Docs / Guide
Matrix policies
Recore always records every piece of detected evidence in the audit files. The matrix policy controls one thing only: which accepted evidence is written into the binary corrected matrix.
Nothing is ever thrown away by a policy. A cell that strict declines to correct is still in gene_presence_absence_recore_states.csv with its state, and still in orthology_evidence.tsv with the reason.
Choosing one#
| Goal | Policy | What it corrects | What stays audit-only |
|---|---|---|---|
| Paper-ready binary PA matrix | balanced | strict same-locus calls, alias clusters, and full-length sequence evidence | pseudogenes, partial homology, context-rejected hits |
| Maximum conservatism | strict | only strict same-locus and alias-cluster evidence | full sequence-only hits without annotation or synteny support |
| Exploratory rescue | sensitive | accepted strict, alias, sequence-only, partial and pseudogene evidence that passes the quality guard | rejected contexts, weak accepted evidence, and homologous or paralogous traps |
balanced is the recommendation for anything that will be published. It rescues the false absences that motivated the tool without letting a disrupted homologue become a clean presence call.
Reach for strict when the downstream analysis is sensitive to false positives in a way that outweighs the false absences, for example a core-genome definition that feeds a phylogeny.
Reach for sensitive when you are exploring, and read the multistate matrix alongside it rather than the binary one.
The adaptive orthology guard#
Policy is not the only gate. Every accepted rescue is also scored by an adaptive guard that combines:
- sequence identity and query coverage;
- locus and synteny evidence;
- ORF integrity;
- the evidence state;
- cluster prevalence;
- representative sequence length.
Rare clusters and very short clusters have to clear a higher sequence bar, because they are the easiest to over-call from weak homology. Same-locus evidence such as a passing synteny context can override that stricter sequence floor for STRICT_PRESENT and ALIAS_CLUSTER calls, but weak sequence-only evidence is held back as audit_only_quality.
The guard shows its work in orthology_evidence.tsv:
| Column | Meaning |
|---|---|
adaptive_identity_floor | the identity this cluster had to clear |
adaptive_coverage_floor | the coverage this cluster had to clear |
adaptive_sequence_pass | whether the sequence evidence cleared both |
orthology_confidence_score | combined score |
orthology_confidence_tier | the score bucketed for reading |
risk_flags | what made this call risky |
So a cell that was not corrected under balanced can be traced to either the policy or the guard, and the two are distinguishable.
Reporting it#
The wording that matches what balanced actually does:
We used Recore with the balanced matrix policy, which rescues full-length false absences while retaining partial or disrupted homologues as multistate audit-only evidence.
If a specific cluster needs its own sentence, orthology_evidence.tsv carries per-cluster wording guidance, and the run report shows it in the inspector under recommended wording.