Skip to main content Crate ternlang_ml Copy item path Source coherence spectra_compat BenchmarkResult Summary statistics for a benchmark run. CoalitionMember One agent’s vote in a coalition. CoalitionResult Coalition voting statistics. DeliberationEngine Multi-round evidence accumulation engine. DeliberationResult Result of a full deliberation run. DeliberationRound One round of a deliberation trace. GateDimension One dimension in an action gate check. GateResult Result of an action gate evaluation. HallucinationScore Measures internal consistency of evidence signals about a claim. ScalarTemperature Maps a ternary decision to a recommended LLM sampling temperature. TernaryMLP A 2-layer ternary multi-layer perceptron. TimedResult Wall-clock timed benchmark result for one matrix size. TritBlock A single Transformer block (Attention + FeedForward). TritEvidenceVec A named, weighted multi-dimensional evidence vector. TritMatrix A flat row-major ternary matrix (rows × cols). TritScalar A continuous ternary confidence scalar, clamped to [-1.0, +1.0]. TritTransformer The full TritTransformer model. TritTransformerConfig GateVerdict The outcome of an action gate evaluation. TEND_BOUNDARY Zone boundary: 1/3 of the full scale. action_gate Evaluate an action through a multi-dimension policy gate. benchmark bitnet_matrix Generate a TritMatrix with exactly target_sparsity fraction of zero entries. bitnet_threshold Compute the BitNet-style threshold: 0.5 × mean(|weights|) coalition_vote Aggregate a coalition of agent votes into a single ternary decision. dense_matmul Dense ternary matrix multiply: C = A × B
No skipping — every element is computed regardless of zero state.
Use this as the baseline for benchmark comparisons. evaluate Evaluate MLP accuracy on a dataset.
Returns (correct, total, accuracy). hallucination_score linear BitNet-style ternary linear layer: output = sparse_matmul(input, W) majority Majority vote across a row of trits — reduces a vector to one trit.
Returns the sign of the sum: positive majority → +1, negative → -1, tie → 0. parity_dataset 3-bit parity dataset: 8 inputs → label 0 (even parity) or 1 (odd parity) print_benchmark_table Print a formatted benchmark table to stdout. quantize Quantize a slice of f32 weights to balanced ternary using threshold tau. scalar_temperature sparse_matmul Sparse ternary matrix multiply: C = A × B, skipping zero-weight elements. timed_benchmark Run timed dense vs sparse matmul across multiple square matrix sizes. timed_benchmark_at_sparsity Benchmark at an arbitrary target sparsity (0.0 = dense, 1.0 = all zeros). timed_benchmark_bitnet Benchmark at a given sparsity level. trit_activation Ternary threshold activation: maps accumulator trit to output trit.
sign(x): +1 → +1, 0 → 0, -1 → -1. Identity on Trit — but useful as a
named function to clarify intent in MLP forward passes. trit_from_f32_approx trit_to_f32 xor_dataset All 4 XOR inputs as ternary rows: {-1,+1} × {-1,+1} → {-1,+1}
Input encoding: -1 = False, +1 = True