Feature layout

Source: packages/features. Requirements: SRD §4, spec 001 FR-010–FR-012.

Both model variants read the same tensors; they differ only in whether the encoder may look ahead. Two feature sets exist: raw (below, iterations 1–2 phase A) and derived (raw plus three channels, below).

Inputs (raw set) — [N, 240, 10]

One row per 10 s sample, padded with zero rows to 240.

#	Channel	Encoding
0–1	bearing_sin, bearing_cos	sin/cos of the observed bearing; 0, 0 at a dropout
2–3	own_x_scaled, own_y_scaled	ownship position in km ÷ position_scale_km
4–5	own_course_sin, own_course_cos	ownship course
6	own_speed_scaled	ownship speed ÷ speed_scale_mps
7	time_scaled	seconds since window start ÷ time_scale_s (not normalised to window length)
8	valid	1 if a bearing was observed
9	window_mask	1 inside the window, 0 in padding

Derived feature set

Requirements: SRD §4 (derived features reserved for when raw features stall), spec 003 US3, research R6. Source: packages/features/src/derived.ts. Built with tma-features norm|build --feature-set derived; the default is raw.

The derived set is the 10 raw channels followed by three more, so its inputs are [N, 240, 13]:

#	Channel	Encoding
10	bearing_rate_scaled	trailing least-squares bearing rate ÷ bearing_rate_scale_rad_s
11	rate_valid	1 if the rate was computed (≥ 3 valid samples in its window)
12	intercept_scaled	bearing-line intercept in km ÷ position_scale_km

All three are 0 on dropout rows (no bearing) and on padding rows.

Bearing rate. At row t with a bearing, take the valid samples k among the trailing 18 rows (t − 17 … t, 3 min at 10 s). With τₖ = tₖ − tₜ and uₖ the bearing difference bₖ − bₜ wrapped to [−π, π] (so a track crossing north has no 2π jump), the rate is the least-squares slope

ḃ = (m Σ τu − Σ τ Σ u) / (m Σ τ² − (Σ τ)²), m = number of valid samples.

With m < 3 the rate is 0 and rate_valid is 0. A 10 s finite difference would be useless at σ = 2° (≈ 0.28°/s of noise against true rates of ≈ 0.005–0.3°/s); the 18-sample slope has ≈ 0.009°/s of noise at σ = 2°.

Bearing-line intercept. c = x_o cos b − y_o sin b (ownship x east, y north; b clockwise from north). A target at (x, y) on the bearing line satisfies x cos b − y sin b = c, the pseudo-linear measurement: with constant target velocity, every row is one linear equation in the target's initial position and velocity given sin b, cos b, t and c. This is the structure a classical pseudo-linear estimator uses, offered to the network directly.

Causality. Both channels at row t read only rows ≤ t (the rate's window is trailing; the intercept uses row t alone), so the instantaneous models stay causal. A test perturbs every row after t (bearings and dropout flags) and checks that rows ≤ t are byte-identical.

Normalisation, inputs only. bearing_rate_scale_rad_s is the 90th percentile (nearest rank) of |ḃ| over every train row that has a rate, float32. It reads bearings, dropout flags and times only; a counterfactual test scrambles target truth and checks that the constant and every derived input are byte-identical. build --feature-set derived refuses a norm file without the constant. On the iteration-2 training split (seed 201) it is 0.00161 rad/s (0.092°/s).

One set per model. The layout JSON names the feature set and its channel list. A model's input width is its features file's; training refuses a file whose set differs from the config's feature_set, and prediction refuses a file of another set than the model was trained on.

Targets — [N, 240, 6]

log_range_km (natural log), course_sin, course_cos, speed_scaled (÷ speed_scale_mps), x_km, y_km (target position, reported only).

Loss mask — [N, 240]

1 inside the window after the zig (rows > zig_index), else 0. Before the zig the range is unobservable, so penalising it would only teach the prior.

Normalisation (inputs only)

position_scale_km, speed_scale_mps and time_scale_s are the maxima of |ownship x|, |ownship y|, ownship speed and timestamp over the train split. Target truth is never read. A test shuffles target tracks across scenarios and checks that the constants do not change.

speed_scale_mps comes from ownship speeds, but it also scales the target-speed target. Target speeds go up to about 1.33 × the scale. This does no harm: the constant is still computed from inputs only, and the output head is linear, so nothing clips (G2 panel, round 1).

File

.features.bin: magic TMAFEAT\0, version, FeatureLayout JSON, NormalisationConstants JSON, then (format 2) the source dataset's DatasetHeader JSON and the SHA-256 of the dataset file, then scenario indices, then the three tensors as little-endian float32, each scenario-major. Format 2 (iteration 2) carries the dataset's provenance into training, where tma-train checks it; the layout names its feature_set (raw or derived) and input channels, and the inputs tensor is [N, 240, 10] or [N, 240, 13] accordingly. Format-1 files (iteration 1) are still read, but training refuses them (lineage-missing).