vynly/mica-flame-w-0.3.3

MICA Flame-W 0.3.3

A newer release is available: Flame-W0.3.5 includes a verified CPU notebook and improved held-out text prediction. This historical checkpoint remains available for reproducibility. Sentence quality and broad reasoning remain experimental.

Experimental English word-symbol model built around the original B-740 Minimal Inference Cellular Automaton with learned integer rules. This update leaves the cellular rules, vocabulary, long-memory bank and sentence decoder unchanged. It blends two integer short-memory readouts for the first 64 fed symbols: floor((3 * released + fitted + 2) / 4). After 64 symbols, inference uses the unchanged 0.3.2 long bank. No transformer, GRU, external language model or retrieval generator runs inside MICA.

The release is a small word-efficiency improvement with a long-context tradeoff, not a claim of better conversation. The model checkpoint SHA-256 remains 1f4d550386930c8485534368033532301f2db9a49a14a69fd35b702d9e1a1f6d. Memory SHAs: released short 5fda5a6c744749932bfba7f18352ce532405e60fca7d64fd99b705c28b928be7, fitted short 5564ba558e7c1a6e594e324575333a626df16cc8605c9a23534d880acebab4ac, long f2abb7913dd626fb75cfeb96dd97c17c7d7ecbaed88f54c95fdb986734469f3b. Checkpoint plus three memory files total 36,132,566 bytes (36.13 MB decimal). The additional bank adds 673,650 bytes over 0.3.2.

Measured results versus 0.3.2

Exact exported-integer bits per word symbol plus EOS from BOS, lower is better. The candidate was chosen on a new fourth development block disjoint from memory fitting and previous development blocks.

Set	0.3.2	0.3.3	Candidate minus 0.3.2, paired 95% CI
Fresh short development, 1,000 records	5.379107	5.358054	−0.021054 [−0.022820,−0.019287]
Fresh long development, 128 records	8.329670	8.334585	+0.004915 [+0.002938,+0.007085]
Equal-weight development domains			−0.008069 [−0.009427,−0.006711]
Repeated clean short500, 7,320 targets	5.208428	5.189430	−0.018998 [−0.021376,−0.016611]
Repeated clean long500, 100,618 targets	8.082371	8.086188	+0.003818 [+0.002786,+0.004814]
Equal-weight clean domains			−0.007590 [−0.008872,−0.006321]

The long-record regression is statistically visible and was below the predeclared +0.005-bits cap. The clean sets had been inspected in earlier research, so the fresh development block is the main model-selection evidence.

A blinded single-rater sentence check on 20 untouched short and 20 untouched long prompts found 1 win, 1 loss, 38 ties and one new unusable output. It passed a no-regression gate but shows no demonstrated sentence-quality gain. Raw outputs remain often repetitive or incoherent. Unedited changed examples from short prompts:

• I miss my old friends. — 0.3.2: But I really wanted to go to the park with my friends and friends.; 0.3.3: I don't know what I would do without you.
• I want to make her jealous — 0.3.2: of his friends.; 0.3.3: of the new York.

On the public 2,000-item Tiny Theory of Mind benchmark with unchanged answer mode, word-normalized accuracy was 632/2000 (31.60%) → 636/2000 (31.80%), paired +0.20 percentage points [−0.25,+0.65]. That interval includes no improvement. The five-task leaderboard was not rerun; no improved ToM or leaderboard claim is made. The sealed test set was not accessed.

Use

pip install numpy
python generate.py "I owe my life to you"
python generate.py --mode suggest "Can you send me"

run_tom.py reproduces the public Tiny ToM protocol. package.json lists the cellular geometry and every memory route. Full frozen protocol, per-record integer scores, blind ratings and conformance checks are in the MICA project. See the MICA model collection. LICENSE contains the non-commercial terms.