Skip to content

SPEC.md

Latest commit

 

History

History
85 lines (65 loc) · 2.74 KB

SPEC.md

File metadata and controls

85 lines (65 loc) · 2.74 KB

QRES v19 Protocol Specification

Overview

QRES v19 is a protocol for decentralized neural consensus. While it produces .qres artifacts (saved model bytecode packets), its primary function is defining the SwarmNeuron trait for behavior and the Gossip headers used for model bytecode propagation.

Core Specifications

  1. The Neuron Protocol: Defines how nodes compute residuals (I16F16) and BFP gradients (Bfp16Vec).
  2. The Model Bytecode Format: A bytecode serialization standard for transmitting learned strategies across the gossip network.
  3. Consensus: A deterministic, reputation-weighted agreement mechanism using Trimmed Mean Aggregation.

2. Header Structure (24 bytes)

Offset Length Type Description
0 4 [u8; 4] Magic Bytes (QRES)
4 2 u16 Major Version (19)
6 2 u16 Minor Version (0)
8 4 u32 Flags (Bitmask)
12 8 u64 Total Uncompressed Size
20 4 u32 Header Checksum (CRC32)

Flags

  • 0x01: Solid Archive (Single stream, no random access)
  • 0x02: Encrypted (AES-256-GCM)
  • 0x04: Checksummed (Each block has CRC32)

3. Block Structure

A QRES file is a stream of Blocks. Blocks are either "Epiphanies" (Model Updates) or "Residuals" (Data).

3.1 Epiphany Block (Type 0x0E)

Contains new weights for the predictor model.

Field Size Details
Block ID 1 byte 0x0E
Length 2 bytes Size of weight payload
Predictor ID 1 byte ID of predictor to update (e.g. 1=Linear)
Weights N bytes Q16.16 fixed-point weights

3.2 Residual Block (Type 0x0D)

Contains compressed residuals (prediction errors).

Field Size Details
Block ID 1 byte 0x0D
Compressed Len 4 bytes Size of bit-packed payload

3.3 Summary Bytecode Packet (Type 0x13) - v19.0

Contains rapid onboarding state.

Field Size Details
Block ID 1 byte 0x13
Round 8 bytes u64 Round Index
Hash 32 bytes History Hash
Consensus N bytes BFP-16 Encoded Vector
Variance N bytes BFP-16 Encoded Vector

4. Deterministic Math Specification

4.1 State Consensus (Q16.16)

All consensus state must use i32 fixed point:

// Current prediction value (16 bits integer, 16 bits fraction)
type Q16 = i32;
// 0.1 + 0.2 = round(6553.6 + 13107.2) = 19661

4.2 Gradient Updates (BFP-16) - v19.0

All gradients must use Block Floating Point to preserve dynamic range at low learning rates:

struct Bfp16Vec {
    exponent: i8,      // Shared exponent
    mantissas: Vec<i16> // Signed 16-bit integers
}
// Value[i] = mantissa[i] * 2^(exponent)