Algorithm versions

Versioned algorithms and the versioning policy.

Every score in DataCourt comes from a named, versioned algorithm. Each audit stores:

algorithm_versions: the version of every stage that ran;
config and config_hash: the fully resolved configuration and its SHA-256;
the embedding backend identifier, including its model revision.

The same registry is served live at GET /api/v1/algorithms and shown on the Models & algorithms page. It is defined in services/api/datacourt/registry.py, which imports each version constant from its implementing module, so the registry cannot drift from the code.

Current versions

Area	Version	Output type	Summary
Configuration	`audit-config-v1`	measured	Versioned default thresholds; project overrides are validated and stored with each audit
Ingestion	`phash-dct32-v1`	measured	SHA-256 identity; 64-bit DCT perceptual hash and dHash
Ingestion	`pixel-sha256-rgb8-v1`	measured	SHA-256 of the decoded, normalised pixels: the same picture in any file format
Quality	`attributes-v1`	measured	Per-image brightness, contrast, normalised sharpness, entropy, edges, colour, border, aspect
Quality	`quality-rules-v1`	heuristic	Absolute and dataset-relative rules for potential blur, exposure, contrast, near-empty, resolution, aspect, extension mismatch
Embeddings	`dc-descriptor-v1`	measured	2,588-d CPU descriptor (default)
Embeddings	`dinov2-small` (`facebook/dinov2-small@<revision>`)	measured	Optional self-supervised ViT-S/14
Duplicates	`dup-families-v2`	measured	Candidates from hashes and kNN, structural verification, families, lineage
Leakage	`leakage-v1`	measured	Cross-split families and adaptive similar-subject clusters
Baseline	`baseline-logreg-v2`	model prediction	Class-balanced logistic head; OOF probabilities; temperature scaling
Training dynamics	`cartography-v1`	model estimate	Confidence, variability, correctness, forgetting
Influence	`tracin-linear-v1`	model estimate	TracIn on the linear head
Labels	`label-forensics-v1`	heuristic	Reliability-weighted witness evidence score
Labels	`rare-or-wrong-v1`	heuristic	Competing hypothesis scores
Shortcuts	`shortcut-detective-v1`	measured	Cramér's V, lift, cue-only accuracy, perturbation tests
Coverage	`coverage-v1`	measured	Map, regions, gaps, collection planner
Privacy	`privacy-scan-v1`	heuristic	Haar face and MSER text-region detectors
Court	`jury-v1`	heuristic	Reason codes, ordered rules R1–R10, uncertainty, priority
Review	`review-budget-v1`	heuristic	Greedy budgeted selection with diminishing returns
Review	`counterfactual-proxy-v1`	model estimate	First-order action effect estimates from the TracIn basis
Experiments	`what-if-v1`	model estimate	Paired retraining with training and evaluation bootstraps
Governance	`debt-v1`	heuristic	Eight debt dimensions with published formulas
Governance	`preflight-v1`	heuristic	Blocking and warning rules
Governance	`contract-v1`	measured	Declarative contract rules over 15 metrics
Versions	`dna-v1`	measured	Descriptive profile and JS-divergence drift
Versions	`version-diff-v1`	measured	SHA-matched version comparison
Versions	`contamination-v2`	measured	Exact and verified near-duplicate overlap with a reference version
Export	`export-v1`	measured	Deterministic clean export with manifest and checksums
Retention	`retention-v1`	measured	Time-based purge of redundant archives

Versioning policy

Patch-level changes (bug fixes that do not change outputs on valid input) keep the version string.
Any change to outputs, including thresholds, formulas, features or ordering, bumps the version: jury-v1 → jury-v2. The previous version is documented here with the reason for the change.
Old audits keep their stored versions and configuration. Comparing a new audit with an old one shows both versions, so differences caused by algorithm upgrades are not mistaken for data changes.
Default configuration changes bump audit-config-vN.

History

Version	Replaced by	Reason
`dup-families-v1`	`dup-families-v2`	Files with identical decoded pixels in different formats (a HEIC and a lossless PNG of it) were near duplicates; v2 makes them exact duplicates. Families otherwise unchanged.
`contamination-v1`	`contamination-v2`	Exact overlap also matches identical decoded pixels, as in `dup-families-v2`.
`baseline-logreg-v1` (SGD for everything)	`baseline-logreg-v2`	SGD optimiser noise made What-if deltas unstable between seeds. v2 uses converged L-BFGS for evidence, metrics and experiments, and keeps SGD only where the training trajectory is itself the evidence (dynamics, TracIn).