A CPU-only virtual machine for self-modifying neural architectures.
Signal: s = p * m p in {-1, 0, +1} m in {0..255}
Format: [ExtID:2][OpCode:2][A:1][B:1][C:1][D:1] 8 bytes fixed
Two bytes per weight. Integer arithmetic. No floats. No GPU. Hot-reloadable.
TVMR is a virtual mainframe — a fixed instruction set that runs programs defining neural architectures, learning rules, structural plasticity, inter-region communication, and lifecycle management. The Rust code is the kernel. Everything else is a program.
Rust Kernel (fixed, rarely recompiled)
├── TVMR instruction dispatch
├── Extension registry
├── Hot-reload engine
└── Substrate I/O bridges
Programs (.ternsig files, hot-reloadable)
├── Region firmware (forward pass, learning, gating)
├── Boot sequences (region topology, wiring)
├── Orchestration scripts (model loading, chaining)
└── Test harnesses (assertions, snapshots)
Why this matters: Changing a brain region's behavior means editing a .ternsig file and hot-reloading. No Rust recompilation. No cargo build. No waiting.
Every instruction is exactly 8 bytes:
[ExtID:2][OpCode:2][A:1][B:1][C:1][D:1]
bytes 0-1 2-3 4 5 6 7
- ExtID (u16 BE): Extension identifier.
0x0000= core ISA. - OpCode (u16 BE): Extension-local opcode. 65,536 per extension.
- A, B, C, D: Operand bytes. Meaning defined per-instruction.
Hot (0x00-0x3F) Activations, intermediates (volatile)
Cold (0x40-0x7F) Weights, Signal storage (persistent via Thermogram)
Param (0x80-0xBF) Scalars (learning_rate, babble_scale)
Shape (0xC0-0xFF) Dimension metadata
4 banks x 64 registers = 256 total.
Programs declare their dependencies. Extensions are registered at boot. Each extension owns a 2-byte ID and up to 65,536 opcodes.
.requires
ternary 0x0002
activation 0x0003
learning 0x0004
.registers
C0: ternary[32, 12] key="region.w1"
H0: i32[12]
H1: i32[32]
.program
load_input H0
ternary_matmul H1, C0, H0
activation.relu H1, H1
store_output H1
halt| ExtID | Name | Ops | Status | Domain |
|---|---|---|---|---|
| 0x0000 | core | 104 | Built-in | System, registers, arithmetic, control flow, debug |
| 0x0001 | tensor | 18 | Functional | Matrix ops, reductions, slicing |
| 0x0002 | ternary | 14 | Functional | Signal matmul, quantization, gating |
| 0x0003 | activation | 5 | Functional | ReLU, sigmoid, tanh, softmax, GELU |
| 0x0004 | learning | 20 | Partial | Mastery, CHL, babble, eligibility |
| 0x0005 | neuro | 8 | Planned | Chemical/field/substrate I/O |
| 0x0006 | arch | 11 | Functional | Structural plasticity (grow/prune/wire) |
| 0x0007 | orchestration | 8 | Planned | Model table, chaining, hot-reload |
| 0x0008 | lifecycle | 8 | Planned | Boot phases, tick/level reads, logging |
| 0x0009 | ipc | 8 | Planned | Inter-region signals, barriers |
| 0x000A | test | 8 | Planned | Assertions, snapshots, test harnesses |
212 total instructions. See specs/ for complete per-extension documentation.
48-byte header. Auto-detect deserializer reads both TVMR and legacy TERN formats.
HEADER (48 bytes)
Magic "TVMR" | Version | Flags
Extension table offset/count
Register defs offset/count
Instruction offset/count
xxhash64 checksum | Reserved
EXTENSION TABLE (8 bytes per dep)
REGISTER DEFS (variable, TypeId-based)
INSTRUCTIONS (8 bytes each)
Types are u16 IDs, portable across any TVMR engine implementation:
0x0000-0x00FF Primitives (void, bool, u8..u64, i8..i64, f16..f64)
0x0100-0x01FF Domain types (signal, packed_signal, chemical)
0x0200+ User types (future)
The core data type. Two bytes. Integer-only.
pub struct Signal {
pub polarity: i8, // -1, 0, or +1
pub magnitude: u8, // 0-255
}
// Effective value = polarity * magnitude
// Range: -255 to +255Temperature-gated signal flow:
for each output[i]:
accumulator = 0
for each input[j]:
signal = weights[i][j]
conductance = temperature[i][j].conductance() // 0.0-1.0
if conductance > activation_threshold:
accumulator += signal.polarity * signal.magnitude * input[j]
output[i] = accumulator
Weights that haven't been used cool down. Cold weights conduct less. This is biological — unused synapses weaken.
Pure integer adaptive learning. No gradients. No backpropagation.
- Peak-relative gating: Only neurons above
max_activation / 4participate - Sustained pressure: Changes require accumulated evidence via pressure registers
- Weaken before flip: Magnitude depletes before polarity changes
- Dopamine gating: Learning requires
chemical_state.dopamine >= threshold - Temperature lifecycle: HOT (learning) -> WARM -> COOL -> COLD (frozen)
[dependencies]
ternsig = "0.5"use ternsig::vm::{assemble, Interpreter};
let program = assemble(include_str!("region.ternsig"))?;
let mut vm = Interpreter::from_program(&program);
let output = vm.forward(&input)?;use ternsig::vm::{HotReloadManager, ReloadableInterpreter};
let mut reloadable = ReloadableInterpreter::new(program);
let mut reload_mgr = HotReloadManager::watch("firmware/")?;
// In your loop:
if let Some(event) = reload_mgr.poll() {
reloadable.reload(event)?;
}ternsig/
├── src/
│ ├── lib.rs # Crate root
│ ├── ternary.rs # Signal type and operations
│ ├── thermo.rs # Thermogram integration
│ ├── learning.rs # Mastery learning algorithm
│ ├── loader.rs # Program loader (.card format)
│ ├── validate.rs # Source validation
│ └── vm/
│ ├── mod.rs # VM module root, re-exports
│ ├── instruction.rs # 8-byte TVMR instruction format
│ ├── interpreter/ # Execution engine
│ ├── assembler.rs # .ternsig source -> program
│ ├── binary.rs # TVMR v1 binary serialization
│ ├── extension.rs # Extension trait, ExecutionContext
│ ├── registry.rs # ExtensionRegistry
│ ├── types.rs # Cross-language TypeId system
│ ├── validator.rs # Pre-execution validation
│ ├── hot_reload.rs # File-watching hot reload
│ └── extensions/ # Standard extension implementations
│ ├── tensor.rs # 0x0001: Matrix operations
│ ├── ternary.rs # 0x0002: Signal operations
│ ├── activation.rs # 0x0003: Activation functions
│ ├── learning.rs # 0x0004: Learning algorithms
│ ├── neuro.rs # 0x0005: Substrate I/O
│ ├── arch.rs # 0x0006: Structural plasticity
│ ├── orchestration.rs # 0x0007: Model management
│ ├── lifecycle.rs # 0x0008: Boot/phase management
│ ├── ipc.rs # 0x0009: Inter-region comms
│ └── test_ext.rs # 0x000A: Testing/assertions
├── specs/ # Instruction specifications
└── Cargo.toml
MIT OR Apache-2.0
