vynly/mica-ember-0.3.8

MICA Ember v0.3.8 — better word completion, same cellular architecture

Ember provides next-word suggestions and typed-word completion. Flame-W handles sentence continuation. This is a research model, not a general assistant.

Architecture

The Minimal Inference Cellular Automaton reads UTF-8 bytes into a ring of 768 cells with 112 integer channels. Sixteen phases select learned integer rules; 240 probes feed integer readouts. The unchanged word component retains 32,768 words, two previous-word-start anchors, and 16 native word-rule ticks whose six signed terms read current probes and fixed codes of the previous three words. Winner rules write six learned integer values. No neural layers, transformer, GRU, external language model or corpus retrieval were introduced.

0.3.8 refits existing prefix2 context coefficients and biases, combines the two existing word heads 9:7, and uses a training-calibrated 12/16 probability gain. Cellular transitions, geometry, word-rule arrays, vocabulary, next-word head, byte checkpoint and integer byte-memory are unchanged. Optional probabilities use floating-point softmax; the scores and cellular updates remain integer. Restricted prefix scoring and optional bounded typing caches avoid unnecessary work. No new general speed benchmark is claimed.

Matched comparison with actual published 0.3.7

After two typed letters; top1 includes out-of-vocabulary targets. Conditional bits/word normalize over only legal prefix words, excluding OOV. These are not bits/byte or unconditional language-model perplexities. Lower bits is better.

Frozen word-stage confirmation	Prompts	Top1 0.3.7 → 0.3.8	Conditional bits/word 0.3.7 → 0.3.8
TinyStories, varied word positions	2000	70.65% → 71.75%	2.530207 → 1.826396
Conversation, first eligible word	2000	80.40% → 81.05%	1.753580 → 1.268459
Everyday, first eligible word	1999	60.58% → 63.03%	4.240615 → 2.977478

Equal-domain accuracy gain +1.4004pp, paired95%CI[+0.8503,+1.9345]. Conditional loss difference −0.817356bits, CI[−0.880697,−0.757175]. These are the existing frozen candidate-confirmation sets, transparently reused for direct public-version comparison, not a newly blind release test. One of 2,000 raw everyday records has no eligible ASCII word; both models use the same1999. Loss denominators are 1,997/1,972/1,982; one probe per normalized-unique record.

Clean reporting only	Prompts	Top1 0.3.7 → 0.3.8	Conditional bits/word 0.3.7 → 0.3.8
Conversation	968	68.90% → 70.56% (667 → 683)	3.228380 → 2.250338
Everyday	937	60.73% → 63.61% (569 → 596)	4.302341 → 3.026711

Equal-domain clean accuracy gain +2.2672pp, CI[+1.2601,+3.3056]. Clean conditional loss difference −1.126836bits, CI[−1.262410,−0.993176]. Loss denominators 953/926. Earlier0.3.7 release tables used a smaller 400-prompt sample and are not directly interchangeable with this full eligible-set table.

Decoder versus newly fitted coefficients

Most gains over public0.3.7 come from the improved decoder and calibration. Compared with ORIGINAL 0.3.7 weights using the same improved decoder, the new coefficients lower confirmation conditional loss by 0.015394bits CI[−0.022877,−0.008062], but the additional accuracy change is inconclusive (+0.1668 pp, CI[−0.2333,+0.5502]). On clean sets, coefficient-only loss changes by −0.022628bits CI[−0.037214,−0.008116]; chat accuracy is 0.7231 pp lower and everyday 1.1740 pp higher than that decoder-only baseline. All domains improve against the ACTUAL public0.3.7 version. This is not a claim that every step or every prompt improves.

What did not improve

Next-word/one-letter behavior and byte loss are unchanged by file/array identity. Full2000-item public Tiny Theory of Mind: 524/2000=26.20% with byte or character length normalization (Wilson95%CI24.32–28.17%). Secondary word-token normalization:545/2000=27.25%. The completion heads are unused by this no-hint byte-scoring task, so0.3.7 and0.3.8 have identical scores. One exact full pass plus byte-path identity audit; no claim of separate independent reruns or ToM improvement. Do not compare normalization conventions without disclosure. No sentence-coherence or free-generation improvement is established.

Data and verification

Learning uses original training text only. Word-stage holdouts exclude original word training, additional fitting/screens and earlier word holdouts, with normalized-overlap checks; their source population was historically used for byte-trainer validation. Candidate fixed before direct comparison; no clean retuning. Paired record bootstrap 5,000 replicates; equal weight per domain. Sealed completion test never accessed. Detailed hashes, outcomes, protocol and ToM audit are in evidence/. Native runtime ranks/probabilities and invariant components are checked separately in evidence/package_verification.json.

Run locally

Only NumPy is required; use the provided native runtime, not AutoModel.

pip install numpy
python test_word_model.py
python test_memory.py
python word_model.py --complete "Please send me the add"
python word_model.py "Thank you for "

from word_model import EmberWord
model = EmberWord()
print(model.complete("The little rabbit went into the fo"))
print(model.complete_with_probabilities("Please send me the add"))
print(model.next_words("Thank you for "))

Completion probabilities are conditional over ALL matching vocabulary words; top 5 probabilities need not sum to 1. One-letter uses unchanged next-word scores. Longer prefixes filter the same causal two-letter score. Vocabulary/English coverage is limited, and suggestions can be wrong. MICA AI Non-Commercial License: personal and research use; commercial use forbidden.

Byte checkpoint SHA256: c2bef375c8dcd6f944e1a44f4b92869645ddf97eb386382fa322cf9519eda506

Word-head SHA256: 44c60a2e64a5d671a03dfc162fca0c9c435b75c2059ebca95eec47ab4a2a92ce

Byte-memory SHA256: 3827c2a38981347bff17f9666543c333178c5a89fdd089d216e5c6f8ab8847c8

Playground · GitHub release · Hugging Face model · Collection