v0.1.1 -- Mathematical Correctness Hardening

Summary

F.R.I.D.A.Y. 3-body parallel review found 34 mathematical bugs across 11 source files. All fixed with 3-round review-fix loop (2 consecutive CLEAN).

Critical (5)

• omega Nyquist aliasing -- high-k oscillators silently aliased, halving representational capacity
• alpha tensor reuse in associative_scan -- potential in-place mutation corrupting imaginary decay
• W=0 all-inf mask producing NaN in SDPA
• Dead StateTokenAdapter consuming optimizer budget
• trn_ratio=1.0 (float) treated as int 1 instead of 100% TRN

High (9)

• L-inf normalization replaced with complex modulus (phase preservation)
• bf16 latch bug in both forward() and step_single() paths
• Cumsum double-correction producing O(1) error instead of O(alpha_cum)
• SafeCumprod backward contract violation
• gate_bias_init default mismatch (0.85 vs 0.65)
• PE truncation guard for both forward paths
• Backward-compat shim for retrieval_temperature rename

Medium/Low (20)

Amplitude gradient vanishing, odd d_model PE, deque.sort() crash, std Bessel correction, assert->ValueError, thread safety docs, and more.

Benchmark (RTX 5090, synthetic data)

Config	Params	Final Loss	Throughput	Stability
Toy (d=128, K=64, L=2)	509K	5.5456	93K tok/s	std=0.001
TRN-100M (d=512, K=256, L=8)	40.5M	10.3879	6.8K tok/s	--

Breaking Changes

• Existing checkpoints are incompatible due to normalization (H2) and amplitude (M1) changes. Retrain from scratch.
• retrieval_temperature parameter renamed to retrieval_sharpness (backward-compat shim provided).

v0.1.0 -- initial public release

TriMemory v0.1.0

Three-path memory architecture for LLM agents.

What's included

• TriMemoryEngine: KV window + retrieval index + TRN recurrent state with 3-way softmax gate
• AgentMemory: Stateful per-token streaming wrapper (save/load across turns)
• TRNModel: Standalone recurrent model with O(1) memory per step
• PolicyBench N=10: Document authority QA benchmark (Japanese corporate IT security policy)
• Config presets: toy (1 KB), trn_100m (16 KB), trn_400m (64 KB), trn_1b (96 KB)

Benchmark highlights

• Config D (Full) composite: 0.676, D >= max(A,B,C) in 10/10 seeds
• TRN throughput: 231 tps at T=10,000 (vs 15.5 tps Transformer+KV)
• TRN state: 8 KB constant (vs 156 MB KV cache at T=10,000)
• 1,000 agents: 15.6 MB TRN total vs 31 GB KV (fp32)

Known limitations

• Toy-scale models only (1--100M params). Scaling to 1B+ is untested.
• TRN selective copy accuracy: 8.8% (vs Transformer 96.2%). Needle-in-Haystack: 0.0.
• Alpha status. No production deployment.

277 tests. Python 3.10+. PyTorch 2.1+. Apache-2.0.