L-S51: Trinity Loss — φ-prior-aware ternary contrastive loss for JEPA-T

[gHashTag]

Overview

Implement the Trinity loss function: a φ-prior-aware ternary contrastive loss for JEPA-T training.

Formula

sim(a,b)       = dot_ternary(a,b) / 64
L_triplet      = max(0, margin + sim(a,n) - sim(a,p))   [margin=0.5]
L_phi_prior    = phi_inv_sq * (zero_count(a) + zero_count(p) + zero_count(n)) / 192   [phi_inv_sq≈0.382]
L_total        = L_triplet + lambda * L_phi_prior   [lambda=0.1]

Deliverables

• crates/trinity_loss/Cargo.toml
• crates/trinity_loss/src/lib.rs — 5 public functions: dot_ternary, sim, zero_count, phi_prior_term, trinity_loss
• crates/trinity_loss/tests/loss.rs — 10 hand-computed + 50 LFSR stability tests
• python_ref/trinity_loss_ref.py — reference only, not in cargo build
• README.md

Constraints

• Apache-2.0
• Author: Dmitrii Vasilev admin@t27.ai
• Rust-only in cargo build (R1 CROWN)
• All functions deterministic, no allocations beyond Vec init, no std::time

References

• DOI: 10.5281/zenodo.19227877
• φ² + φ⁻² = 3; φ⁻² ≈ 0.382 used as soft margin anchor