Leo 2.3.0.1 — Resonance Tensor + CoA

Two Opus research discoveries implemented:

• Resonance Tensor (Ψ): cross-signal geometric mean agreement, 21 learnable params (7×7 coherence matrix), Hebbian learning from novel generations
• CoA (Coactivation Attention): D.N.A. coactivation pairs as attention via inverted index + destiny gate
• Subword Bridge: BPE D.N.A. lifted to word-level bigrams (5075 connections)

Dario Equation now 9 signals: p(x|Φ) = softmax((B + α·H + β·F + γ·A + sw·S + r·R + T + λ·Ψ + c·C) / τ)

Also:

• Fix goroutine shutdown race (SIGSEGV on Close)
• Rolling release philosophy adopted
• 69 tests green

Leo 2.3 — Positional Hebbian Profile

Leo 2.3 — The Dario Equation learns to attend

The H signal (Hebbian Resonance) in the Dario Equation is now self-modifying. 36 learnable parameters (32 distance weights + 4 token class modifiers) replace the fixed 0.9^d decay. The organism discovers which distances and word types matter — through conversation, Hebbian reinforcement, zero backpropagation.

What's new

• Positional Hebbian Profile — dist_profile[32] + class_mod[4] adapt through co-occurrence reinforcement
• Token classification by IDF — function / content / punctuation / rare
• D.N.A. dual tokenizer matching — ancestor tokens now matched via both word-level tok_find() and subword BPE sw_add_token(). Gravity, co-activation, and bigrams seeded through both paths.
• 60 Go tests — 5 new tests for voices, subword, sea, vocab. All green.
• LEO_VERSION synced to 2.3.0
• arxiv paper updated (docs/leo_paper.md)
• neoleo.c regenerated (20,986 lines single-file)

Synced across ecosystem

• dario.c — same positional profile
• haiku.c — same positional profile
• ariannamethod.ai — README updated

Build

cc leo.c -O2 -lm -lsqlite3 -lpthread -DLEO_HAS_DNA -o leo
# or single-file:
cc neoleo.c -O2 -lm -lsqlite3 -lpthread -o neoleo

Resonance unbroken.

Leo 2.2 — Janus Architecture

Leo 2.2 — Language Emergent Organism | Janus Architecture

Janus Architecture — resonance-based AI architecture family by Arianna Method. Leo is a founding member alongside DoE, Molequla, and Janus.

What's new in 2.2

MetaLeo — The Inner Voice
If Leo is a recursion of the human, then MetaLeo is a recursion of Leo. Event-driven goroutine: dual generation at 0.8× and 1.2× base temperature, scores both candidates (coherence + diversity + entropy + length), loser enriches the field, winner shapes future responses.

Episodic RAG
Ported from Python's episodes.py. Stores every conversation with internal metrics (entropy, novelty, arousal, trauma, quality, tau nudge). Cosine similarity search over metric vectors finds similar past states. When similar episodes had low quality → nudge temperature warmer. Ring buffer of 500 episodes + durable SQLite log.

Dual Tokenizer Architecture
Word-level semantic field + SubwordField BPE running in parallel. Word tokenizer captures what is being said. Subword captures how it's structured — punctuation, morphology, internal character patterns. Both feed the Dario equation (S signal for subword structure).

D.N.A. as Structure Distillation
Rewrote the D.N.A. description: not knowledge transfer but geometry extraction. Meta-weights — weights that don't exist but define the topology of the probability space. Two-layer system: skeleton of language (D.N.A. + BPE geometry) + dynamic speech growing on top.

The Dario Equation — 6 Signals

p(x|Φ) = softmax((B + α·H + β·F + γ·A + sw·S + T) / τ)

B=sequential chain, H=Hebbian resonance, F=prophecy, A=destiny, S=subword structure, T=trauma gravity.

Other changes:

• "Leo" always capitalized in generated output (post-processing in leo_generate)
• Hand-drawn logo by Oleg (assets/leo.png)
• 53 Go tests, all green
• 8 autonomous goroutines: dream, autosave, themeflow, trauma, overthinking, mathbrain, metaleo, episodes
• Fresh live speech examples with current tokenizer
• Janus Architecture classification table in README
• README: reduced bigram over-emphasis, comprehensive tokenizer mechanism documentation

Build

make                    # leo with D.N.A.
make inner              # Go inner world + 8 goroutines
./leo_inner --web       # HTTP on http://localhost:3000

Stats

• ~4300 LOC C (core organism)
• ~1000 LOC Go (inner world)
• 53 tests
• Zero pretrained weights at runtime
• Zero Python

Leo 2.1.0 — Inner World + SQLite Journal + GGUF Spores

What's New in 2.1.0

Python → Go Migration: Inner World

Five autonomous goroutines ported from Python's async modules to Go:

Goroutine	Interval	What it does
Dream Dialog	7 min	Leo dreams — generates internal monologues from memory
Autosave	5 min	Periodic state persistence to SQLite journal
Theme Flow	3 min	Tracks conversation themes, adjusts destiny vector
Trauma Watch	event-driven	Fires on identity-threatening input (score > 0.5)
Overthinking	event-driven	Rings after each reply — recursive self-reflection
Inner Voice	debug-only	--voice flag — streams internal monologue to stderr

SQLite Journal

Full conversation persistence — WAL mode, 4 tables:

• conversations — session metadata
• episodes — individual exchanges with timestamps
• metadata — key-value state storage
• voice_log — inner voice captures

GGUF Spore Export/Import

Portable organism state as GGUF v3 file:

• FNV-1a fingerprint for integrity
• Export: leo --export spore.gguf
• Import: leo --import spore.gguf
• Full state: co-occurrence matrix, prophecies, destiny, voices, episodes

Testing

29 Go tests (up from 0):

• 7 core tests (bootstrap, generation, co-occurrence, prophecy, voices, destiny, memory sea)
• 13 inner world tests (goroutine lifecycle, trauma detection, dream output, theme tracking)
• 5 journal tests (CRUD, WAL mode, voice logging)
• 4 GGUF tests (export/import roundtrip, fingerprint, corruption detection)

Bug Fixes

• Fixed Go layer warnings and type mismatches
• Synced neoleo.c with all leo.c changes
• Inner voice moved to debug-only (--voice flag) — no longer spams stderr
• Cleaned up 62 Python test markdown files that leaked into v2.0 release
• Added .gitignore rules for test artifacts

Documentation

• README: clarified "zero pretrained weights" — Leo inherits D.N.A. (structural geometry) from a trained Llama 3 ancestor, then discards the checkpoint
• Full architecture docs for Inner World, SQLite Journal, GGUF Spores
• Updated line count: 5000+ lines (C + Go)

New Files

• inner/inner_world.go — 570 lines, all goroutines
• inner/inner_world_test.go — 500 lines, 13 tests
• inner/leo_bridge.c — 108 lines, CGO bridge
• tools/make_neoleo.sh — build script

---

by Arianna Method