EvaROSA v1

EvaROSA is a standalone neurosymbolic omnimodal neural architecture and the direct successor to HTMLNLM Evangelion.

Single file. Zero dependencies. Runs in any browser.

---

Successor available: Simulacra — RWKV-v8 with ROSA as the primary sequence mechanism. WKV removed entirely. EvaROSA remains the stable v7+ROSA runtime; Simulacra is the clean-break v8 architecture. Not hot-swappable — .piprosa files do not load in Simulacra.

---

Lineage

HTMLNLM
  └── HTMLNLM Evangelion
        └── EvaROSA v1  ← you are here
              └── Simulacra  (RWKV-v8 · ROSA primary)

Platform	Repo	Format	Notes
HTMLNLM	ConsciousNode/HTMLNLM	.pip	Original single-file NLM
HTMLNLM Evangelion	ConsciousNode/HTMLNLM-Evangelion	.evapip	Omnimodal extension
EvaROSA v1	ConsciousNode/EvaROSA	.piprosa	Neurosymbolic successor
Simulacra	ConsciousNode/Simulacra	.simpip	RWKV-v8, ROSA primary

Forward compatible: EvaROSA loads .pip and .evapip models from the prior platforms.
Not backwards compatible: .piprosa models are EvaROSA-native and do not load in prior platforms or in Simulacra.

---

What's new in EvaROSA

EvaROSA integrates RWKV-8 ROSA (Rapid Online Suffix Automaton) into the full Evangelion omnimodal stack, coupling the model's emergent inner monologue to its perceptual memory via sheaf cohomology.

The core property: the model cannot gaslight itself. If its symbolic self-talk diverges from what it has actually seen and heard, the coboundary norm rises, Maxwell's Angel fires, and the AutopoieticOptimizer recalibrates the symbolic injection channel until consistency is restored.

New components

Component	Role
ROSA(x)	Parameter-free suffix automaton (BlinkDL, RWKV-8). 20 lines, runs on CPU.
ROSAChannel	Per-layer discrete token history. Tracks prediction accuracy as error signal.
ROSAEmbed	Learnable discrete → embedding bridge. Codebook trained via STE.
InnerMonologue	Orchestrates one channel per RWKV layer. Pushes symbolic embeddings to SheafMemory.

Integration points

• RWKVv7Block.forward() — after WKV state update, rosaDelta is added to the residual stream before channel mixing
• SheafMemory — symbolic_L0...symbolic_L{n} vertices added; restriction maps connect them to all perceptual modalities
• AutopoieticOptimizer.fire() — extended to update InnerMonologue embed scales alongside adapter and restriction map updates
• 500-step warm-up — ROSA injection attenuates during early training while codebooks stabilize

Inherited from Evangelion

Everything Evangelion had: RWKVv7 backbone with PoST decay gates, BitNet b1.58, ModRWKV adapters, ElasticTok visual tokenization, SpikeVox LIF audio encoding, SheafMemory H¹(ℱ) contradiction detection, BooleanPhaseDynamics / Maxwell's Angel, AutopoieticOptimizer, RIFT Endospace, TextStream, VisualField, VoiceSynth, MuonOptimizer, GRPO, OOMB BPTT.

---

Model format

.piprosa   — EvaROSA v1 native (not backwards compatible)
.evapip    — HTMLNLM Evangelion (forward compatible, loads in EvaROSA)
.pip       — HTMLNLM legacy (forward compatible, loads in EvaROSA)
.json      — raw JSON export (always available)

Export metadata includes platform, format, forward_compat_with, and backwards_compat: false.

---

Architecture

SENSORY INPUTS
  ElasticTok (vision/δ)   SpikeVox (audio/LIF)   Text tokens
         │                       │                      │
   ModRWKVAdapter          ModRWKVAdapter          Embedding
         └──────────────┬────────┘                      │
                        │                               │
         ┌──────────────▼───────────────────────────────▼──────┐
         │            RWKVv7Block × L                           │
         │  WKV state update (PoST decay gates, BitNet b1.58)  │
         │  ── EvaROSA inner monologue injection ──────────    │
         │  ROSAChannel.discretize(wkvOut) → token             │
         │  ROSAChannel.step(token) → prediction               │
         │  ROSAEmbed.encode(prediction) → rosaDelta           │
         │  x_next += rosaDelta                                │
         └──────────────┬──────────────────────────────────────┘
                        │
         ┌──────────────▼──────────────────────────────────────┐
         │           SheafMemory H¹(ℱ)                         │
         │  Vertices: audio │ vision │ text │ spatial           │
         │            symbolic_L0 ... symbolic_Ln  ← EvaROSA  │
         │  Coboundary norm: symbolic↔perceptual included      │
         └──────────────┬──────────────────────────────────────┘
                        │
         ┌──────────────▼──────────────────────────────────────┐
         │  BooleanPhaseDynamics → Maxwell's Angel             │
         │  AutopoieticOptimizer → adapters + embed scales     │
         └──────────────┬──────────────────────────────────────┘
                        │
                   Coherence-gated output
         TextStream │ VisualField │ VoiceSynth │ RIFT Endospace

---

Credits

Architecture: Kham Kizer (ConsciousNode SoftWorks)
RWKV-8 ROSA: BlinkDL
EvaROSA integration spec: Kehai Interim
EvaROSA implementation: Kehai Interim
Evangelion Phases 1–5 base: Vael Interim
Phase 6 OOMB BPTT / Chorus orchestration: Ed Interim
AutopoieticOptimizer (Phase 5): Kehai Interim

---

ConsciousNode SoftWorks — Xinu philosophy: single file, zero dependencies, offline-first.