Skip to main content

Critical Bug Fix: Synaptic Contribution Calculation

Date: November 4, 2025
Severity: 🔴 CRITICAL
Status: ✅ FIXED
Impact: Synaptic propagation and sensory injection

Bug Description​

During Phase 2 refactoring, I introduced a critical behavioral change in synaptic contribution calculation that broke synaptic propagation.

Root Cause​

Incorrect assumption: I assumed synaptic weights were normalized (0.0-1.0 range)
Actual behavior: FEAGI uses direct cast (0-65,025 range)

Impact​

Before Fix (WRONG):

// feagi-synapse v1 (BROKEN)
let w = weight as f32 / 255.0;        // 255 → 1.0 (NORMALIZED)
let c = conductance as f32 / 255.0;   // 255 → 1.0 (NORMALIZED)
contribution = w * c * sign;          // 1.0 × 1.0 = 1.0

After Fix (CORRECT):

// feagi-synapse v2 (FIXED)
let w = weight as f32;                // 255 → 255.0 (DIRECT CAST)
let c = conductance as f32;           // 255 → 255.0 (DIRECT CAST)
contribution = w * c * sign;          // 255.0 × 255.0 = 65,025.0

Magnitude Difference: 65,025× smaller with bug!

Why This Broke Everything​

Synaptic Propagation​

  • Synaptic contributions were 65,000× too weak
  • Neurons never received enough input to fire
  • Network appeared "dead"

Sensory Injection​

  • Same calculation used for sensory input
  • Injected potentials 65,000× too weak
  • Sensors had no effect on neurons

Fix Applied​

File: feagi-core/crates/feagi-synapse/src/contribution.rs

Changed: Lines 56-58

// WRONG (introduced in Phase 2):
let w = weight as f32 / 255.0;  // ❌ Normalization
let c = conductance as f32 / 255.0;

// CORRECT (matches original FEAGI):
let w = weight as f32;  // ✅ Direct cast
let c = conductance as f32;

Tests Updated: Lines 120-142 (test expectations corrected)

Canonical Contract (Determinism Across Backends)​

To keep FEAGI deterministic across CPU, WGPU, and CUDA backends, synaptic math must follow one canonical contract:

  • weight: stored as u8 (0..255), treated as absolute units
  • postsynaptic potential (PSP / conductance): stored as u8 (0..255), treated as absolute units
  • contribution:
    • (contribution = sign \times weight \times psp)
    • Range per synapse: ([-65{,}025, +65{,}025])
  • No implicit normalization (/255.0) or scaling (*255.0) inside synaptic compute loops.
    • Conversions are allowed only at IO boundaries (genome/API/UI) and must be explicit.

Verification​

Test Results​

$ cargo test -p feagi-synapse --lib
running 11 tests
...........
test result: ok. 11 passed ✅

Integration Test​

$ cargo test -p feagi-burst-engine test_synaptic --lib
running 1 test
test synaptic_propagation::tests::test_synaptic_propagation ... ok ✅

Build Status​

$ cd feagi && cargo build --release
    Finished `release` profile [optimized] target(s) in 1m 47s ✅

Lesson Learned​

Always verify behavioral equivalence when refactoring:

  • ✅ Check tests pass
  • ✅ Check function signatures match
  • ⚠️ Check actual numeric output (this was missed!)

Why tests didn't catch it:

  • New tests in feagi-synapse were written with wrong assumptions
  • Integration tests don't verify exact magnitudes
  • Need end-to-end validation with real connectomes

Action Items​

  1. ✅ Fixed synaptic contribution formula
  2. ✅ Updated tests
  3. ✅ Updated documentation
  4. ✅ Rebuilt main application
  5. ⏳ User needs to retest

Status: Bug fixed and verified. Ready for testing again.