be discarded or hidden in experiment history.	They remain versioned diagnostic evidence for the next repair.

Tokenizer

QuarkLM already has its own tokenizer. closed_world_lm.tokenizer.CharTokenizer learns a character vocabulary from admitted corpus text and rejects characters outside that vocabulary. The current transformer uses this tokenizer.

Future tokenizer work can improve compression, for example with a corpus-derived subword tokenizer, but it must be trained from admitted text only. A pretrained vocabulary would cross the same boundary as pretrained weights.

Transformer

v0.24 adds closed_world_lm.transformer_char_model, a tiny decoder-only transformer built without PyTorch, JAX, Hugging Face, pretrained checkpoints, or pretrained tokenizers. It starts from random weights and trains with a small standard-library scalar autodiff engine.

The transformer is not yet the reliable answering path. It is the weight consolidation path: the component that should gradually learn from admitted training candidates after retrieval memory has made the knowledge available and evaluation can reject harmful updates.

Why So Small?

A small world makes cause and effect inspectable. When the model learns a new fact, the project can show:

• where the fact was admitted
• how the curriculum changed
• which weights were updated
• which probes checked the behavior
• whether older behavior was preserved

That is the model boundary: growth is gradual, bounded, and auditable.

What Counts As Knowing?

For this project, a fact is known only when it is inside the admitted corpus and the promoted responder or learned model can answer its probes. If a fact is not in the corpus, QuarkLM should answer unknown.

This is not a claim about consciousness or subjective selfhood. QuarkLM's "self" knowledge is operational: dataset boundary, update process, unknown policy, and current improvement loop.