Data dictionary

Generated

Generated from schema/tma.yaml at build time by gen-markdown-datadict. Do not edit; change the schema.

Units

Class	Slot	Unit (UCUM)
BearingErrorModel	ar_sd_rad	rad
BearingErrorModel	ar_tau_s	s
BearingErrorModel	bias_sd_rad	rad
BearingErrorModel	bias_tau_s	s
CurvePoint	t_since_zig_s	s
DetectionRow	timestamp_s	s
MetricsRecord	position_window_m_median	m
MetricsRecord	position_window_m_p90	m
NaiveMeans	speed_mps	m/s
NormalisationConstants	bearing_rate_scale_rad_s	rad/s
Observation	bearing_rad	rad
Observation	timestamp_s	s
ParameterRanges	target_manoeuvre_after_zig_s	s
ParameterRanges	target_manoeuvre_margin_s	s
ParameterRanges	target_manoeuvre_spacing_s	s
ParameterRanges	zig2_delay_s	s
PlatformState	course_rad	rad
PlatformState	speed_mps	m/s
PlatformState	timestamp_s	s
PlatformState	x_m	m
PlatformState	y_m	m
PredictionRow	course_rad	rad
PredictionRow	course_sd_rad	rad
PredictionRow	range_m	m
PredictionRow	speed_mps	m/s
PredictionRow	speed_sd_mps	m/s
PredictionRow	timestamp_s	s
PredictionRow	vx_sd_mps	m/s
PredictionRow	vy_sd_mps	m/s
PredictionRow	x_m	m
PredictionRow	y_m	m
ScenarioMetadata	mean_range_m	m
ScenarioMetadata	min_range_m	m
ScenarioMetadata	zig_time_s	s
ScenarioParameters	dt_s	s
ScenarioParameters	noise_sigma_rad	rad
ScenarioParameters	ownship_course0_rad	rad
ScenarioParameters	ownship_speed_mps	m/s
ScenarioParameters	ownship_turn_rate_rad_s	rad/s
ScenarioParameters	target_bearing0_rad	rad
ScenarioParameters	target_course_rad	rad
ScenarioParameters	target_manoeuvre_angle_rad	rad
ScenarioParameters	target_manoeuvre_time_s	s
ScenarioParameters	target_range0_m	m
ScenarioParameters	target_speed_mps	m/s
ScenarioParameters	target_turn_rate_rad_s	rad/s
ScenarioParameters	window_s	s
ScenarioParameters	zig2_angle_rad	rad
ScenarioParameters	zig2_delay_s	s
ScenarioParameters	zig_angle_rad	rad
TargetManoeuvre	start_time_s	s

metamodel version: 1.12.0

version: 0.21.0

Single source of truth for every data type in the TMA via ML proof of concept (constitution principle I). Frame: local flat-earth Cartesian, x east, y north, metres, origin at ownship's window-start position. Bearings and courses are true, clockwise from north, radians in [0, 2π). Speeds in m/s. Times in seconds since window start. All stored floats are float32-rounded.

ERD Diagrams

Component 1 (BearingErrorModel, Continuation, DatasetHeader...)

Component 2 (ChangeDecision, DetectorThreshold, ModelDigest...)

Component 3 (Detection, DetectionRow)

Component 4 (DetectionHeader, PredictionHeader, PredictionInput)

Component 5 (GateCheck, GateVerdict)

Component 6 (GateRecomputation, GateScoring)

Component 7 (NaiveMeans, NaiveModel)

Component 8 (Prediction, PredictionRow)

Standalone Classes

These classes are completely isolated with no relationships and are not used as base classes:

Class	Description
CurvePoint	One point of an error-vs-time-since-zig curve (curves.csv row).
FeatureLayout	Documents the model tensors (not stored as records). inputs [N, 240, 10]: bearing_sin, bearing_cos, own_x_km, own_y_km, own_course_sin, own_course_cos, own_speed_scaled, time_scaled, valid, window_mask; the derived feature set appends bearing_rate_scaled, rate_valid, intercept_scaled [N, 240, 13]. targets [N, 240, 6]: log_range_km, course_sin, course_cos, speed_scaled, x_km, y_km. loss_mask [N, 240]: window_mask and t ≥ zig_time. Features file format 2 also embeds the source DatasetHeader and the source dataset SHA-256.
MetricsRecord	Metric values for one estimator on one slice at one evaluation point.
PairedComparison	A model and a reference estimator compared on rows both answered (converged and finite), with each estimator's coverage reported separately (spec 003 R9).

Classes

BearingErrorModel

Realistic bearing error (generator 0.6.0, SRD Annex C §C.14), added to the white noise of noise_sigma_rad: a first-order autoregressive (Gauss–Markov) term and a slowly varying bias, itself a Gauss–Markov process with a long time constant. Both start in their stationary distribution and are drawn every sample (dropped ones included) from the scenario's own stream (sub-index 2), so the geometry and the white noise are those of 0.5.0 for the same seed.

Attributes

Name	Cardinality:	Type	Description
ar_sd_rad	1..1	float	Stationary sd of the correlated term.
ar_tau_s	1..1	float	Time constant of the correlated term.
bias_sd_rad	1..1	float	Stationary sd of the drifting bias.
bias_tau_s	1..1	float	Time constant of the drifting bias.

Referenced by:

• ParameterRanges : parameterRanges__bearing_error _0..1
• ScenarioParameters : scenarioParameters__bearing_error _0..1

ChangeDecision

The single change of the iteration's change round, fixed from validation evidence and committed before phase B trains (spec 003 R5).

Attributes

Name	Cardinality:	Type	Description
calibrations	0..*	UncertaintyCalibration	The frozen post-hoc interval multipliers of each estimator and output a G7 decision names (schema 0.20.0, spec 018), fitted on validation; tma-g7 refuses others.
carried_model	0..1	string	G4: the manifest id of the one carried network (Annex B §B.4); tma-g4 refuses network predictions from another model.
carried_models	0..*	string	G5 (spec 012, schema 0.16.0): the manifest ids of the carried trackers (one or both of T1 and T2, Annex C §C.5), each with its own plan in plans, in the same order; tma-g5 refuses tracker predictions from any other model.
carried_set_rule	1..1	string
change	1..1	string	e.g. feature_set: derived
decided_at	1..1	string
detector_thresholds	0..*	DetectorThreshold	The frozen alarm threshold of each detector a G6 decision names (schema 0.19.0, spec 017), set on validation at the matched false-alarm rate; tma-g6 refuses others.
fresh_test_count	0..1	integer	The dataset count whose test split a re-score scores (the plan's dataset.fresh_test.count; schema 0.11.1); tma-g2 refuses a plan with another count.
git_commit	1..1	string
model_digests	0..*	ModelDigest	The frozen carried models a re-score scores, by manifest id, with the SHA-256 of each one's weights.bin, manifest.json and model.json; tma-g2 refuses a re-score whose carried set or any of those files differs.
phase_a_experiment	1..1	string
plan	0..1	string	The plan this decision pre-registers (configs/<…>/plan.json). Required from iteration 3 (spec 007): tma-g2 scores a test split only with the decision that names its plan.
plans	0..*	string	G5: the plan of each carried tracker (configs/phase4//plan.json), in the order of carried_models; they share one dataset and one fresh test seed.
rationale	1..1	string	Validation evidence cited.
reference_estimator	0..1	string	G4 (spec 010, schema 0.13.0): the frozen reference tracker's estimator name, which encodes every tuning parameter (e.g. imm-cv0-ct0.03-w0.003-s0.0001-r4); tma-g4 refuses reference predictions with another name.
rescore_of	0..1	string	The experiment directory of the inconclusive gate this decision re-scores (SRD §7: one re-score on a larger fresh-seed test set; e.g. experiments/20261024-i4-g2). When present, tma-g2 runs in re-score mode (--rescore-of): the frozen carried models are bound by model_digests instead of by training ancestry, the re-scored gate must be inconclusive and not itself a re-score, and the fresh test seed must differ from its test seed.
stalled	1..1	boolean	Any gated criterion failed or was inconclusive on validation.
test_seed	0..1	string	Dataset seed of the test split the decision commits to (the plan's fresh_test seed, or its dataset seed); tma-g2 checks it against the plan.

Continuation

The committed model a seeded continuation started from (scripts/train-continue.mjs; spec 003 decision 14, spec 004 decision 15). Its epoch records up to best_epoch are kept in the continued run's manifest, with their val_loss as recorded under the earlier config.

Attributes

Name	Cardinality:	Type	Description
best_epoch	1..1	integer
git_commit	0..1	string	The source model's training-code commit, when it recorded one.
rescored_val_loss	0..1	float	The source's best weights re-scored under the continued run's loss, written by tma-train when it resumed (spec 004 FR-002). With best_epoch and best_val_loss it makes the kept weights auditable: a run whose best_epoch is the source's must have best_val_loss equal to this value and weights_digest equal to the source's.
val_loss	1..1	float	The source's best-epoch validation loss under the source's config.
weights_digest	1..1	string	SHA-256 of the weights the continuation resumed from (the source's best epoch).

Referenced by:

• ModelManifest : modelManifest__continued_from _0..1

CurvePoint

One point of an error-vs-time-since-zig curve (curves.csv row).

Attributes

Name	Cardinality:	Type	Description
course_deg_median	0..1	float
estimator	1..1	string
gross_error_rate	1..1	float
n	1..1	integer
non_converged_rate	0..1	float
range_rel_median	0..1	float	Absent when not finite (more than half the rows non-converged), as in MetricsRecord.
range_rel_p90	0..1	float
speed_kt_median	0..1	float
t_since_zig_s	1..1	float

DatasetHeader

First record of every dataset file.

Attributes

Name	Cardinality:	Type	Description
created_by	1..1	string	Command line that produced the file.
dataset_seed	1..1	string	Decimal uint64 string.
generator_git_commit	0..1	string	Commit of the checkout that ran tma-generate (absent before schema 0.8.0).
generator_git_dirty	0..1	boolean	True when that checkout had uncommitted changes (only allowed with --allow-dirty; absent before schema 0.8.0).
generator_version	1..1	string
index_end	1..1	integer	Last scenario index (exclusive).
index_start	1..1	integer	First scenario index (inclusive).
parameter_ranges	1..1	ParameterRanges
record	1..1	RecordKind
redraw_count	1..1	integer	Geometries rejected for violating min_range_m.
scenario_count	1..1	integer
schema_version	1..1	string
split	1..1	Split

Referenced by:

• DatasetRef : datasetRef__header _1..1
• ModelManifest : modelManifest__source_dataset_header _0..1

DatasetRef

Attributes

Name	Cardinality:	Type	Description
digest	0..1	string	SHA-256 of the dataset file (absent in manifests written before schema 0.4.0).
header	1..1	DatasetHeader
path	1..1	string

Referenced by:

• ExperimentManifest : experimentManifest__datasets _1..*

Detection

One detector's per-timestep scores for one scenario (schema 0.19.0).

Attributes

Name	Cardinality:	Type	Description
detector	1..1	string
record	1..1	RecordKind
rows	1..*	DetectionRow
scenario_index	1..1	integer

DetectionHeader

First record of every detection file (schema 0.19.0).

Attributes

Name	Cardinality:	Type	Description
dataset_digest	1..1	string	SHA-256 of the truth file.
detector	1..1	string
inputs	0..*	PredictionInput	The prediction files the scores were computed from, with their SHA-256s.
record	1..1	RecordKind
schema_version	1..1	string

DetectionRow

One detector's score at one timestep (schema 0.19.0, spec 017). Larger means a manoeuvre is more likely; an alarm is the score at or above the detector's threshold.

Attributes

Name	Cardinality:	Type	Description
score	1..1	float	The detector's statistic at this timestep (units depend on the detector).
timestamp_s	1..1	float

Referenced by:

• Detection : detection__rows _1..*

DetectorThreshold

One detector's frozen alarm threshold (schema 0.19.0, spec 017).

Attributes

Name	Cardinality:	Type	Description
detector	1..1	string
threshold	1..1	float

Referenced by:

• ChangeDecision : changeDecision__detector_thresholds _0..*

EpochRecord

Attributes

Name	Cardinality:	Type	Description
epoch	1..1	integer
seconds	1..1	float	Cumulative wall time.
train_loss	1..1	float
val_loss	1..1	float

Referenced by:

• ModelManifest : modelManifest__epochs _1..*

EstimatorRef

Attributes

Name	Cardinality:	Type	Description
name	1..1	string
settings_json	1..1	string	Estimator settings as JSON.

Referenced by:

• ExperimentManifest : experimentManifest__estimators _1..*

ExperimentManifest

Everything needed to reproduce an experiment run.

Attributes

Name	Cardinality:	Type	Description
id	1..1	string
command_lines	1..*	string
created	1..1	string	ISO 8601 timestamp.
datasets	1..*	DatasetRef
estimators	1..*	EstimatorRef
generator_version	1..1	string
git_commit	1..1	string
git_dirty	0..1	boolean	True when the working tree had uncommitted changes (absent before schema 0.4.0).
node_version	0..1	string	Node.js version that ran the experiment (byte identity is promised per pinned Node).
normalisation	0..1	NormalisationConstants
schema_version	1..1	string

FeatureLayout

Documents the model tensors (not stored as records). inputs [N, 240, 10]: bearing_sin, bearing_cos, own_x_km, own_y_km, own_course_sin, own_course_cos, own_speed_scaled, time_scaled, valid, window_mask; the derived feature set appends bearing_rate_scaled, rate_valid, intercept_scaled [N, 240, 13]. targets [N, 240, 6]: log_range_km, course_sin, course_cos, speed_scaled, x_km, y_km. loss_mask [N, 240]: window_mask and t ≥ zig_time. Features file format 2 also embeds the source DatasetHeader and the source dataset SHA-256.

Attributes

Name	Cardinality:	Type	Description
feature_set	0..1	FeatureSet	Absent means raw.
input_channels	1..*	string
rows	1..1	integer
target_channels	1..*	string

FloatRange

Closed interval [min, max].

Attributes

Name	Cardinality:	Type	Description
max	1..1	float
min	1..1	float

Referenced by:

• ParameterRanges : parameterRanges__dropout_fraction _1..1
• ParameterRanges : parameterRanges__los_offset_rad _0..1
• ParameterRanges : parameterRanges__ownship_speed_mps _1..1
• ParameterRanges : parameterRanges__ownship_turn_rate_rad_s _0..1
• ParameterRanges : parameterRanges__target_manoeuvre_angle_abs_rad _0..1
• ParameterRanges : parameterRanges__target_range0_m _1..1
• ParameterRanges : parameterRanges__target_speed_change_abs_mps _0..1
• ParameterRanges : parameterRanges__target_speed_change_rate_mps2 _0..1
• ParameterRanges : parameterRanges__target_speed_mps _1..1
• ParameterRanges : parameterRanges__target_turn_rate_rad_s _0..1
• ParameterRanges : parameterRanges__window_s _1..1
• ParameterRanges : parameterRanges__zig2_delay_s _0..1
• ParameterRanges : parameterRanges__zig_angle_abs_rad _1..1
• ParameterRanges : parameterRanges__zig_fraction _1..1

GateCheck

Attributes

Name	Cardinality:	Type	Description
name	1..1	string
ci_high	0..1	float	Upper bound of the bootstrap interval.
ci_level	0..1	float	Interval coverage, e.g. 0.95.
ci_low	0..1	float	Lower bound of the bootstrap interval.
detail	1..1	string
result	1..1	GateResult
threshold	0..1	float	Limit the statistic is tested against.
value	0..1	float	Statistic tested (e.g. median range error).

Referenced by:

• GateVerdict : gateVerdict__checks _1..*

GateRecomputation

One estimator re-computed by a gate on a sample of scenarios (schema 0.18.0).

Attributes

Name	Cardinality:	Type	Description
estimator	1..1	string
sample_sha256	1..1	string	SHA-256 (hex) of the sampled scenario indices, comma-separated in ascending order.
scenarios	1..1	integer	Number of scenarios re-computed.

Referenced by:

• GateScoring : gateScoring__recomputed _0..*

GateScoring

What a test-split gate verdict was scored against (schema 0.18.0; tma-g5 writes it as scoring.json beside verdict.json). tma-g5 refuses to score a decision and test seed that another experiment's scoring.json already records, unless it names that experiment (repeat_of).

Attributes

Name	Cardinality:	Type	Description
decision	1..1	string	Path of the decision record, from the repository root.
decision_sha256	1..1	string	SHA-256 (hex) of the decision record.
gate	1..1	GateId
recomputed	0..*	GateRecomputation	The estimators whose rows the gate re-computed on a seeded sample of scenarios from their declared settings, byte-identical to their files (absent in records backfilled for experiments scored before schema 0.18.0).
repeat_of	0..1	string	The earlier experiment that scored the same decision and test seed, named with tma-g5 --repeat-of (a reproduction, not a second chance at the gate).
test_seed	1..1	string
truth_sha256	1..1	string	SHA-256 (hex) of the test split file.

GateVerdict

Machine verdict for a gate; the team review records the final decision.

Attributes

Name	Cardinality:	Type	Description
checks	1..*	GateCheck
decided_by	1..1	string
gate	1..1	GateId
result	1..1	GateResult

ManoeuvreSpan

Where one target manoeuvre lies in a scenario's samples (generator 0.5.0).

Attributes

Name	Cardinality:	Type	Description
end_index	1..1	integer	First timestep after it completes.
kind	1..1	ManoeuvreKind
start_index	1..1	integer	First timestep after it starts.

Referenced by:

• ScenarioMetadata : scenarioMetadata__manoeuvres _0..*

MetricsRecord

Metric values for one estimator on one slice at one evaluation point.

Attributes

Name	Cardinality:	Type	Description
course_deg_median	0..1	float
course_deg_p90	0..1	float
estimator	1..1	string
excluded	1..1	integer
gross_error_rate	0..1	float
n	1..1	integer
non_converged_rate	0..1	float	Fraction of rows with no estimate (scored +∞; explains absent medians/P90).
point	1..1	string	"final" or "zig+s".
position_m_mean	0..1	float	Mean position error over converged rows (non-converged rows count in gross_error_rate).
position_window_m_median	0..1	float	Median over scenarios of each scenario's mean position error over its predicted rows (time-late window-mean position, SRD §5; reported, not gated).
position_window_m_p90	0..1	float	90th percentile over scenarios of the same per-scenario mean.
range_rel_median	0..1	float
range_rel_p90	0..1	float
slice	1..1	string	"overall" or e.g. "range_band=5-10km".
speed_kt_median	0..1	float
speed_kt_p90	0..1	float

ModelDigest

A trained model named by its manifest id and the SHA-256 of its files (schema 0.11.0; manifest and model.json digests from 0.11.1).

Attributes

Name	Cardinality:	Type	Description
manifest_sha256	1..1	string	SHA-256 (hex) of the model's manifest.json (schema 0.11.1).
model	1..1	string	The model's manifest id.
model_json_sha256	1..1	string	SHA-256 (hex) of the model's model.json, its topology (schema 0.11.1).
weights_sha256	1..1	string	SHA-256 (hex) of the model's weights.bin; equals its manifest weights_digest.

Referenced by:

• ChangeDecision : changeDecision__model_digests _0..*

ModelManifest

Everything needed to reproduce and identify a trained model.

Attributes

Name	Cardinality:	Type	Description
id	1..1	string
best_epoch	1..1	integer
best_val_loss	0..1	float	The validation loss a later epoch must beat, under this manifest's config: the best epoch's val_loss, or, after a continuation changed the loss, the best weights re-scored under the new loss (spec 004 FR-002). Persisted so a chained job resumes against the same value. Absent before schema 0.8.2 (then the best epoch's recorded val_loss).
complete	1..1	boolean	False while training is in progress (checkpoint).
config	1..1	TrainingConfig
continued_from	0..1	Continuation	Present when the run was seeded from a committed model (absent before schema 0.8.2).
cpu_model	1..1	string
epochs	1..*	EpochRecord
g0_experiment	0..1	string	Id of the G0 experiment whose verdict the training guard accepted.
generator_version	1..1	string
git_commit	0..1	string	Commit of the training code, read when training started (absent before schema 0.6.0).
git_dirty	0..1	boolean	True when the working tree had uncommitted changes when training started (absent before schema 0.6.0).
node_version	1..1	string
normalisation	1..1	NormalisationConstants
parameter_count	1..1	integer
schema_version	1..1	string
source_dataset_header	0..1	DatasetHeader	Header of the dataset the train features came from (lineage; absent before schema 0.8.0).
stop_reason	0..1	StopReason
tf_interop_threads	0..1	integer	TF_NUM_INTEROP_THREADS in force when training ran (absent: unset).
tf_intraop_threads	0..1	integer	TF_NUM_INTRAOP_THREADS in force when training ran (absent: unset, so the TensorFlow default of one per core).
tfjs_node_version	1..1	string
train_features_digest	1..1	string
val_features_digest	1..1	string
wall_s	1..1	float
weights_digest	1..1	string	SHA-256 of best-epoch weights.

NaiveMeans

Training-set means used by the naive baseline for one prediction mode.

Attributes

Name	Cardinality:	Type	Description
course_cos	1..1	float
course_sin	1..1	float
log_range_km	1..1	float	Mean of ln(range in km).
n	1..1	integer	Number of labels averaged.
speed_mps	1..1	float

Referenced by:

• NaiveModel : naiveModel__final _1..1
• NaiveModel : naiveModel__per_timestep _1..1

NaiveModel

Naive baseline (SRD §5): predicts training-set means. final uses each scenario's last-sample label; per_timestep uses every post-zig label.

Attributes

Name	Cardinality:	Type	Description
final	1..1	NaiveMeans
per_timestep	1..1	NaiveMeans
schema_version	1..1	string
train_digest	1..1	string	SHA-256 of the train file.

NormalisationConstants

Fixed scaling constants computed from the train split's input channels only.

Attributes

Name	Cardinality:	Type	Description
bearing_rate_scale_rad_s	0..1	float	90th percentile of |trailing least-squares bearing rate| over train inputs; required for the derived feature set.
computed_from	1..1	string	Always "inputs_only".
position_scale_km	1..1	float
speed_scale_mps	1..1	float
time_scale_s	1..1	float
train_header_digest	1..1	string	SHA-256 of the train header.

Referenced by:

• ExperimentManifest : experimentManifest__normalisation _0..1
• ModelManifest : modelManifest__normalisation _1..1

Observation

One bearing from ownship to target. Stone Soup analogue Detection. When valid is false (dropout) bearing_rad is 0 and must be ignored.

Attributes

Name	Cardinality:	Type	Description
bearing_rad	1..1	float
timestamp_s	1..1	float
valid	1..1	boolean

Referenced by:

• Scenario : scenario__observations _1..*

PairedComparison

A model and a reference estimator compared on rows both answered (converged and finite), with each estimator's coverage reported separately (spec 003 R9).

Attributes

Name	Cardinality:	Type	Description
course_deg_median	0..1	float
course_deg_median_ref	0..1	float
coverage_estimator	1..1	float	Answered / scored.
coverage_reference	1..1	float	Answered / scored.
estimator	1..1	string
gross_error_rate	0..1	float
gross_error_rate_ref	0..1	float
n_common	1..1	integer
point	1..1	string
range_rel_median	0..1	float
range_rel_median_ref	0..1	float
range_rel_p90	0..1	float
range_rel_p90_ref	0..1	float
reference	1..1	string
slice	1..1	string
speed_kt_median	0..1	float
speed_kt_median_ref	0..1	float

ParameterRanges

Ranges from which ScenarioParameters are drawn. Initial ownship course, target bearing and target course are always uniform on [0, 2π); zig sign is uniform ±.

Attributes

Name	Cardinality:	Type	Description
bearing_error	0..1	BearingErrorModel	Generator 0.6.0: the same realistic bearing error for every scenario with σ > 0, recorded in its parameters; a noise-free scenario (σ = 0) has none. Absent: white noise only, as 0.5.0.
dropout_fraction	1..1	FloatRange
dt_s	1..1	float
los_offset_rad	0..1	FloatRange	Angle off the line of sight, on the other bow, after an across_los turn; required when zig_rule is across_los.
min_range_m	1..1	float	Geometries closer than this at any sample are redrawn.
noise_sigma_rad_values	1..*	float	σ drawn uniformly from this set.
ownship_speed_mps	1..1	FloatRange
ownship_turn_rate_rad_s	0..1	FloatRange	Turn rate drawn uniformly per scenario; absent means an instantaneous zig.
range_distribution	1..1	RangeDistribution
target_course_change_count_weights	0..*	float	Generator 0.5.0 (SRD Annex C §C.2): probabilities of 0, 1, 2, … target course changes (summing to 1). Present: every scenario draws its manoeuvres this way, with target_manoeuvre_angle_abs_rad, target_turn_rate_rad_s, target_manoeuvre_after_zig_s and target_manoeuvre_margin_s, and target_manoeuvre_fraction must be absent.
target_manoeuvre_after_zig_s	0..1	float	Earliest start of the target's turn, after the first sample on leg 2.
target_manoeuvre_angle_abs_rad	0..1	FloatRange	Absolute target course change; the sign is uniform ±.
target_manoeuvre_fraction	0..1	float	Probability a scenario's target makes one course change (generator 0.4.0, SRD Annex B §B.2); absent means the target always holds course, as generators 0.1.0–0.3.0. When present, every scenario draws the manoeuvre parameters (so the PRNG is consumed alike) and the window must fit a turn whether or not one is made.
target_manoeuvre_margin_s	0..1	float	Least time from the end of the target's turn to the end of the window, so the gate point (manoeuvre + 5 min) is reached.
target_manoeuvre_spacing_s	0..1	float	Least time from the end of one target manoeuvre to the start of the next.
target_range0_m	1..1	FloatRange
target_speed_change_abs_mps	0..1	FloatRange	Absolute speed change; its sign is uniform ±, flipped if the new speed would leave target_speed_mps (redrawn if both signs would).
target_speed_change_fraction	0..1	float	Probability of one target speed change (generator 0.5.0).
target_speed_change_rate_mps2	0..1	FloatRange	Magnitude of the target's acceleration during a speed change.
target_speed_mps	1..1	FloatRange
target_turn_rate_rad_s	0..1	FloatRange	Target turn rate, drawn uniformly per scenario.
window_s	1..1	FloatRange
zig2_delay_s	0..1	FloatRange	Delay from the start of leg 2 to the second turn; absent means one zig. The second zig follows zig_rule (and uses the same turn rate as the first).
zig_angle_abs_rad	1..1	FloatRange
zig_fraction	1..1	FloatRange
zig_rule	0..1	ZigRule	How each zig is chosen; absent means random (generators 0.1.0–0.2.0). Under across_los zig_angle_abs_rad is still drawn but not used.

Referenced by:

• DatasetHeader : datasetHeader__parameter_ranges _1..1

PlatformState

One platform state. Stone Soup analogue GroundTruthState.

Attributes

Name	Cardinality:	Type	Description
course_rad	1..1	float
speed_mps	1..1	float
timestamp_s	1..1	float
x_m	1..1	float
y_m	1..1	float

Referenced by:

• Track : track__states _1..*

Prediction

One estimator's output for one scenario.

Attributes

Name	Cardinality:	Type	Description
estimator	1..1	string
mode	1..1	PredictionMode
record	1..1	RecordKind
rows	1..*	PredictionRow
scenario_index	1..1	integer

PredictionHeader

First record of every prediction file.

Attributes

Name	Cardinality:	Type	Description
dataset_digest	1..1	string	SHA-256 of the truth file.
estimator	1..1	string
inputs	0..*	PredictionInput	The prediction files a derived estimator was built from, such as a hybrid's tracker and network files (schema 0.17.0; written by tma-hybrid-track). Gates check each against the files they are given.
mode	1..1	PredictionMode
model_weights_sha256	0..1	string	SHA-256 (hex) of the weights.bin of the model that made these predictions (schema 0.17.0; written by tma-model-predict). Gates check it against the bound model.
record	1..1	RecordKind
schema_version	1..1	string

PredictionInput

One input file of a derived prediction file: its estimator and the SHA-256 of the whole file (schema 0.17.0).

Attributes

Name	Cardinality:	Type	Description
estimator	1..1	string
sha256	1..1	string	SHA-256 (hex) of the input file.

Referenced by:

• DetectionHeader : detectionHeader__inputs _0..*
• PredictionHeader : predictionHeader__inputs _0..*

PredictionRow

An estimator's target estimate at one timestep.

Attributes

Name	Cardinality:	Type	Description
converged	1..1	boolean
course_rad	1..1	float
course_sd_rad	0..1	float	Stated standard deviation of the course estimate (schema 0.20.0, spec 018): from a tracker's covariance by the delta method, written only under the -sd name suffix.
iterations	0..1	integer
log_range_sd	0..1	float	Standard deviation of the estimator's log range (natural log; about the relative range error), from a velocity_gaussian network (spec 012). Absent for other estimators.
p_manoeuvre	0..1	float	The estimator's probability that the target's turn has started by this timestep (velocity_two_hypothesis networks only; spec 011). Absent for other estimators.
range_m	1..1	float
speed_mps	1..1	float
speed_sd_mps	0..1	float	Stated standard deviation of the speed estimate (schema 0.20.0, spec 018).
timestamp_s	1..1	float
vx_sd_mps	0..1	float	Standard deviation of the estimated east velocity (velocity_gaussian only).
vy_sd_mps	0..1	float	Standard deviation of the estimated north velocity (velocity_gaussian only).
x_m	1..1	float
y_m	1..1	float

Referenced by:

• Prediction : prediction__rows _1..*

Scenario

One ownship-vs-target run over a window.

Attributes

Name	Cardinality:	Type	Description
metadata	1..1	ScenarioMetadata
observations	1..*	Observation
ownship	1..1	Track
record	1..1	RecordKind
target	1..1	Track

ScenarioMetadata

Seed, parameter draws and derived slicing keys for one scenario.

Attributes

Name	Cardinality:	Type	Description
index	1..1	integer	Scenario index within the dataset.
manoeuvre_index	0..1	integer	First timestep after the target's turn completes (generator 0.4.0; absent when the target holds course). Gate G4 counts time from here, as earlier gates count from zig_index. Under generator 0.5.0, after the last manoeuvre completes.
manoeuvre_start_index	0..1	integer	First timestep after the target's turn starts (generator 0.4.0; absent when the target holds course). Under generator 0.5.0, after the first manoeuvre starts.
manoeuvres	0..*	ManoeuvreSpan	Generator 0.5.0: each target manoeuvre's sample span, in time order (present, possibly empty, only under 0.5.0).
mean_range_m	1..1	float
min_range_m	1..1	float
parameters	1..1	ScenarioParameters
samples	1..1	integer	Number of timesteps (60–240).
seed	1..1	string	Scenario seed as a decimal uint64 string.
turn2_start_index	0..1	integer	First timestep after the second turn starts (absent with one zig or no turn rate).
turn_start_index	0..1	integer	First timestep after the turn starts (absent for an instantaneous zig).
zig2_index	0..1	integer	First timestep on leg 3, after the second turn completes (absent with one zig). Gate times still count from zig_index.
zig_index	1..1	integer	First timestep on leg 2 (with a rate-limited turn, the first after it completes).
zig_time_s	1..1	float

Referenced by:

• Scenario : scenario__metadata _1..1

ScenarioParameters

The random draws that define one scenario.

Attributes

Name	Cardinality:	Type	Description
bearing_error	0..1	BearingErrorModel	Realistic bearing error (generator 0.6.0); absent = white noise only.
dropout_fraction	1..1	float	Probability each bearing is missing.
dt_s	1..1	float	Bearing interval.
noise_sigma_rad	1..1	float	Bearing noise σ; 0 = noise-free.
ownship_course0_rad	1..1	float	Ownship course on leg 1.
ownship_speed_mps	1..1	float
ownship_turn_rate_rad_s	0..1	float	Ownship turn rate at the zig (rate-limited turn, spec 003 US6). Absent or 0: the course changes instantaneously, as in generator 0.1.0.
target_bearing0_rad	1..1	float	True bearing at window start.
target_course_rad	1..1	float
target_manoeuvre_angle_rad	0..1	float	Signed target course change; positive is clockwise.
target_manoeuvre_time_s	0..1	float	Target course change (generator 0.4.0, SRD Annex B §B.2): time the target's turn starts, a whole number of samples. Absent: the target holds course.
target_manoeuvres	0..*	TargetManoeuvre	Target manoeuvres in time order (generator 0.5.0, SRD Annex C §C.2), none overlapping. Present (possibly empty) only under generator 0.5.0 ranges; then the single-turn fields above are absent.
target_range0_m	1..1	float	Range at window start.
target_speed_mps	1..1	float
target_turn_rate_rad_s	0..1	float	Target turn rate through its course change.
window_s	1..1	float	Window length; multiple of dt_s.
zig2_angle_rad	0..1	float	Signed course change at the second zig; positive is clockwise.
zig2_delay_s	0..1	float	Second zig (generator 0.3.0, spec 007): time from the first sample on leg 2 to the start of the second turn. Absent: one zig, two legs.
zig_angle_rad	1..1	float	Signed course change at the zig; positive is clockwise.
zig_fraction	1..1	float	Zig time as a fraction of the window.

Referenced by:

• ScenarioMetadata : scenarioMetadata__parameters _1..1

TargetManoeuvre

One target manoeuvre (generator 0.5.0). A course change is an arc at constant speed; a speed change is constant acceleration on a constant course.

Attributes

Name	Cardinality:	Type	Description
change	1..1	float	Signed change: course in rad (positive clockwise) or speed in m/s (positive faster).
kind	1..1	ManoeuvreKind
rate	1..1	float	Magnitude of the rate of change, rad/s (course) or m/s² (speed).
start_time_s	1..1	float	Time the manoeuvre starts, a whole number of samples.

Referenced by:

• ScenarioParameters : scenarioParameters__target_manoeuvres _0..*

Track

Time-ordered states of one platform. Stone Soup analogue GroundTruthPath.

Attributes

Name	Cardinality:	Type	Description
id	1..1	TrackId
states	1..*	PlatformState

Referenced by:

• Scenario : scenario__ownship _1..1
• Scenario : scenario__target _1..1

TrainingConfig

Hyperparameters and seeds for one training run (SRD §4; constitution III).

Attributes

Name	Cardinality:	Type	Description
id	1..1	string	Estimator name used in predictions.
batch_size	1..1	integer
bucket_batches	0..1	integer	Batches per length-bucketing group (default 16).
feature_set	0..1	FeatureSet	Input channel set; absent means raw. Must match the features files.
kernel_size	0..1	integer	Causal CNN kernel size.
layers	0..1	integer	Stacked layers: causal CNN conv layers (no dilation; default 3), or recurrent layers for instantaneous and time_late (default 1; each returns sequences; spec 003 R20).
learning_rate	1..1	float
length_bucketing	0..1	boolean	Length bucketing (research R15; default false). Each group of bucket_batches batches of the epoch's seeded permutation is sorted by valid length, cut into batches whose rows are trimmed to the longest window, and the batch order is shuffled with a seed derived from train_seed and the epoch. The model takes a variable time dimension. Not for mlp.
loss_weight_course	1..1	float
loss_weight_manoeuvre	0..1	float	Weight of the turn-probability cross-entropy term (velocity_two_hypothesis only; spec 011). Absent means 1.
loss_weight_range	1..1	float
loss_weight_speed	1..1	float
max_epochs	1..1	integer
max_wall_s	1..1	float	Wall-clock cap in seconds.
model_kind	1..1	ModelKind
output_parameterisation	0..1	OutputParameterisation	Output encoding; absent means course_speed. With velocity, loss_weight_speed weights each velocity component and loss_weight_course is unused (spec 008).
patience	1..1	integer	Epochs without validation improvement before stopping.
shuffle_labels	1..1	boolean	Permute targets across scenarios (leakage check).
tfjs_threads	0..1	integer	TensorFlow intra-op and inter-op thread count, set by tma-train through TF_NUM_INTRAOP_THREADS and TF_NUM_INTEROP_THREADS before tfjs-node loads. Absent: the TensorFlow default (one per core) or whatever the environment sets.
train_limit	0..1	integer	Use only the first N training scenarios (tests and smoke runs).
train_seed	1..1	string	Decimal uint64; derives initializer seeds and batch order.
units	1..1	integer	LSTM units (per direction), conv filters or dense width.
velocity_scale_mps	0..1	float	Velocity scale in m/s for the velocity parameterisation (required with it): the top of the generator's target speed range, a constant of the problem, never a truth statistic.

Referenced by:

• ModelManifest : modelManifest__config _1..1

UncertaintyCalibration

One estimator's post-hoc interval multipliers for one output (schema 0.20.0, spec 018): the calibrated q-interval is the estimate ± multiplier × the stated standard deviation.

Attributes

Name	Cardinality:	Type	Description
estimator	1..1	string
output	1..1	string	"log_range", "course" or "speed".
q50	1..1	float	Multiplier for the 50% interval.
q90	1..1	float	Multiplier for the 90% interval.

Referenced by:

• ChangeDecision : changeDecision__calibrations _0..*

Slots

Name	Cardinality/Range	Used By
bearingErrorModel__ar_sd_rad Stationary sd of the correlated term.	1..1 float
bearingErrorModel__ar_tau_s Time constant of the correlated term.	1..1 float
bearingErrorModel__bias_sd_rad Stationary sd of the drifting bias.	1..1 float
bearingErrorModel__bias_tau_s Time constant of the drifting bias.	1..1 float
changeDecision__calibrations The frozen post-hoc interval multipliers of each estimator and output a G7 decision names (schema 0.20.0, spec 018), fitted on validation; tma-g7 refuses others.	0..* UncertaintyCalibration
changeDecision__carried_model G4: the manifest id of the one carried network (Annex B §B.4); tma-g4 refuses network predictions from another model.	0..1 string
changeDecision__carried_models G5 (spec 012, schema 0.16.0): the manifest ids of the carried trackers (one or both of T1 and T2, Annex C §C.5), each with its own plan in plans, in the same order; tma-g5 refuses tracker predictions from any other model.	0..* string
changeDecision__carried_set_rule	1..1 string
changeDecision__change e.g. feature_set: derived	1..1 string
changeDecision__decided_at	1..1 string
changeDecision__detector_thresholds The frozen alarm threshold of each detector a G6 decision names (schema 0.19.0, spec 017), set on validation at the matched false-alarm rate; tma-g6 refuses others.	0..* DetectorThreshold
changeDecision__fresh_test_count The dataset count whose test split a re-score scores (the plan's dataset.fresh_test.count; schema 0.11.1); tma-g2 refuses a plan with another count.	0..1 integer
changeDecision__git_commit	1..1 string
changeDecision__model_digests The frozen carried models a re-score scores, by manifest id, with the SHA-256 of each one's weights.bin, manifest.json and model.json; tma-g2 refuses a re-score whose carried set or any of those files differs.	0..* ModelDigest
changeDecision__phase_a_experiment	1..1 string
changeDecision__plan The plan this decision pre-registers (configs/<…>/plan.json). Required from iteration 3 (spec 007): tma-g2 scores a test split only with the decision that names its plan.	0..1 string
changeDecision__plans G5: the plan of each carried tracker (configs/phase4//plan.json), in the order of carried_models; they share one dataset and one fresh test seed.	0..* string
changeDecision__rationale Validation evidence cited.	1..1 string
changeDecision__reference_estimator G4 (spec 010, schema 0.13.0): the frozen reference tracker's estimator name, which encodes every tuning parameter (e.g. imm-cv0-ct0.03-w0.003-s0.0001-r4); tma-g4 refuses reference predictions with another name.	0..1 string
changeDecision__rescore_of The experiment directory of the inconclusive gate this decision re-scores (SRD §7: one re-score on a larger fresh-seed test set; e.g. experiments/20261024-i4-g2). When present, tma-g2 runs in re-score mode (--rescore-of): the frozen carried models are bound by model_digests instead of by training ancestry, the re-scored gate must be inconclusive and not itself a re-score, and the fresh test seed must differ from its test seed.	0..1 string
changeDecision__stalled Any gated criterion failed or was inconclusive on validation.	1..1 boolean
changeDecision__test_seed Dataset seed of the test split the decision commits to (the plan's fresh_test seed, or its dataset seed); tma-g2 checks it against the plan.	0..1 string
continuation__best_epoch	1..1 integer
continuation__git_commit The source model's training-code commit, when it recorded one.	0..1 string
continuation__rescored_val_loss The source's best weights re-scored under the continued run's loss, written by tma-train when it resumed (spec 004 FR-002). With best_epoch and best_val_loss it makes the kept weights auditable: a run whose best_epoch is the source's must have best_val_loss equal to this value and weights_digest equal to the source's.	0..1 float
continuation__val_loss The source's best-epoch validation loss under the source's config.	1..1 float
continuation__weights_digest SHA-256 of the weights the continuation resumed from (the source's best epoch).	1..1 string
curvePoint__course_deg_median	0..1 float
curvePoint__estimator	1..1 string
curvePoint__gross_error_rate	1..1 float
curvePoint__n	1..1 integer
curvePoint__non_converged_rate	0..1 float
curvePoint__range_rel_median Absent when not finite (more than half the rows non-converged), as in MetricsRecord.	0..1 float
curvePoint__range_rel_p90	0..1 float
curvePoint__speed_kt_median	0..1 float
curvePoint__t_since_zig_s	1..1 float
datasetHeader__created_by Command line that produced the file.	1..1 string
datasetHeader__dataset_seed Decimal uint64 string.	1..1 string
datasetHeader__generator_git_commit Commit of the checkout that ran tma-generate (absent before schema 0.8.0).	0..1 string
datasetHeader__generator_git_dirty True when that checkout had uncommitted changes (only allowed with --allow-dirty; absent before schema 0.8.0).	0..1 boolean
datasetHeader__generator_version	1..1 string
datasetHeader__index_end Last scenario index (exclusive).	1..1 integer
datasetHeader__index_start First scenario index (inclusive).	1..1 integer
datasetHeader__parameter_ranges	1..1 ParameterRanges
datasetHeader__record	1..1 RecordKind
datasetHeader__redraw_count Geometries rejected for violating min_range_m.	1..1 integer
datasetHeader__scenario_count	1..1 integer
datasetHeader__schema_version	1..1 string
datasetHeader__split	1..1 Split
datasetRef__digest SHA-256 of the dataset file (absent in manifests written before schema 0.4.0).	0..1 string
datasetRef__header	1..1 DatasetHeader
datasetRef__path	1..1 string
detectionHeader__dataset_digest SHA-256 of the truth file.	1..1 string
detectionHeader__detector	1..1 string
detectionHeader__inputs The prediction files the scores were computed from, with their SHA-256s.	0..* PredictionInput
detectionHeader__record	1..1 RecordKind
detectionHeader__schema_version	1..1 string
detectionRow__score The detector's statistic at this timestep (units depend on the detector).	1..1 float
detectionRow__timestamp_s	1..1 float
detection__detector	1..1 string
detection__record	1..1 RecordKind
detection__rows	1..* DetectionRow
detection__scenario_index	1..1 integer
detectorThreshold__detector	1..1 string
detectorThreshold__threshold	1..1 float
epochRecord__epoch	1..1 integer
epochRecord__seconds Cumulative wall time.	1..1 float
epochRecord__train_loss	1..1 float
epochRecord__val_loss	1..1 float
estimatorRef__name	1..1 string
estimatorRef__settings_json Estimator settings as JSON.	1..1 string
experimentManifest__command_lines	1..* string
experimentManifest__created ISO 8601 timestamp.	1..1 string
experimentManifest__datasets	1..* DatasetRef
experimentManifest__estimators	1..* EstimatorRef
experimentManifest__generator_version	1..1 string
experimentManifest__git_commit	1..1 string
experimentManifest__git_dirty True when the working tree had uncommitted changes (absent before schema 0.4.0).	0..1 boolean
experimentManifest__id	1..1 string
experimentManifest__node_version Node.js version that ran the experiment (byte identity is promised per pinned Node).	0..1 string
experimentManifest__normalisation	0..1 NormalisationConstants
experimentManifest__schema_version	1..1 string
featureLayout__feature_set Absent means raw.	0..1 FeatureSet
featureLayout__input_channels	1..* string
featureLayout__rows	1..1 integer
featureLayout__target_channels	1..* string
floatRange__max	1..1 float
floatRange__min	1..1 float
gateCheck__ci_high Upper bound of the bootstrap interval.	0..1 float
gateCheck__ci_level Interval coverage, e.g. 0.95.	0..1 float
gateCheck__ci_low Lower bound of the bootstrap interval.	0..1 float
gateCheck__detail	1..1 string
gateCheck__name	1..1 string
gateCheck__result	1..1 GateResult
gateCheck__threshold Limit the statistic is tested against.	0..1 float
gateCheck__value Statistic tested (e.g. median range error).	0..1 float
gateRecomputation__estimator	1..1 string
gateRecomputation__sample_sha256 SHA-256 (hex) of the sampled scenario indices, comma-separated in ascending order.	1..1 string
gateRecomputation__scenarios Number of scenarios re-computed.	1..1 integer
gateScoring__decision Path of the decision record, from the repository root.	1..1 string
gateScoring__decision_sha256 SHA-256 (hex) of the decision record.	1..1 string
gateScoring__gate	1..1 GateId
gateScoring__recomputed The estimators whose rows the gate re-computed on a seeded sample of scenarios from their declared settings, byte-identical to their files (absent in records backfilled for experiments scored before schema 0.18.0).	0..* GateRecomputation
gateScoring__repeat_of The earlier experiment that scored the same decision and test seed, named with tma-g5 --repeat-of (a reproduction, not a second chance at the gate).	0..1 string
gateScoring__test_seed	1..1 string
gateScoring__truth_sha256 SHA-256 (hex) of the test split file.	1..1 string
gateVerdict__checks	1..* GateCheck
gateVerdict__decided_by	1..1 string
gateVerdict__gate	1..1 GateId
gateVerdict__result	1..1 GateResult
manoeuvreSpan__end_index First timestep after it completes.	1..1 integer
manoeuvreSpan__kind	1..1 ManoeuvreKind
manoeuvreSpan__start_index First timestep after it starts.	1..1 integer
metricsRecord__course_deg_median	0..1 float
metricsRecord__course_deg_p90	0..1 float
metricsRecord__estimator	1..1 string
metricsRecord__excluded	1..1 integer
metricsRecord__gross_error_rate	0..1 float
metricsRecord__n	1..1 integer
metricsRecord__non_converged_rate Fraction of rows with no estimate (scored +∞; explains absent medians/P90).	0..1 float
metricsRecord__point "final" or "zig+s".	1..1 string
metricsRecord__position_m_mean Mean position error over converged rows (non-converged rows count in gross_error_rate).	0..1 float
metricsRecord__position_window_m_median Median over scenarios of each scenario's mean position error over its predicted rows (time-late window-mean position, SRD §5; reported, not gated).	0..1 float
metricsRecord__position_window_m_p90 90th percentile over scenarios of the same per-scenario mean.	0..1 float
metricsRecord__range_rel_median	0..1 float
metricsRecord__range_rel_p90	0..1 float
metricsRecord__slice "overall" or e.g. "range_band=5-10km".	1..1 string
metricsRecord__speed_kt_median	0..1 float
metricsRecord__speed_kt_p90	0..1 float
modelDigest__manifest_sha256 SHA-256 (hex) of the model's manifest.json (schema 0.11.1).	1..1 string
modelDigest__model The model's manifest id.	1..1 string
modelDigest__model_json_sha256 SHA-256 (hex) of the model's model.json, its topology (schema 0.11.1).	1..1 string
modelDigest__weights_sha256 SHA-256 (hex) of the model's weights.bin; equals its manifest weights_digest.	1..1 string
modelManifest__best_epoch	1..1 integer
modelManifest__best_val_loss The validation loss a later epoch must beat, under this manifest's config: the best epoch's val_loss, or, after a continuation changed the loss, the best weights re-scored under the new loss (spec 004 FR-002). Persisted so a chained job resumes against the same value. Absent before schema 0.8.2 (then the best epoch's recorded val_loss).	0..1 float
modelManifest__complete False while training is in progress (checkpoint).	1..1 boolean
modelManifest__config	1..1 TrainingConfig
modelManifest__continued_from Present when the run was seeded from a committed model (absent before schema 0.8.2).	0..1 Continuation
modelManifest__cpu_model	1..1 string
modelManifest__epochs	1..* EpochRecord
modelManifest__g0_experiment Id of the G0 experiment whose verdict the training guard accepted.	0..1 string
modelManifest__generator_version	1..1 string
modelManifest__git_commit Commit of the training code, read when training started (absent before schema 0.6.0).	0..1 string
modelManifest__git_dirty True when the working tree had uncommitted changes when training started (absent before schema 0.6.0).	0..1 boolean
modelManifest__id	1..1 string
modelManifest__node_version	1..1 string
modelManifest__normalisation	1..1 NormalisationConstants
modelManifest__parameter_count	1..1 integer
modelManifest__schema_version	1..1 string
modelManifest__source_dataset_header Header of the dataset the train features came from (lineage; absent before schema 0.8.0).	0..1 DatasetHeader
modelManifest__stop_reason	0..1 StopReason
modelManifest__tf_interop_threads TF_NUM_INTEROP_THREADS in force when training ran (absent: unset).	0..1 integer
modelManifest__tf_intraop_threads TF_NUM_INTRAOP_THREADS in force when training ran (absent: unset, so the TensorFlow default of one per core).	0..1 integer
modelManifest__tfjs_node_version	1..1 string
modelManifest__train_features_digest	1..1 string
modelManifest__val_features_digest	1..1 string
modelManifest__wall_s	1..1 float
modelManifest__weights_digest SHA-256 of best-epoch weights.	1..1 string
naiveMeans__course_cos	1..1 float
naiveMeans__course_sin	1..1 float
naiveMeans__log_range_km Mean of ln(range in km).	1..1 float
naiveMeans__n Number of labels averaged.	1..1 integer
naiveMeans__speed_mps	1..1 float
naiveModel__final	1..1 NaiveMeans
naiveModel__per_timestep	1..1 NaiveMeans
naiveModel__schema_version	1..1 string
naiveModel__train_digest SHA-256 of the train file.	1..1 string
normalisationConstants__bearing_rate_scale_rad_s 90th percentile of |trailing least-squares bearing rate| over train inputs; required for the derived feature set.	0..1 float
normalisationConstants__computed_from Always "inputs_only".	1..1 string
normalisationConstants__position_scale_km	1..1 float
normalisationConstants__speed_scale_mps	1..1 float
normalisationConstants__time_scale_s	1..1 float
normalisationConstants__train_header_digest SHA-256 of the train header.	1..1 string
observation__bearing_rad	1..1 float
observation__timestamp_s	1..1 float
observation__valid	1..1 boolean
pairedComparison__course_deg_median	0..1 float
pairedComparison__course_deg_median_ref	0..1 float
pairedComparison__coverage_estimator Answered / scored.	1..1 float
pairedComparison__coverage_reference Answered / scored.	1..1 float
pairedComparison__estimator	1..1 string
pairedComparison__gross_error_rate	0..1 float
pairedComparison__gross_error_rate_ref	0..1 float
pairedComparison__n_common	1..1 integer
pairedComparison__point	1..1 string
pairedComparison__range_rel_median	0..1 float
pairedComparison__range_rel_median_ref	0..1 float
pairedComparison__range_rel_p90	0..1 float
pairedComparison__range_rel_p90_ref	0..1 float
pairedComparison__reference	1..1 string
pairedComparison__slice	1..1 string
pairedComparison__speed_kt_median	0..1 float
pairedComparison__speed_kt_median_ref	0..1 float
parameterRanges__bearing_error Generator 0.6.0: the same realistic bearing error for every scenario with σ > 0, recorded in its parameters; a noise-free scenario (σ = 0) has none. Absent: white noise only, as 0.5.0.	0..1 BearingErrorModel
parameterRanges__dropout_fraction	1..1 FloatRange
parameterRanges__dt_s	1..1 float
parameterRanges__los_offset_rad Angle off the line of sight, on the other bow, after an across_los turn; required when zig_rule is across_los.	0..1 FloatRange
parameterRanges__min_range_m Geometries closer than this at any sample are redrawn.	1..1 float
parameterRanges__noise_sigma_rad_values σ drawn uniformly from this set.	1..* float
parameterRanges__ownship_speed_mps	1..1 FloatRange
parameterRanges__ownship_turn_rate_rad_s Turn rate drawn uniformly per scenario; absent means an instantaneous zig.	0..1 FloatRange
parameterRanges__range_distribution	1..1 RangeDistribution
parameterRanges__target_course_change_count_weights Generator 0.5.0 (SRD Annex C §C.2): probabilities of 0, 1, 2, … target course changes (summing to 1). Present: every scenario draws its manoeuvres this way, with target_manoeuvre_angle_abs_rad, target_turn_rate_rad_s, target_manoeuvre_after_zig_s and target_manoeuvre_margin_s, and target_manoeuvre_fraction must be absent.	0..* float
parameterRanges__target_manoeuvre_after_zig_s Earliest start of the target's turn, after the first sample on leg 2.	0..1 float
parameterRanges__target_manoeuvre_angle_abs_rad Absolute target course change; the sign is uniform ±.	0..1 FloatRange
parameterRanges__target_manoeuvre_fraction Probability a scenario's target makes one course change (generator 0.4.0, SRD Annex B §B.2); absent means the target always holds course, as generators 0.1.0–0.3.0. When present, every scenario draws the manoeuvre parameters (so the PRNG is consumed alike) and the window must fit a turn whether or not one is made.	0..1 float
parameterRanges__target_manoeuvre_margin_s Least time from the end of the target's turn to the end of the window, so the gate point (manoeuvre + 5 min) is reached.	0..1 float
parameterRanges__target_manoeuvre_spacing_s Least time from the end of one target manoeuvre to the start of the next.	0..1 float
parameterRanges__target_range0_m	1..1 FloatRange
parameterRanges__target_speed_change_abs_mps Absolute speed change; its sign is uniform ±, flipped if the new speed would leave target_speed_mps (redrawn if both signs would).	0..1 FloatRange
parameterRanges__target_speed_change_fraction Probability of one target speed change (generator 0.5.0).	0..1 float
parameterRanges__target_speed_change_rate_mps2 Magnitude of the target's acceleration during a speed change.	0..1 FloatRange
parameterRanges__target_speed_mps	1..1 FloatRange
parameterRanges__target_turn_rate_rad_s Target turn rate, drawn uniformly per scenario.	0..1 FloatRange
parameterRanges__window_s	1..1 FloatRange
parameterRanges__zig2_delay_s Delay from the start of leg 2 to the second turn; absent means one zig. The second zig follows zig_rule (and uses the same turn rate as the first).	0..1 FloatRange
parameterRanges__zig_angle_abs_rad	1..1 FloatRange
parameterRanges__zig_fraction	1..1 FloatRange
parameterRanges__zig_rule How each zig is chosen; absent means random (generators 0.1.0–0.2.0). Under across_los zig_angle_abs_rad is still drawn but not used.	0..1 ZigRule
platformState__course_rad	1..1 float
platformState__speed_mps	1..1 float
platformState__timestamp_s	1..1 float
platformState__x_m	1..1 float
platformState__y_m	1..1 float
predictionHeader__dataset_digest SHA-256 of the truth file.	1..1 string
predictionHeader__estimator	1..1 string
predictionHeader__inputs The prediction files a derived estimator was built from, such as a hybrid's tracker and network files (schema 0.17.0; written by tma-hybrid-track). Gates check each against the files they are given.	0..* PredictionInput
predictionHeader__mode	1..1 PredictionMode
predictionHeader__model_weights_sha256 SHA-256 (hex) of the weights.bin of the model that made these predictions (schema 0.17.0; written by tma-model-predict). Gates check it against the bound model.	0..1 string
predictionHeader__record	1..1 RecordKind
predictionHeader__schema_version	1..1 string
predictionInput__estimator	1..1 string
predictionInput__sha256 SHA-256 (hex) of the input file.	1..1 string
predictionRow__converged	1..1 boolean
predictionRow__course_rad	1..1 float
predictionRow__course_sd_rad Stated standard deviation of the course estimate (schema 0.20.0, spec 018): from a tracker's covariance by the delta method, written only under the -sd name suffix.	0..1 float
predictionRow__iterations	0..1 integer
predictionRow__log_range_sd Standard deviation of the estimator's log range (natural log; about the relative range error), from a velocity_gaussian network (spec 012). Absent for other estimators.	0..1 float
predictionRow__p_manoeuvre The estimator's probability that the target's turn has started by this timestep (velocity_two_hypothesis networks only; spec 011). Absent for other estimators.	0..1 float
predictionRow__range_m	1..1 float
predictionRow__speed_mps	1..1 float
predictionRow__speed_sd_mps Stated standard deviation of the speed estimate (schema 0.20.0, spec 018).	0..1 float
predictionRow__timestamp_s	1..1 float
predictionRow__vx_sd_mps Standard deviation of the estimated east velocity (velocity_gaussian only).	0..1 float
predictionRow__vy_sd_mps Standard deviation of the estimated north velocity (velocity_gaussian only).	0..1 float
predictionRow__x_m	1..1 float
predictionRow__y_m	1..1 float
prediction__estimator	1..1 string
prediction__mode	1..1 PredictionMode
prediction__record	1..1 RecordKind
prediction__rows	1..* PredictionRow
prediction__scenario_index	1..1 integer
scenarioMetadata__index Scenario index within the dataset.	1..1 integer
scenarioMetadata__manoeuvre_index First timestep after the target's turn completes (generator 0.4.0; absent when the target holds course). Gate G4 counts time from here, as earlier gates count from zig_index. Under generator 0.5.0, after the last manoeuvre completes.	0..1 integer
scenarioMetadata__manoeuvre_start_index First timestep after the target's turn starts (generator 0.4.0; absent when the target holds course). Under generator 0.5.0, after the first manoeuvre starts.	0..1 integer
scenarioMetadata__manoeuvres Generator 0.5.0: each target manoeuvre's sample span, in time order (present, possibly empty, only under 0.5.0).	0..* ManoeuvreSpan
scenarioMetadata__mean_range_m	1..1 float
scenarioMetadata__min_range_m	1..1 float
scenarioMetadata__parameters	1..1 ScenarioParameters
scenarioMetadata__samples Number of timesteps (60–240).	1..1 integer
scenarioMetadata__seed Scenario seed as a decimal uint64 string.	1..1 string
scenarioMetadata__turn2_start_index First timestep after the second turn starts (absent with one zig or no turn rate).	0..1 integer
scenarioMetadata__turn_start_index First timestep after the turn starts (absent for an instantaneous zig).	0..1 integer
scenarioMetadata__zig2_index First timestep on leg 3, after the second turn completes (absent with one zig). Gate times still count from zig_index.	0..1 integer
scenarioMetadata__zig_index First timestep on leg 2 (with a rate-limited turn, the first after it completes).	1..1 integer
scenarioMetadata__zig_time_s	1..1 float
scenarioParameters__bearing_error Realistic bearing error (generator 0.6.0); absent = white noise only.	0..1 BearingErrorModel
scenarioParameters__dropout_fraction Probability each bearing is missing.	1..1 float
scenarioParameters__dt_s Bearing interval.	1..1 float
scenarioParameters__noise_sigma_rad Bearing noise σ; 0 = noise-free.	1..1 float
scenarioParameters__ownship_course0_rad Ownship course on leg 1.	1..1 float
scenarioParameters__ownship_speed_mps	1..1 float
scenarioParameters__ownship_turn_rate_rad_s Ownship turn rate at the zig (rate-limited turn, spec 003 US6). Absent or 0: the course changes instantaneously, as in generator 0.1.0.	0..1 float
scenarioParameters__target_bearing0_rad True bearing at window start.	1..1 float
scenarioParameters__target_course_rad	1..1 float
scenarioParameters__target_manoeuvre_angle_rad Signed target course change; positive is clockwise.	0..1 float
scenarioParameters__target_manoeuvre_time_s Target course change (generator 0.4.0, SRD Annex B §B.2): time the target's turn starts, a whole number of samples. Absent: the target holds course.	0..1 float
scenarioParameters__target_manoeuvres Target manoeuvres in time order (generator 0.5.0, SRD Annex C §C.2), none overlapping. Present (possibly empty) only under generator 0.5.0 ranges; then the single-turn fields above are absent.	0..* TargetManoeuvre
scenarioParameters__target_range0_m Range at window start.	1..1 float
scenarioParameters__target_speed_mps	1..1 float
scenarioParameters__target_turn_rate_rad_s Target turn rate through its course change.	0..1 float
scenarioParameters__window_s Window length; multiple of dt_s.	1..1 float
scenarioParameters__zig2_angle_rad Signed course change at the second zig; positive is clockwise.	0..1 float
scenarioParameters__zig2_delay_s Second zig (generator 0.3.0, spec 007): time from the first sample on leg 2 to the start of the second turn. Absent: one zig, two legs.	0..1 float
scenarioParameters__zig_angle_rad Signed course change at the zig; positive is clockwise.	1..1 float
scenarioParameters__zig_fraction Zig time as a fraction of the window.	1..1 float
scenario__metadata	1..1 ScenarioMetadata
scenario__observations	1..* Observation
scenario__ownship	1..1 Track
scenario__record	1..1 RecordKind
scenario__target	1..1 Track
targetManoeuvre__change Signed change: course in rad (positive clockwise) or speed in m/s (positive faster).	1..1 float
targetManoeuvre__kind	1..1 ManoeuvreKind
targetManoeuvre__rate Magnitude of the rate of change, rad/s (course) or m/s² (speed).	1..1 float
targetManoeuvre__start_time_s Time the manoeuvre starts, a whole number of samples.	1..1 float
track__id	1..1 TrackId
track__states	1..* PlatformState
trainingConfig__batch_size	1..1 integer
trainingConfig__bucket_batches Batches per length-bucketing group (default 16).	0..1 integer
trainingConfig__feature_set Input channel set; absent means raw. Must match the features files.	0..1 FeatureSet
trainingConfig__id Estimator name used in predictions.	1..1 string
trainingConfig__kernel_size Causal CNN kernel size.	0..1 integer
trainingConfig__layers Stacked layers: causal CNN conv layers (no dilation; default 3), or recurrent layers for instantaneous and time_late (default 1; each returns sequences; spec 003 R20).	0..1 integer
trainingConfig__learning_rate	1..1 float
trainingConfig__length_bucketing Length bucketing (research R15; default false). Each group of bucket_batches batches of the epoch's seeded permutation is sorted by valid length, cut into batches whose rows are trimmed to the longest window, and the batch order is shuffled with a seed derived from train_seed and the epoch. The model takes a variable time dimension. Not for mlp.	0..1 boolean
trainingConfig__loss_weight_course	1..1 float
trainingConfig__loss_weight_manoeuvre Weight of the turn-probability cross-entropy term (velocity_two_hypothesis only; spec 011). Absent means 1.	0..1 float
trainingConfig__loss_weight_range	1..1 float
trainingConfig__loss_weight_speed	1..1 float
trainingConfig__max_epochs	1..1 integer
trainingConfig__max_wall_s Wall-clock cap in seconds.	1..1 float
trainingConfig__model_kind	1..1 ModelKind
trainingConfig__output_parameterisation Output encoding; absent means course_speed. With velocity, loss_weight_speed weights each velocity component and loss_weight_course is unused (spec 008).	0..1 OutputParameterisation
trainingConfig__patience Epochs without validation improvement before stopping.	1..1 integer
trainingConfig__shuffle_labels Permute targets across scenarios (leakage check).	1..1 boolean
trainingConfig__tfjs_threads TensorFlow intra-op and inter-op thread count, set by tma-train through TF_NUM_INTRAOP_THREADS and TF_NUM_INTEROP_THREADS before tfjs-node loads. Absent: the TensorFlow default (one per core) or whatever the environment sets.	0..1 integer
trainingConfig__train_limit Use only the first N training scenarios (tests and smoke runs).	0..1 integer
trainingConfig__train_seed Decimal uint64; derives initializer seeds and batch order.	1..1 string
trainingConfig__units LSTM units (per direction), conv filters or dense width.	1..1 integer
trainingConfig__velocity_scale_mps Velocity scale in m/s for the velocity parameterisation (required with it): the top of the generator's target speed range, a constant of the problem, never a truth statistic.	0..1 float
uncertaintyCalibration__estimator	1..1 string
uncertaintyCalibration__output "log_range", "course" or "speed".	1..1 string
uncertaintyCalibration__q50 Multiplier for the 50% interval.	1..1 float
uncertaintyCalibration__q90 Multiplier for the 90% interval.	1..1 float

Enums

FeatureSet

Model input channel sets (spec 003 R6).

Text	Meaning:	Description
derived	None	raw plus bearing_rate_scaled, rate_valid and intercept_scaled (13 channels), computed causally from bearings and ownship states up to each row.
raw	None	Iteration-1 layout, 10 input channels.

Used by

• FeatureLayout featureLayout__feature_set _0..1
• TrainingConfig trainingConfig__feature_set _0..1

GateId

Text	Meaning:	Description
G0	None	Pipeline verified.
G1	None
G2	None
G3	None
G4	None	Better than the tuned tracker after a target turn (SRD Annex B §B.4).
G5	None	An end-to-end tracker better than the re-tuned reference overall, at most 10% worse when settled (SRD Annex C §C.5).
G6	None	A manoeuvre detector from frozen models faster than a CUSUM on the reference's innovations at matched false alarms (SRD Annex C §C.12).
G7	None	Calibrated uncertainty: the fused tracker's recalibrated intervals cover their nominal level and are narrower than the reference's (SRD Annex C §C.13).

Used by

• GateScoring gateScoring__gate _1..1
• GateVerdict gateVerdict__gate _1..1

GateResult

Text	Meaning:	Description
fail	None
inconclusive	None	The 95% bootstrap interval straddles the threshold (G1 report, T033).
pass	None

Used by

• GateCheck gateCheck__result _1..1
• GateVerdict gateVerdict__result _1..1

ManoeuvreKind

What a target manoeuvre changes (generator 0.5.0, SRD Annex C §C.2).

Text	Meaning:	Description
course	None	A rate-limited course change at constant speed (an arc).
speed	None	A constant-acceleration speed change on a constant course.

Used by

• ManoeuvreSpan manoeuvreSpan__kind _1..1
• TargetManoeuvre targetManoeuvre__kind _1..1

ModelKind

Model families of SRD §4.

Text	Meaning:	Description
causal_cnn	None	Causal 1-D CNN control for the instantaneous variant.
instantaneous	None	Causal single-layer LSTM (operator's running solution).
learned_gain_filter	None	Phase 4 tracker T1 (spec 012, SRD Annex C §C.4): a constant-velocity filter with a bearings-only update whose gain a GRU of units computes each step, in one declared filter layer (constitution 1.3.0). Reads the raw feature set; outputs the velocity parameterisation.
mlp	None	MLP over the flattened window, control for the time-late variant.
time_late	None	Bidirectional LSTM over the whole window.

Used by

• TrainingConfig trainingConfig__model_kind _1..1

OutputParameterisation

How the model's four output channels encode the target's course and speed (spec 008).

Text	Meaning:	Description
course_speed	None	log_range_km, course_sin, course_cos, speed / speed_scale_mps (iterations 1–3).
velocity	None	log_range_km, vx / velocity_scale_mps, vy / velocity_scale_mps, and a fourth channel that carries no loss and is ignored; course and speed are derived from (vx, vy).
velocity_gaussian	None	Spec 012 (phase 4, tracker T2): six channels per row. log_range_km, vx, vy (as velocity), then the log-variance of each, trained by the Gaussian negative log-likelihood. The reported estimate is the mean; prediction rows carry the standard deviations (log_range_sd, vx_sd_mps, vy_sd_mps), reported and their calibration measured, never gated.
velocity_two_hypothesis	None	Spec 011 (phase 3b): seven channels per row. log_range_km, vx, vy (as velocity) under the hypothesis that the target has not turned; the same under the hypothesis that it has; and a logit of P(the target's turn has started by this row). Each hypothesis is trained only on rows where it is true (the label comes from the stored truth course); the logit by binary cross-entropy. The reported estimate is the more probable hypothesis's, never an average.

Used by

• TrainingConfig trainingConfig__output_parameterisation _0..1

PredictionMode

Text	Meaning:	Description
final	None	One row at the last timestep of the window.
per_timestep	None	One row per post-zig timestep using bearings up to that time.

Used by

• PredictionHeader predictionHeader__mode _1..1
• Prediction prediction__mode _1..1

RangeDistribution

Text	Meaning:	Description
log_uniform	None
uniform	None

Used by

• ParameterRanges parameterRanges__range_distribution _1..1

RecordKind

Discriminator for records in NDJSON files.

Text	Meaning:	Description
detection	None	One Detection (schema 0.19.0, spec 017).
header	None	Dataset or prediction file header (first line).
prediction	None	One Prediction.
scenario	None	One Scenario.

Used by

• DatasetHeader datasetHeader__record _1..1
• DetectionHeader detectionHeader__record _1..1
• Detection detection__record _1..1
• PredictionHeader predictionHeader__record _1..1
• Prediction prediction__record _1..1
• Scenario scenario__record _1..1

Split

Dataset split. Train/validation/test are 80/10/10 by scenario index.

Text	Meaning:	Description
heldout	None	Held-out geometry with parameter ranges outside training.
test	None
train	None
validation	None

Used by

• DatasetHeader datasetHeader__split _1..1

StopReason

Text	Meaning:	Description
diverged	None
early_stopping	None
max_epochs	None
wall_clock	None

Used by

• ModelManifest modelManifest__stop_reason _0..1

TrackId

Text	Meaning:	Description
ownship	None
target	None

Used by

• Track track__id _1..1

ZigRule

How each zig's direction and size are chosen (generator 0.3.0, spec 007).

Text	Meaning:	Description
across_los	None	Turn across the line of sight: the new course puts the target on the other bow at an angle off drawn from los_offset_rad, by the shorter turn. Uses the true bearing at the last sample before the turn.
random	None	Size from zig_angle_abs_rad, side uniformly ±, regardless of the target.

Used by

• ParameterRanges parameterRanges__zig_rule _0..1