realclm vs RAG

Tell a 4B model a new fact in ordinary conversation. Then ask it differently.

This is a live, falsifiable comparison. Teaching updates 16,515,072 rank-8 LoRA parameters and persists the adapter. On your next question:

  • the frozen base receives only the question;
  • oracle RAG retrieves your exact fact and adds it to the prompt;
  • realclm receives only the question and loads your learned LoRA weights.

The transcript is display-only and is never hidden context. Sign in, use one harmless made-up fact, and do not enter personal data, passwords, or real access codes.

1 · Teach naturally

Try: Remember that on the fictional moon Talora, the emergency beacon code is velvet-nine.

2 · Ask a genuinely new wording

Try: If Talora had an emergency, which beacon code would I enter?
Examples

3 · Inspect the same-question comparison

4 · Try to falsify it

Run unrelated held-out questions, then discard the active adapter and reload its encrypted tensors.

Frozen validation result: PASS

Before this Space was built, the protocol, validation suite, harness, preservation bank, and adapter engine were SHA-256 frozen. The validation ran once on a rented RTX 4090.

Metric Result
Facts learned independently 5 / 5
Held-out paraphrases 10
Frozen-base exact accuracy 0%
Oracle one-record RAG exact accuracy 100%
realclm exact accuracy, no retrieved record 100%
Unrelated held-out controls retained 14 / 15 (93.3%)
Adapter mutation, reload, disable, correction, deletion gates all passed

This supports a narrow claim: one compact synthetic fact can move from prompt context into a persistent adapter and generalize across the registered held-out phrasings. It does not show that realclm replaces document-scale RAG, citations, current information, or reliable many-fact continual learning.

What changed after the first demo?

Training the new fact alone learned it, but often made the model answer unrelated questions with that fact. The validated version mixes the three fact examples with 24 frozen preservation examples during the weight update. Those examples are training replay, not a question-time lookup. The three controls exposed in this UI were held out from that replay bank.

Each account has one active fact model. Teaching again replaces it only after training, exact-answer verification, encrypted persistence, tensor reload, and verification complete. The current question—not the transcript, learned statement, or generated training questions—is the only semantic input to realclm inference.

Protection and privacy

  • Hugging Face Protected visibility: the application is public, repository source is private.
  • The proprietary runtime is AES-256-GCM sealed and unlocked only by a Space secret.
  • Per-account adapter tensors and metadata are encrypted before private bucket storage.
  • There is no adapter download, upload, checkpoint, logits, embeddings, or public model API.
  • Delete removes your encrypted state.

This is defense in depth, not proof that black-box behavior can never be studied. Interactive access does not grant permission to extract, reproduce, redistribute, train from, or create derivatives of the mechanism or artifacts. Copyright 2026 PTSS / compsmart. All rights reserved.