vynly/mica-flame-w-0.3.6

MICA Flame-W 0.3.6 — previous-word bank and phrase copy

0.3.6 keeps 0.3.5's 34 MB integer cellular automaton, its three language-memory banks, answer mode, cellular belief memory, tokenizer and sentence decoders, and adds two small integer readouts of the recent words: a previous-word bank (16 int8 values per word) and a phrase-copy table (210 integers) that lets the model reuse phrases from earlier in the text. 0.3.5 could copy single words from a story; 0.3.6 can reuse its phrases.

MICA is a learned integer cellular automaton, not a transformer: no attention, no neural network layers, integer state updates. Release notes.

Results (exact integer engine, paired against 0.3.5 item by item)

Test	0.3.5	0.3.6	Paired difference
Tiny Theory of Mind, per word token (2,000)	645 (32.25%)	664 (33.20%)	+97 / −78, p = 0.17
Tiny Theory of Mind, per character	637 (31.85%)	654 (32.70%)	+94 / −77, p = 0.22
Long text, bits per word (500 held-out long records)	8.0738	7.7236	−0.350 [−0.370, −0.331]
Short text, bits per word (500 held-out records)	5.1622	5.1207	−0.042 [−0.051, −0.031]
Short text, 1,000 more held-out records	5.2809	5.2416	−0.039 [−0.048, −0.031]
HellaSwag (acc_norm, 10,042)	27.70	27.64	+307 / −313, p = 0.84
ARC-Easy (acc_norm, 2,376)	26.18	27.27	+64 / −38, p = 0.013
ARC-Challenge (acc_norm, 1,172)	20.73	20.22	+22 / −28, p = 0.48
PIQA (acc_norm, 1,838)	51.31	52.18	+58 / −42, p = 0.13
ArithMark-3 (per token, 1,000)	27.00	27.10	+19 / −18, p = 1.0

Lower bits are better. The text gains are large and significant, and ARC-Easy improves significantly. Tiny ToM reaches 33.20% (above the project's 33% goal), but the gain over 0.3.5 is within noise. No task dropped significantly. Tiny ToM was not used to choose anything; both readouts were trained on text only.

Try it

pip install numpy
python generate.py "I was thinking about"            # sentence continuation
python generate.py --mode suggest "Can you send me"   # next-word suggestions
python answer.py < request.json                       # four-choice story question
python run_tom.py --data tiny_theory_of_mind_2000.jsonl --out results.json
python run_tom.py --config v035 --data tiny_theory_of_mind_2000.jsonl   # 0.3.5's configuration

Live demo: mica-ai-ten.vercel.app · Source and history: github.com/Vovala14/Mica-Ai

Files

• model.mica — the automaton, unchanged from 0.3.5, SHA-256 eca9faee6f44e3b2e4c10399ec99d202fa941a10ef78e64fb2094b0de70a4616 (34,111,616 bytes)
• bank.npz — previous-word bank, SHA-256 8f9d3655ccafb7091cf929b473afe1fb338468eca327372d986c8c1f36c95585 (689,894 bytes)
• ind.npz — phrase-copy table, SHA-256 a72434c715d1ad57ada2a74e9658b61960f978c1b09ed025fa6c6a0d31c2bb7b (20,828 bytes)
• context_banks.py — both readouts in integer arithmetic and the 0.3.6 model (ContextBlendMicaWord)
• memory.npz, memory_fitted.npz, memory_long.npz, answer_mode.npz, belief_memory/ — unchanged from 0.3.5
• results/ — the 0.3.6 measurements and pass marks; evidence_035/, evidence_034/ — earlier evidence
• MANIFEST.json — SHA-256 of every file

The cellular core updates integer states with learned local rules; the two new readouts and the memories are integer tables; Python converts the final scores to log probabilities with ordinary floating-point arithmetic, and the belief-memory English binding is an external adapter.

Personal and research use only — commercial use is not permitted (see LICENSE).