Crosswalk: this storage schema ↔ BIDS Stats Models and NIDM-Results

Written against the BIDS Stats Models specification and NIDM-Results 1.3.0, read at field level. Companion to ars-crosswalk.md, which covers CDISC's Analysis Results Standard; that document is referred to here as the ARS crosswalk and its findings are not repeated.

Neither standard is in a producer/consumer relationship with this schema, and nothing here is imported. This is a crosswalk for design comparison, not an executable map. The two were named in ModelEstimation.see_also from the beginning and never read against the schema; this is that reading.

---

1. Three different things are being represented

Almost every difference below follows from one asymmetry, so it is worth stating once rather than per row.

	unit of representation	when it is written	can it be re-executed
BIDS Stats Models	a recipe: how to fit a model over a BIDS dataset	before fitting	yes — that is the point
NIDM-Results	a record of an execution: the PROV graph SPM/FSL/AFNI emitted	after fitting, by the tool	no, but every input is hashed
CDISC ARS	a planned analysis and its reported table cells	before and after	partly (WhereClause, programmingCode)
here	a reported tested effect, recovered from a published document	after publication, by an extractor	no

BIDS Stats Models says so in its own words: it is "prescriptive (as in a recipe), rather than descriptive (as in summary of already fit models)."

Three consequences run through everything that follows.

A prescriptive standard has no results. BSM has no threshold, no correction, no statistic value, no direction of outcome—inference is the fitting tool's business. BSM is therefore less expressive than both this schema and NIDM-Results across the whole inference surface, and the comparison there is one-sided.

A tool-exported standard has no silence. Neither BSM nor NIDM has any counterpart to not_reported, because a pipeline that ran knows what it did. Every field in NIDM-Results is either present or the exporter chose not to emit it, and there is no way to say "the source did not report this." That single fact accounts for a whole class of fields this schema has and they cannot.

A standard bound to data does not need semantics in the document. BSM Model.X entries are column names in events.tsv and participants.tsv; NIDM is explicit that design-matrix regressor names are abstract identifiers. What a regressor means is recoverable by opening the data. This schema has no data, so the meaning has to be in the record — which is why FactorLevel points at typed study entities and ModelTerm carries variation_level, unit, and assessment.

A subject-matter vocabulary is not a fourth row. ONVOC, Cognitive Atlas, CogPO, BrainMap's taxonomy, MeSH and UBERON are a different kind of artifact: they supply values, where the three standards above supply structure. None of this table's columns means anything for a SKOS concept scheme, so they are deliberately not crosswalked here, and no field in this schema binds one — storage-schema-design-notes.md §"Conditions belong to a task" gives the reason. What a vocabulary can be checked against instead is whether a field exists whose free-text value a later normalization step could map onto it, which is a per-field audit rather than a crosswalk.

---

2. The crosswalk

2.1 BIDS Stats Models

Containers and the stage graph

BSM	Here	Fit
BIDSStatsModel{Name, BIDSModelVersion, Description, Input, Nodes, Edges}	Study + its ModelEstimation records	weak. A BSM document is one model spanning every stage; a Study is a publication containing many models.
Input (entity filter selecting the data)	Analysis.acquisitions, .tasks	partial. Both say what the model ran on — an entity query there, a reference to a described protocol here.
Node{Name, Level, GroupBy, Transformations, Model, Contrasts, DummyContrasts}	ModelEstimation	the central row, and now largely parity. A Node is one stage and so is a ModelEstimation; what a Node holds and this does not is GroupBy and Transformations. See finding 5.
Node.Level ∈ Run, Session, Subject, Dataset	ModelEstimation.stage (free text)	weak, and deliberately. Same concept, closed vocabulary there and open string here — the stage order is inputs_from, so nothing reads the label.
Node.GroupBy	—	no equivalent. How inputs partition into estimation units is not represented.
Edges[{Source, Destination, Filter}]	ModelEstimation.inputs_from	partial. Carries "this stage consumed that stage," pointing from consumer to consumed; Filter, which of the source's contrasts feed on, has no counterpart. See finding 5.

The model

BSM	Here	Fit
Model.X (design matrix variables)	ModelEstimation.terms : ModelTerm[]	parity of role, and this schema is richer. Both are the columns. X is a list of name strings; ModelTerm carries type, levels, unit, provenance and variation level.
Model.Formula (Wilkinson notation)	ModelTerm.interaction_with, .functional_form	partial. A composable algebra there — nesting, expansion, crossing — against two slots here. See finding 8.
Model.Type ∈ glm, meta	ModelEstimation.model_family (open) + which AnalysisDetails subclass is filled	partial, in this schema's favour. Two values there; seven families and eight method payloads here.
Model.HRF{Variables, Model, Parameters}	ModelEstimation.hrf_model (free text)	weak. Structured basis and parameters there, one string here.
Model.Options (e.g. high-pass cutoff, mask, serial-correlation handling)	ModelEstimation.model_settings (free text)	weak. See finding 7.
Model.Software	ModelEstimation.software, .estimator	parity.
Transformations{Transformer, Instructions}	—	no equivalent. See finding 8.

The contrast

BSM	Here	Fit
Contrast{Name, ConditionList, Weights, Test}	Effect{cells, statistic} + Analysis.name	parity of shape, divergence of content. Both select over the model's columns. ConditionList + Weights is a name-keyed numeric vector; cells is a list of {term, level, direction}.
Contrast.ConditionList (matched to X by name string)	Cell.term (matched by identifier)	this schema is more robust. Carried over from the ARS crosswalk's finding 1.
Contrast.Weights	Cell.direction (positive/negative/undirected/held)	deliberate divergence. Signs, not magnitudes. See finding 6.
Contrast.Test ∈ t, F (plus a pass-through that runs no test)	Statistic.family (open: t, z, f, chi_square, likelihood_ratio, beta, correlation)	this schema is broader, as it must be: papers report statistics BSM has no reason to prescribe.
DummyContrasts{Contrasts, Test}	—	not applicable. A convenience for generation; there is nothing to generate from a document.
—	Effect.mediation, Analysis.prespecification, Measure, InferenceSettings, Region, spatial_scope, spatial_unit	no counterpart in BSM.

2.2 NIDM-Results

The provenance chain

NIDM-Results	Here	Fit
ModelParameterEstimation (activity) used Data, DesignMatrix, ErrorModel, MaskMap	ModelEstimation	parity of role. Both are "how the model was fit," shared by every contrast taken from it.
ContrastEstimation (activity) used DesignMatrix, ParameterEstimateMap, ContrastWeightMatrix	Effect	parity of role. One per reported contrast in both.
Inference (activity) used StatisticMap, thresholds, criteria	Analysis.inference_settings	partial, and the cardinality differs. A separate act there, an inlined single-valued object here. See finding 4.
the PROV edges themselves (used, wasGeneratedBy)	implicit in the reference structure	weak. Traversable there; here the chain is recovered from which record points at which.
Data{grandMeanScaling, targetIntensity, hasMRIProtocol}	Acquisition, Preprocessing	partial. Global scaling and target intensity have no structured home here.
StudyGroupPopulation{groupName, numberOfSubjects}	AnalysisGroup{group, n} + Group	this schema is far richer — Group carries a recruitment funnel, demographics, clinical characterization.

The design matrix

NIDM-Results	Here	Fit
DesignMatrix{regressorNames, hasHRFBasis, hasDriftModel, format, fileName}	ModelEstimation.terms, .hrf_model	partial. NIDM says outright that regressorNames are abstract identifiers; the matrix itself is attached as a file. This schema has no file and so has to carry the semantics.
hasHRFBasis (a closed ontology: canonical, gamma-difference, gamma, Gaussian, FIR, Fourier, spline, sine, derivatives, custom)	hrf_model (free text)	weak. See finding 7.
hasDriftModel (DCTDriftModel{cutoffPeriod}, GaussianRunningLineDriftModel{cutoffPeriod})	—	gap, folded into model_settings. See finding 7.
ErrorModel{errorDistribution, errorVarianceHomogeneous, errorDependence, varianceMapWiseDependence, dependenceMapWiseDependence}	ModelEstimation.model_settings (free text)	gap. See finding 7.
withEstimationMethod (OLS, GLS, WLS, IRWLS)	ModelEstimation.estimator	parity.
ContrastWeightMatrix{contrastName, statisticType, value}	Effect.cells + Statistic.family	deliberate divergence, as with BSM. See finding 6.

Inference and results

NIDM-Results	Here	Fit
StatisticMap{statisticType, contrastName, effectDegreesOfFreedom, errorDegreesOfFreedom, ...}	Statistic{family, degrees_of_freedom_numerator, degrees_of_freedom_denominator}	parity, and closer than it looks: NIDM's effect/error split is exactly the numerator/denominator split, for the same reason.
hasAlternativeHypothesis ∈ one-tailed, two-tailed	—	gap. See finding 3.
HeightThreshold{value, correctionMethod, equivalentZ, pValue}	InferenceSettings.height_threshold_value, .height_threshold_type, .multiple_comparison_method	partial, and the attachment differs. NIDM binds the correction to the threshold; here it floats beside several thresholds. See finding 2.
ExtentThreshold{value, equivalentZ, pValue}	.cluster_extent_threshold, .clusterwise_threshold_value	partial, same attachment problem.
ClusterDefinitionCriteria{clusterConnectivityCriterion}	.neighborhood_definition	parity.
PeakDefinitionCriteria{peakStrategy, minDistanceBetweenPeaks, maxNumberOfPeaks}	—	no equivalent, and appropriately: a table's peaks are what the authors chose to print.
SearchSpaceMaskMap{searchVolumeInVoxels/InUnits/InResels, reselSizeInVoxels, noiseFWHM, heightCriticalThresholdFWE05/FDR05, expectedNumberOfClusters}	.search_volume (free text), .number_of_tests	gap. RFT quantification has no structured home.
ExcursionSetMap	StatisticalMap{map_type, is_thresholded, url}	partial. is_thresholded: true is roughly an excursion set.
SupraThresholdCluster{clusterSizeInVoxels, clusterSizeInResels, pValueUncorrected, pValueFWER, qValueFDR}	— (delegated to Table)	gap within the schema. See finding 1.
Peak{coordinate, value, pValueUncorrected, pValueFWER, equivalentZStatistic}	— (delegated to Table)	gap within the schema. See finding 1.
Coordinate{coordinateVector, coordinateUnit}, CoordinateSpace{dimensionsInVoxels, voxelSize, voxelUnits, voxelToWorldMapping, numberOfDimensions, inWorldCoordinateSystem}	Table.coordinate_space / Analysis.coordinate_space (string)	partial → gap. Only the world coordinate system survives. See finding 3.
ParameterEstimateMap, ContrastMap, ContrastStandardErrorMap, ResidualMeanSquaresMap, GrandMeanMap, ReselsPerVoxelMap, MaskMap, DisplayMaskMap	StatisticalMap (one class, map_type free text)	partial. Eight typed map classes there, one class with a type string here — right for a schema that records whichever maps a study happens to share.
per-map sha512, format, location, MapHeader	StatisticalMap.url	weak. No checksum.
ConjunctionInference, PartialConjunctionInference{partialConjunctionDegree}	ConjunctionDetails{null_hypothesis, implementation, components}	mostly this schema's favour, with one exception. global_null/conjunction_null and the implementation are richer than NIDM's split; NIDM's u-of-n degree has no home here.

2.3 No counterpart in any of the three

BSM / NIDM only	Here only
Node.GroupBy, Edges — the stage graph	Cell.direction as the sign of a fitted coefficient
Transformations — a variable-manipulation DSL	FactorLevel.{conditions, groups, timepoints, arms} — typed level referents
Model.Formula — a composable design algebra	ModelTerm.variation_level, and the EffectKind derivation
ErrorModel, hasDriftModel	Measure{family, type, source_label, specific_metric, unit}
Peak, SupraThresholdCluster, Coordinate values	Region.definition_method — independence as a claim
sha512 and file-level integrity	Analysis.prespecification
hasAlternativeHypothesis	Effect.mediation{path, mediator}
SearchSpaceMaskMap RFT quantities	DecodingDetails, SimilarityDetails, ConnectivityDetails, LatentDecompositionDetails
CoordinateSpace geometry	NotStructurableDetails, model_representation_notes, undirected vs held vs a not_reported direction

---

3. Findings

1. On results, NIDM-Results is far more expressive, and this is a boundary rather than a gap

ExcursionSetMap → SupraThresholdCluster → Peak → Coordinate is a complete, typed, queryable results payload: cluster size in voxels and resels, corrected and uncorrected p-values at both cluster and peak level, equivalent z, coordinate vectors with units. The analysis module here has none of it. Table carries a caption, a footer, column headings and a coordinate_count; the points themselves live in the repository's coordinate representation, reached through source_path.

Worth stating plainly, because it is easy to miss while reading the schema: there is no effect size, p-value, or confidence interval anywhere on an Analysis. PerformanceMetric.value — a decoding accuracy — is the only reported number in the entire analysis module. ARS stores its numbers inline in OperationResult; NIDM stores them per peak and per cluster; this schema stores none.

That is a boundary, not an oversight: the module's job is to decide whether two maps are poolable, and the coordinates are handled by a layer built for them. It is worth recording because the boundary is invisible from inside analysis.yaml, and because it means a query that filters on statistic values cannot be answered from this module alone.

2. The correction method floats free of the threshold it corrected

NIDM attaches correctionMethod to a HeightThreshold or ExtentThreshold, so each threshold carries its own. InferenceSettings has three threshold slots — height_threshold_value, clusterwise_threshold_value, cluster_extent_threshold — and one multiple_comparison_method beside them.

The commonest sentence in the fMRI literature has two corrections in it: voxelwise p < 0.001 uncorrected, clusterwise FWE-corrected p < 0.05. That is none at the voxel level and FWE at the cluster level, and there is one field for both. An extractor must pick one, and whichever it picks, the record asserts something false about the other threshold.

This was found while writing the crosswalk rather than being in the plan, and it is the cheapest of the gaps to close: a small Threshold class with {level, value, type, correction_method}, multivalued on InferenceSettings, subsumes all three current slots and the correction. It is also the most invasive to extraction-to-storage.map.yaml for the same reason.

Half of it has since been done, for a different reason. voxelwise_threshold_value and cluster_forming_threshold_value were one fact under two names — the height threshold, called after whichever of its two jobs a paper was describing — and no record in the corpus ever filled both, so they are now one height_threshold_value. That leaves the level count at three and the correction still floating; the Threshold class remains the fix for the attachment.

3. Two facts a meta-analysis needs are not recorded: tailedness and voxel size

Tailedness. NIDM has hasAlternativeHypothesis — one-tailed or two-tailed. Nothing here does. This is not a reporting nicety: converting a reported p to a signed z requires knowing the tail, and that conversion is a routine step in coordinate-based meta-analysis. Cell.direction implies a one-tailed test in the ordinary case and an omnibus implies two, but neither is a statement, and papers do report two-tailed directional tests. A closed one_tailed/two_tailed enum on Effect would cost one slot.

Voxel size. NIDM's CoordinateSpace carries dimensionsInVoxels, voxelSize, voxelUnits, voxelToWorldMapping and numberOfDimensions alongside the world coordinate system. Here, coordinate_space is a string naming the template — the world system and nothing else. The consequence is direct: InferenceSettings.cluster_extent_threshold is documented as a count of "voxels, vertices, or the unit stated in the source," and a 20-voxel cluster at 2 mm isotropic and a 20-voxel cluster at 4 mm differ eightfold in volume. The threshold as recorded is not comparable across studies, which is what the field exists for.

Acquisition.acquisition_voxel_size_mm is not the answer: it is the acquired resolution, before normalization, and analyses are thresholded in the resampled space.

4. NIDM separates estimating a contrast from drawing inference on it; here they are one record

ContrastEstimation and Inference are distinct PROV activities, so n inferences hang off one contrast estimate. Analysis.inference_settings is inlined and single-valued.

A paper that reports one contrast at whole-brain FWE and at an uncorrected exploratory threshold — routine, and often the whole point of a supplementary table — has two inferences over one effect. The schema's options are to mint two Analysis records that duplicate the effect, the sample, the model reference and the measure, or to drop one threshold. Neither is right, and the separate-Analysis rule in Analysis.description does not cover this case: it lists direction, group, method, cell pattern, seed, decoded variable, component identity and spatial scope, and thresholding is not among them, correctly — two thresholdings of one map are not two tested effects.

Making inference_settings multivalued is the small fix and breaks nothing. Note that it composes with finding 2 rather than substituting for it: multivalued InferenceSettings fixes two inferences on one contrast, and a Threshold class fixes two corrections within one inference.

5. The stage graph is represented after all — the edges, not the recipe

Reversed. This finding used to read "the flattening is deliberate" and sat on the not-to-be-relitigated list. ModelEstimation.inputs_from was added because the price it recorded came due; what follows is what changed, and what did not.

BSM models the hierarchy explicitly: Node.Level ∈ Run/Session/Subject/Dataset, Node.GroupBy saying how inputs partition into estimation units, and Edges[{Source, Destination, Filter}] saying which node's contrasts feed which. A three-stage analysis is three nodes and two edges.

Here, each stage the source describes takes a ModelEstimation and the higher one names the lower in inputs_from — Edges without Filter. A model's terms are its own plus, transitively, those of the stages it names, so the mixed term list is unmixed and every derivation reads the whole chain.

What forced it. The old defence was that a paper states its group-level design matrix and rarely its first-level one in enough detail to constitute a separate model, so two linked records would come out one populated and one full of not_reported. Extraction found the opposite often enough to matter. In pmid 24600410 the model produced exactly two records — a FEAT first level holding the amygdala seed regressors and the WM/CSF/motion nuisance covariates, and a FLAME group model holding diagnosis, age, sex and scanner — and with no link between them the first was referenced by nothing, no Cell could name the seed the maps were of, and all four contrasts derived as unadjusted for motion. The mixing this finding called "not inert" is the same problem seen from the other side: whether the stages are flattened into one record or split into two unlinked ones, the adjustment set comes out wrong, and the schema had no third option.

The cheapness of the fix is the other half of the argument. inputs_from is optional and adds one slot: a paper that describes one design matrix still takes one record, and level is still free text with no ordering read from it.

Still not represented, and still deliberately:

• GroupBy has no counterpart, so "one model per run, then averaged within subject" and "one model over concatenated runs" remain the same record. That is a statement about how estimation units partition, which is recipe rather than result.
• BSM's Edges[].Filter — which of a node's contrasts feeds the next — has no counterpart either. The link says the stage was consumed, not which output of it was.
• There is still no general "this result was computed from that one" mechanism. ARS has one (ReferencedOperationRelationship); this schema now has four specific ones — interaction_with, ConjunctionComponent, Mediation.mediator, and inputs_from. inputs_from relates models; the analysis→analysis case, where a second model is fitted to a first analysis's map, is F3 in the expressivity probe and remains open. The two are complementary: a first-level model no paper reports a contrast from has no Analysis for an analysis→analysis link to point at.

6. Signs instead of weights: the trade, and its price

BSM Contrast.Weights and NIDM ContrastWeightMatrix.value are numeric vectors. Cell.direction is a sign. Direction's own description states the trade — [1, -1/2, -1/2] and [1, -1, 0] over three levels are both "one positive level against two others" — and the justification is that weights are not recoverable from prose. That is right, and both peers confirm it is a real fork rather than an accident: nobody else records signs, and nobody else is reading a PDF.

The price is worth writing down next to it, in one place:

• A multi-row contrast matrix collapses. An F-contrast in BSM or NIDM is a matrix, and its rank says how many comparisons it spans. Here an omnibus F is a set of cells with direction: undirected, which records that the factor was tested undirectionally but not which comparisons the test spanned. A 3-level factor's omnibus F and a specific 2-row subset of it are the same record.
• Unequal weights collapse. A linear trend across four dose levels and a comparison of the extremes are the same cells. FactorLevel.order recovers the sequence, which is what the ARS crosswalk added it for, but not the coefficients.

Both losses are bounded by what a paper reports, which is the argument for accepting them. Neither is recorded anywhere in the schema at present.

7. On how the model was fit, NIDM is structured where this schema has prose

Three things NIDM makes queryable and ModelEstimation.model_settings holds as free text:

• ErrorModel.errorDependence ∈ independent, exchangeable, Toeplitz, compound symmetry, unstructured — with errorVarianceHomogeneous as a separate boolean, and varianceMapWiseDependence / dependenceMapWiseDependence saying whether each was estimated per element or pooled.
• hasDriftModel — a DCT basis with a cutoff period, or a Gaussian running line with one. BSM reaches the same fact through Model.Options (a high-pass cutoff in Hz).
• hasHRFBasis — a closed ontology of eleven basis sets, against hrf_model as a string.

Of the three, error dependence has the strongest claim on a slot: whether a repeated-measures analysis assumed sphericity is a fact that decides comparability, papers do state it (a Greenhouse-Geisser correction is already accommodated by degrees_of_freedom_denominator being a float, which is a partial admission that this matters), and model_settings cannot be filtered on.

Drift and high-pass filtering are COBIDAS items and are reported often. hrf_model as free text is defensible: the wording is what papers give, and the vocabulary is long-tailed.

8. On specifying a design, BSM is more expressive, and most of it does not transfer

Model.Formula is Wilkinson notation — crossing, nesting, polynomial expansion, all composable. Transformations is a variable-manipulation DSL (Scale, Threshold, Orthogonalize, Convolve, Factor, Product, Sum, Lag, Split). Against that, this schema has ModelTerm.functional_form with four values plus an open string, and interaction_with for products.

Most of the difference is inert here, because a paper does not report its transformation pipeline and a schema cannot extract what is not written. Two exceptions:

• Orthogonalization of parametric modulators is reported, does change what a coefficient means, and has nowhere to go — not functional_form (which is the term's shape, not its relation to other columns) and not interaction_with (which is what makes a column a product). It would currently land in source_definition as prose.
• Random effects structure is conceded rather than represented, which the schema says out loud: model_representation_notes names "random slopes, latent variables, dynamic connectivity, and dependencies between terms" as the things it exists to record the absence of. That is the honest version of the gap, and it is a better answer than a half-built Formula would be.

Also absent, and following from the same root: event timing. BSM works from events.tsv onsets, durations and amplitudes. Condition here is a name and a description; Task.design_type is a free-text "event-related, block, mixed, or naturalistic." A block design and an event-related design of the same contrast are distinguishable; the timing is not.

9. Where this schema is more expressive than both, and why each one had to be

Ranked by how much of the schema's purpose depends on it rather than by size.

1. Direction of a fitted coefficient. All three peers record the contrast specification. On a continuous term, Cell.direction is the sign the coefficient came out with — an outcome, not a specification. With PerformanceMetric.relation beside it, this schema is the only one of the four that answers "which way did it go." It has to be, because that is the finding a paper reports and a synthesis pools. Cell.direction is the whole of it: the connectivity payload used to carry a whole-map sign of its own, and two routes to one direction is one too many.
2. Typed level referents. FactorLevel.{conditions, groups, timepoints, arms, regions} says what a level is. No peer can distinguish a cohort comparison from a pre-post change from a crossover comparison structurally: BSM has column-name strings, NIDM has declared-abstract identifiers, ARS has an untyped WhereClause.
3. ModelTerm.variation_level, and therefore the parametric-modulation vs cross-subject-regression distinction. Absent from all three, and it separates two effects that are otherwise the same record.
4. The EffectKind derivation. No peer answers "what kind of effect is this" at all. With weights and a design matrix you could work out a crossing; without variation_level you could not reach the modulation/regression split at any effort.
5. Method families beyond the mass-univariate GLM. NIDM-Results is mass-univariate only. BSM Type is glm or meta. The seven AnalysisDetails subclasses — decoding with per-metric reference_value, RSA, connectivity with EdgeDirectionality and InferenceTarget, latent decomposition with second_block and PolaritySemantics — have no counterpart anywhere in the comparison set.
6. Measure. The scientific quantity of the map. For BSM and NIDM it is implicit in the input files; ARS has no imaging measure. A schema deciding poolability cannot leave it implicit — grey-matter volume and grey-matter density are deliberately not collapsed here.
7. Region.definition_method as an independence claim. same_study_analysis marks a circular ROI, and Analysis.defines_regions names the analysis it was drawn from, so the circularity is checkable rather than asserted. No peer represents non-independence in any form, and it is a first-order threat to the validity of a pooled estimate.
8. The epistemic layer. prespecification; interpretations, marked as inferred by which slot it sits in; undirected, held and a not_reported direction as three different facts; NotStructurableDetails; model_representation_notes; reason_first_class_type_not_used. None of this can exist in a prescriptive or tool-exported standard, because neither has an author who might not have said. ARS's reason and purpose are the only partial peers, and the ARS crosswalk already settled why Prespecification is not AnalysisReasonEnum.
9. Effect.mediation — direct, indirect, total. Without it an indirect-effect map is structurally identical to a regression on the same predictor.

One thing carried forward from the ARS crosswalk, now confirmed across the full set: on reference mechanism, BSM binds contrasts to design columns by name string, NIDM by vector position, ARS and this schema by identifier. Of the four, the two document-derived LinkML schemas are the two that use ids.

---

4. What this leaves open

Nothing here was acted on. Five candidates, in the order their claim on the schema's stated purpose runs strongest to weakest:

gap	finding	candidate shape	cost
correction method per threshold	2	a Threshold{level, value, type, correction_method} class, multivalued	one class; subsumes four slots; the largest map change
tailedness	3	Effect.alternative_hypothesis: one_tailed/two_tailed, closed	one slot, one enum
n inferences per contrast	4	make inference_settings multivalued	loosens a cardinality; breaks nothing
voxel size of the analysed space	3	voxel_size_mm on Table, or a CoordinateSpace class	one slot, or one class done properly
error dependence	7	error_dependence + error_variance_homogeneous on ModelEstimation	two slots, one enum

Settled and recorded, not to be relitigated: signs rather than weights and the two losses that follow (finding 6), Formula and Transformations (finding 8), event timing (finding 8), and the delegation of all result values to the coordinate layer (finding 1).

Noted and not worth a slot: NIDM's partialConjunctionDegree, grand-mean scaling and target intensity, per-map checksums, RFT search-space quantities, and PeakDefinitionCriteria.