# Changelog

## v0.9.0 current-candidate API hardening (2026-07-25)

- Made `api/v1/candidates.json` current-first: current state and action are
  top-level, `current_rank` is null and historical v0.6 fields are nested under
  `historical_evidence_v06`.
- Updated the representative gene index so GFOD2 and the other 20 candidates
  display current actions before historical priority labels.
- Added the four evidence-transition endpoints to OpenAPI.
- Added regression checks for current-first API semantics, GFOD2 index wording
  and single-instance generated sections.
- Preserved the v0.9.0 evidence tables, historical ranks and canonical SQLite
  bytes; no new negative-effect or treatment claim was introduced.

## v0.9.0 candidate evidence transition (2026-07-24)

- Removed the GFOD2 “Priority A” homepage demonstration and the single-winner
  framing from current reader-facing surfaces.
- Added current status, status history, external validation registry and
  evidence-transition rule objects in TSV, JSON and SQLite.
- Retained GFOD2 historical computational rank 1 for reproducibility while
  changing its current state to deprioritized pending auditable external
  evidence.
- Recorded the cross-project update as author-reported with quantitative source
  import pending; it is not presented as a validated negative-effect claim.
- Changed the current candidate display to action-then-gene ordering with no
  numeric current rank. Historical A/B/C labels and ranks remain downloadable.

## v1.4.0-candidate evidence architecture and release identity (2026-07-17)

- Added dedicated DMD direction-prior, applicability/evidence-gap,
  models/methods and release/citation pages.
- Renamed the reader-facing candidate route to Experimental Hypotheses and
  separated evidence class, operational action, predictive support, source
  agreement, rank interval, expression, myoblast-screen state, dependency
  caution and missing evidence; no single score is authoritative.
- Made gene records source-conflict-aware with DMD source-agreement modules,
  per-source effects and compact prefix-partitioned JSON endpoints.
- Expanded OpenAPI with servers, reusable parameters, response schemas,
  missing/null semantics and new evidence/release endpoints.
- Added a machine-readable release identity, claim lock and automatic
  manuscript–site fact checker while keeping DOI, licence and contact fields
  explicitly pending.

## v1.2.0 interpretation refresh (2026-07-17)

- Reframed the homepage endpoint card as an audit contrast: 0/16 strict versus
  16/16 legacy, explicitly stating that the discrepancy is the finding.
- Added an intuitive three-endpoint explainer and a visible explanation of why
  RMSE can improve while raw direction fails.
- Promoted the DepMap-ablation interpretation above the detailed benchmark
  tables with exact full, DepMap-only and no-DepMap mean RMSE values.
- Added an accessible warning icon, highlighted row and explicit follow-up text
  for Priority-A candidates below 1 TPM in skeletal muscle; ZFP69B is the
  current affected row.
- Preserved the v1.0.2 scientific content lock, candidate rules, benchmark
  values, database bytes and canonical release checksums.

## v1.1.0 website refresh (2026-07-17)

- Rebuilt the first-visit path around three user tasks, a static GFOD2 evidence
  demonstration, quick gene examples and search-result priority labels.
- Reordered navigation to place Claim boundary before Benchmarks, Genes and
  Candidates.
- Added a two-minute benchmark gate matrix, prominent split-dependence warning,
  collapsible detailed tables and a model-family selector.
- Added pure-CSS confidence-interval plots and highlighted, explicitly sorted
  feature-ablation rows without changing the underlying values.
- Added Priority-A coverage notes for ZFP69B and ZNF236, sortable candidate
  columns and distinct fusion-screen assessed/not-assessed styling.
- Added bounded evidence-level indicators to static and lazy gene pages.
- Expanded the Help glossary to 15 linked concepts and reframed Claim boundary
  supported content as evidence facts with the 16/16-versus-0/16 audit contrast.
- Added task-led download routes for candidate-table reproduction, all-data SQL
  queries and release-integrity verification.
- Retained lazy bucketed gene search, responsive/accessibility controls, the
  v1.0.2 scientific content lock and the unchanged canonical release checksums.

## v0.6.0-v21-evidence-aligned-candidate-20260717

- Built public static database-resource snapshot.
- Added SQLite database and JSON API artifacts.
- Added gene detail pages and searchable index.
- Synced official GEARS 0.1.2 five-seed, 20-epoch V100 benchmark artifacts.
- Added GEARS same-coverage summary, gates, and GEARS-vs-ridge comparisons as
  an official negative/control deep-model baseline.
- Synced official scGPT 0.2.5 five-seed, 15-epoch V100 benchmark artifacts.
- Added scGPT checkpoint-supported same-coverage summary, gates, and
  scGPT-vs-ridge comparisons as an official negative/control foundation-model
  baseline.
- Added NAR Database Issue gap matrix, help pages, maintenance draft, and inquiry draft.
- Added paired bootstrap/randomization, Holm correction, and a strict composite held-out-gene gate.
- Added feature-family ablation, leakage sensitivity, and shuffled negative controls.
- Added GSE293514 human-myoblast fusion phenotype statistics and candidate safety flags.
- Added a same-split Nature Methods linear-PCA strong baseline with explicit target-gene coverage.
- Added source-signature sensitivity, candidate-tier multiverse, 21-gene design
  sensitivity, ridge effect-magnitude, and HepG2-to-muscle applicability audits.
- Added four-source pairwise concordance, GTEx v8 skeletal-muscle expression
  thresholds, exact NMD support-score decomposition, and an explicit
  point-estimate-versus-strict-gate decision audit.
- Removed the word `safe` from A-tier candidate labels; A now means priority for
  muscle-context validation only.
- Added repeated-split overlap/gene-reuse, target-level cell/batch/construct
  reliability, DMD source-specific modules and meta-analysis readiness.
- Added a DepMap 26Q1 leave-HepG2-out rebuild and five-seed sensitivity audit.
- Replaced the primary candidate interpretation with evidence class plus action
  priority; legacy A/B/C fields remain only for release reproducibility.
- Preserved claim boundary: STATE is a negative/control benchmark; direct feature-to-delta heads show preliminary point-estimate improvement but 0/4 pass the strict test gate.

## v0.6.0-v21 candidate — 2026-07-17

- Added 27 v2.1 evidence tables and matching JSON/TSV objects.
- Added full 2,160-target OOF primary results and historical/primary separation.
- Added coverage-selection, reliability-effect and implementation-sensitivity audits.
- This candidate remains deployment- and DOI-pending.

## v0.7.0-v22 candidate — 2026-07-18

- Added ten repeated balanced-fold stability tables in SQLite/JSON/TSV.
- Added joint realization/target hierarchical-bootstrap uncertainty.
- Reclassified raw directional evidence as mixed because its joint interval crosses zero.
- Added Supplementary Figure S14 and independently verified fold-stability objects.
- This candidate remains deployment- and DOI-pending.