Contrast types: baseline differences vs. responses

resilience_classification says what a comparison measured: a resilience outcome, a disease response, a stress response, or exposure alone (resilience_vs_exposure.md). It does not say between whom or when. Two comparisons can both be resilience evidence and still ask different questions:

Do intertidal oysters, which tolerate heat better, express a gene differently from subtidal oysters before either is heated?
How does that gene respond when an oyster is heated?

A breeder can only use the first kind directly. A marker that already differs between resilient and susceptible animals under ordinary conditions can be read before any stress, on animals that are then kept. A gene that responds to heat tells you an animal is being heated.

So every comparison also declares a contrast_type (comparisons[].contrast_type in schemas/study.schema.json, required).

The four contrast types

`contrast_type`	The two arms are	Sampled	Example
`baseline_phenotype_difference`	groups that differ in resilience by origin: line, family, population, ploidy or age	under common, unstressed conditions	LI2018 intertidal vs. subtidal populations at 0 h
`phenotype_difference_under_exposure`	groups that differ in resilience, by origin, outcome or prior conditioning	during or after a challenge	CIBNOR2021 heat-resistant vs. heat-susceptible families during heat; CORPOREAU2014 high vs. low viral load
`exposure_response`	a treatment and its control	any time	heat vs. control; infected vs. uninfected; starved vs. fed
`phenotype_by_exposure_interaction`	the difference between two groups’ responses to one exposure	in a 2 × 2 design	resistant line’s heat response minus susceptible line’s heat response

How to choose:

Do the two arms differ by something done to them? If so, it is an exposure_response, even when sampling happens before a later challenge. DUPERRET2025’s starved vs. fed oysters, sampled before infection, are an exposure response: starvation made them more resistant, but the contrast is a diet treatment, not a difference between resilient and susceptible animals.
Do the arms differ in resilience by origin, and were both sampled under the same unstressed conditions? That is a baseline_phenotype_difference. The resilience difference must be measured, in the study or in a cited source, not assumed; record where in limitations. A baseline contrast is resilience evidence by construction, so validation warns if its resilience_classification is anything but resilience.
Do the arms differ in resilience but were sampled during or after the challenge? That is a phenotype_difference_under_exposure. A difference here can be a consequence of the outcome rather than a cause: dying animals express different genes.

When in doubt, choose exposure_response. It is the type that claims least.

What the pipeline does with it

Pooling. contrast_type is part of the meta-analysis grouping key and of each Benjamini-Hochberg test family. A baseline contrast never pools with an exposure response, even under the same phenotype and feature. Cross-species pooling through orthogroups keeps the same split.
Primary comparisons. meta_analysis_primary is chosen within each contrast type. A study whose primary comparison is a heat response still contributes its baseline contrast to the baseline pool. Before contrast types, LI2018’s intertidal-vs-subtidal contrast was set aside because the study’s primary comparison was the 35 °C response.
Multi-omics convergence counts layers significant under the same phenotype and contrast type. A methylation difference at baseline and an expression response to heat are two questions, not convergence.
Ranking. Every candidate carries its contrast_type. A candidate whose contrast type is baseline_phenotype_difference and whose evidence class is resilience_associated is flagged is_predictive_baseline. The flag does not change its tier or score: tier still says how well replicated it is.
Reporting. aree top-candidates opens with a “Predictive baseline candidates” section. Evidence cards state the contrast type and what it does and does not show. A card for a non-default contrast type has the type as a filename suffix, so it never overwrites the exposure-response card for the same feature, and its per-study table names each comparison, so the sign of the effect can be read against the arms. The query layer and MCP server can filter candidates by contrast_type.
Website. The home page has a “Predictive baseline candidates” section: the significant ones, by tier then score, with gene names and links to their evidence cards, and a plain statement when all rest on a single study. On the Explore page, the Candidates tab filters by contrast type and has a “Predictive baseline only” toggle; the candidates table carries gene_symbol and gene_description, and is_predictive_baseline is a true/false column usable directly in SQL.
Old evidence. An evidence table harmonized before this field existed has no contrast_type. The meta-analysis refuses it and names the studies to re-harmonize, rather than defaulting them to exposure_response.

What replication means for a baseline candidate

Groups that differ in origin differ in more than resilience: genetic background, growth rate, the site they came from. One resistant-vs-susceptible pair cannot tell a resilience marker from a background difference. Treat k_studies for a baseline candidate as the number of independent group pairs that agree, and look for:

the same difference in independent line, family or population pairs (the high-priority tier’s two-study gate);
ideally, individual-level validation: sample animals non-lethally before a challenge, record who survives, and test whether the marker predicts it.

The second is the step that turns a predictive candidate into a marker.

Registered baseline contrasts

Study	Comparison	Status
LI2018_HEAT_TIDAL	`intertidal_vs_subtidal_0h`	harmonized
DECICOMP2025_POMS_AGE	`age28mo_vs_age4mo_basal` (primary), `age16mo_vs_age4mo_basal`	not started
DECICOMP2025_POMS_AGE_EMSEQ	`age28mo_vs_age4mo_basal` (primary), `age16mo_vs_age4mo_basal`	not started

The DECICOMP2025 contrasts compare ages, and age differs in development as well as POMS resistance; their limitations should say so when they are run.