Terasweep

Minimal Rust sweep runner for trillion-scale experiment permutations.

Upstream data is prepared by Limen; Terasweep only runs the hot sweep loop.

This package contains the minimum surface needed to repeat the cached sampled sweep:

• src/sweep.rs: orchestration, config, sampling, and summary output.
• src/results.rs: scoring, price accounting, fees, slippage, and metrics.
• src/ridge.rs: first model architecture.
• src/dlinear.rs: closed-form DLinear architecture.
• src/lightgbm.rs: native histogram GBDT architecture.
• src/tide.rs: minimal dense temporal encoder-decoder architecture.
• data/train.bin: frozen training matrix and target.
• data/test.bin: frozen test matrix and target.
• data/price.bin: frozen open/close backtest prices.
• configs/sweep.json: deterministic sampled sweep space.
• configs/dlinear.json: deterministic DLinear sampled sweep space.
• configs/lightgbm.json: deterministic LightGBM-style sampled sweep space.
• configs/tide.json: deterministic TiDE sampled sweep space.
• golden/1m.json: stable canonical 1M validation result.

Build and smoke test:

make smoke

Run a larger sweep:

make run RUNS=1000000 OUT=runs/1m
make run RUNS=10000000 OUT=runs/10m
make run RUNS=100000000 OUT=runs/100m
make run RUNS=1000000000 OUT=runs/1b

Run DLinear on the same frozen data:

make run RUNS=1000000 OUT=runs/dlinear-1m CONFIG=configs/dlinear.json

Run LightGBM-style GBDT on the same frozen data:

make run RUNS=1000000 OUT=runs/lightgbm-1m CONFIG=configs/lightgbm.json

Run TiDE on the same frozen data:

make run RUNS=1000000 OUT=runs/tide-1m CONFIG=configs/tide.json

Golden validation:

make verify

The runner writes summary.json and time.txt under the selected run directory.

Execution-level comparison

See Ridge execution levels for compact caching, Rust AVX-512 intrinsics, handwritten assembly, validation, and exact-plan replay. make now preserves the historical executable and builds target/terasweep-funnel.