Per spec 2026-05-20-continuous-reasoning-trader-design.md §3.3 and §8:
decision_stride is REMOVED, not deprecated. No backwards-compat shim,
no fallback. Every consumer migrates in this commit per
feedback_no_partial_refactor.
Removed from:
- bin/fxt-backtest: RunArgs CLI flag, SweepBase field, SweepCell
override field, default_decision_stride() function, all three
BacktestHarnessConfig and RunArgs construction sites
- crates/ml-backtesting/src/harness.rs: BacktestHarnessConfig field,
MultiHorizonLoaderConfig decision_stride initializer, `let stride`
local, `if event_count % stride == 0` gate around
step_decision_with_latency; forward_step_into + step_decision now
share a single window-full guard (merged into one `if` block)
- crates/ml-alpha/src/data/loader.rs: MultiHorizonLoaderConfig field,
next_sequence stride logic simplified to stride=1 (consecutive
snapshots only)
- crates/ml-alpha/src/trainer/perception.rs: PerceptionTrainerConfig
field and Default impl; all four dt_s locals replaced with 1.0_f32
(training K-loop, graph-capture K-loop, forward_step_into CfC step,
eval K-loop)
- crates/ml-alpha/examples/alpha_train.rs: CLI flag, trainer_cfg and
both loader configs
- crates/ml/examples/alpha_baseline.rs: CLI flag, train + eval stride
gates replaced with unconditional read_all()
- config/ml/*.yaml: decision_stride: lines removed from
sweep_smoke, sweep_threshold_tuning, sweep_deployability,
sweep_decision_stride_example (file repurposed as generic example)
- tests: forward_step_golden, perception_overfit (×7 structs including
the stride=4 smoke repurposed as a second convergence check),
multi_horizon_loader (stride=4 spacing test repurposed as
ts_ns monotonicity check), ring3_replay, trainer_parity
Harness loop now invokes BOTH forward_step_into AND
step_decision_with_latency on every event whenever the snapshot window
is full. forward_step_into advances SSM state and writes alpha_probs_d;
step_decision_with_latency reads alpha_probs_d immediately after —
no CPU roundtrip, no stride gate.
n_decisions ≈ events_processed - seq_len + 1 after this commit
(vs ~9999 at stride=200 in the S2 baseline).
cargo check --workspace: clean
cargo test -p ml-backtesting --lib: 33 passed
cargo test -p ml-alpha --lib: 33 passed (6 ignored)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
112 lines
4.3 KiB
Rust
112 lines
4.3 KiB
Rust
//! Ring 1b — trainer/backtest parity check. See spec §8 Ring 1b.
|
|
//!
|
|
//! Verifies that the data-loading path used by the backtest harness
|
|
//! (`inference_only=true`) produces an Mbp10RawInput that is byte-equal
|
|
//! to what the trainer's loader produces from the same source.
|
|
//! Since both paths share the same loader code and same Mbp10Snapshot
|
|
//! → Mbp10RawInput conversion, the equality is structurally guaranteed —
|
|
//! but the test guards against any future refactor that breaks the claim.
|
|
//!
|
|
//! The full inference-output parity (asserting [probs; N_HORIZONS] from
|
|
//! the trunk's captured graph is bit-equal between training-time loader
|
|
//! and backtest-time loader paths) requires a real ml-alpha checkpoint,
|
|
//! which is deferred until a checkpoint-format is pinned in ml-alpha.
|
|
//! When a checkpoint is available, FOXHUNT_TEST_CKPT will gate the
|
|
//! gpu-inference assertion below.
|
|
|
|
use anyhow::Result;
|
|
use ml_alpha::data::loader::{
|
|
discover_mbp10_files_sorted, MultiHorizonLoader, MultiHorizonLoaderConfig,
|
|
};
|
|
use std::path::PathBuf;
|
|
|
|
fn try_loader(inference_only: bool) -> Option<MultiHorizonLoader> {
|
|
let root = std::env::var("FOXHUNT_TEST_DATA").ok()?;
|
|
let mbp10 = PathBuf::from(&root).join("ES.FUT");
|
|
if !mbp10.exists() {
|
|
return None;
|
|
}
|
|
let files = discover_mbp10_files_sorted(&mbp10).ok()?;
|
|
let cfg = MultiHorizonLoaderConfig {
|
|
files,
|
|
predecoded_dir: mbp10.clone(),
|
|
seq_len: 1,
|
|
horizons: [30, 100, 300, 1000, 6000],
|
|
n_max_sequences: 1,
|
|
seed: 0xCAFEF00D,
|
|
inference_only,
|
|
};
|
|
match MultiHorizonLoader::new(&cfg) {
|
|
Ok(l) => Some(l),
|
|
Err(e) => {
|
|
eprintln!("skipping: fixture data not usable ({e})");
|
|
None
|
|
}
|
|
}
|
|
}
|
|
|
|
#[test]
|
|
#[ignore = "requires populated FOXHUNT_TEST_DATA"]
|
|
fn peek_first_byte_equal_across_modes() -> Result<()> {
|
|
let Some(loader_train) = try_loader(false) else { return Ok(()); };
|
|
let Some(loader_bt) = try_loader(true) else { return Ok(()); };
|
|
|
|
let first_train = loader_train.peek_first()?;
|
|
let first_bt = loader_bt.peek_first()?;
|
|
|
|
// Inputs must be byte-equal (Mbp10RawInput is Pod via #[derive(Default)]
|
|
// — bid_px/sz/ask_px/sz arrays of f32, plus a handful of scalars).
|
|
assert_eq!(first_train.ts_ns, first_bt.ts_ns);
|
|
assert_eq!(first_train.prev_ts_ns, first_bt.prev_ts_ns);
|
|
for k in 0..10 {
|
|
assert_eq!(
|
|
first_train.bid_px[k].to_bits(),
|
|
first_bt.bid_px[k].to_bits(),
|
|
"bid_px[{k}] bit-mismatch between training and inference loader"
|
|
);
|
|
assert_eq!(first_train.bid_sz[k].to_bits(), first_bt.bid_sz[k].to_bits());
|
|
assert_eq!(first_train.ask_px[k].to_bits(), first_bt.ask_px[k].to_bits());
|
|
assert_eq!(first_train.ask_sz[k].to_bits(), first_bt.ask_sz[k].to_bits());
|
|
}
|
|
for k in 0..6 {
|
|
assert_eq!(
|
|
first_train.regime[k].to_bits(),
|
|
first_bt.regime[k].to_bits(),
|
|
"regime[{k}] bit-mismatch"
|
|
);
|
|
}
|
|
assert_eq!(first_train.prev_mid.to_bits(), first_bt.prev_mid.to_bits());
|
|
assert_eq!(first_train.trade_signed_vol.to_bits(), first_bt.trade_signed_vol.to_bits());
|
|
assert_eq!(first_train.trade_count, first_bt.trade_count);
|
|
|
|
Ok(())
|
|
}
|
|
|
|
#[test]
|
|
#[ignore = "requires populated FOXHUNT_TEST_DATA"]
|
|
fn inference_iteration_matches_chronological_snapshots() -> Result<()> {
|
|
let Some(mut loader) = try_loader(true) else { return Ok(()); };
|
|
|
|
// Pull 5 inference inputs; verify timestamps are monotone non-decreasing
|
|
// and each input's prev_ts_ns equals the previous input's ts_ns
|
|
// (after the first, which uses cur=prev semantics).
|
|
let mut prev_ts_ns = 0u64;
|
|
let mut first_iter = true;
|
|
for i in 0..5 {
|
|
let snap = loader.next_inference_input()?.expect("stream not exhausted");
|
|
if first_iter {
|
|
// peek_first/next-first uses cur==prev → prev_ts == ts.
|
|
assert_eq!(snap.prev_ts_ns, snap.ts_ns, "first snap has prev_ts == ts");
|
|
first_iter = false;
|
|
} else {
|
|
assert_eq!(
|
|
snap.prev_ts_ns, prev_ts_ns,
|
|
"snap {i}: prev_ts_ns should equal previous snap's ts_ns"
|
|
);
|
|
}
|
|
assert!(snap.ts_ns >= prev_ts_ns, "snap {i}: monotonic ts violated");
|
|
prev_ts_ns = snap.ts_ns;
|
|
}
|
|
Ok(())
|
|
}
|