SingularityPrinciple/Gemma-E2B-IT-10M-Chat

NZFC Hybrid Exact Recall 10M

This package distributes a hybrid structural NZFC external-memory exact-recall system.

Safe claim

This is not an internal 10M-token context model. It is an external archive retrieval system using a finite trace-budget NZFC memory readout.

What is included

• Hybrid sparse CSR indexes stored as safetensors blocks
• Lexical, character, and structural/math channels
• SHA-256 and RID based exact recall verification
• Evidence logs for a 10M-token-after complex language-math passage test
• Runtime loader: runtime/nzfc_hybrid_exact_recall.py

Core operator

X_hybrid = 0.25 X_lex ⊕ 0.25 X_char ⊕ 0.50 X_struct
T_mem(q) = diag(w_i(q)) X_hybrid
K(q) = T_mem(q) T_mem(q)^*
T'_mem(q) = Pi_{||T||_* <= tau}(T_mem(q))

Validation basis

• 50,002 records
• at least 10,000,000 tokens after target passage
• target RID: RID_000000_COMPLEX_MATH_CANONICAL_EXACT
• target SHA-256: 03026df135358211a326a95b99da799c065441b47fdc568eb2a9c8a362c5a638
• hybrid safetensors load tested

Quickstart

import sys
sys.path.append('runtime')
from nzfc_hybrid_exact_recall import NZFCHybridExactRecall10M

mem = NZFCHybridExactRecall10M('.')
query = '라그랑주-베셀 NZFC 기억정리에서 T_mem(q), K(q), 핵노름 사영, rank_eff 조건을 설명한 원문 passage를 정확히 다시 가져와.'
strict, selected, diag = mem.query(query, tau_trace=0.3)
print(strict[0]['rid'])
print(strict[0]['exact_text_match'], strict[0]['exact_target_sha_match'])
print(mem.render_pack(query, strict, diag)[:4000])

Non-claims

• This package does not claim universal hallucination elimination.
• This package does not claim that an LLM internally remembered 10M tokens.
• This package does not redistribute Gemma base model weights.