Changelog

Unreleased

Reorganized the repository documentation around quick understanding, reproducible demos, code reading, evidence boundaries and interview preparation.
Added a full mining-safety system design that clearly separates implemented label-governance capabilities from planned RTSP, RAG and Agent evolution.
Added a Chinese core-code walkthrough, reproducible no-GPU acceptance guide, project-evidence matrix and high-frequency interview Q&A.
Added repository-local Markdown link validation to CI plus structured bug, feature and pull-request templates.
Expanded the project development and interview handbook with a GitHub-to-code learning route, failure-injection labs and evidence-based presentation guidance.

Added a Spring Boot 4, Vue 3 and MySQL collaboration platform for multi-user human review.
Added JWT authentication, role-based access control and project-level membership isolation.
Added pessimistic task claiming, renewable leases and optimistic versions to prevent duplicate and stale decisions.
Added idempotent bounded-batch import from the existing Python review queue plus read-only visual serving.
Added Docker Compose deployment, Flyway schema management, integration tests and real production review screenshots.

Promoted the offline reviewer into an image-grouped, bilingual human-in-the-loop review application.
Added one-image/many-candidate navigation so overlapping and cross-class evidence can be judged in context.
Added action-aware controls that disable invalid add, replace and evaluation decisions for each case.
Added append-only JSONL journaling, periodic atomic CSV checkpoints and crash-safe resume.
Added deterministic train-first, hard-case and rare-class queue ordering for large review jobs.
Added portable Chinese and English Windows launchers with working-directory and quoted-path safeguards.
Added a grouped synthetic fixture, a real application screenshot and focused persistence/grouping tests.

Added exhaustive GT0_AUTO0, GT1_AUTO0, GT0_AUTO1 and GT1_AUTO1 image/class accounting.
Added pure-Python IoU, IoS, normalized center-distance and area-ratio decision rules for same-target and cross-class conflicts.
Added review-build for portable visual review bundles, full audit CSVs, decision templates and HTML summaries.
Added review-ui, an offline autosaving Tk reviewer with explicit add, replace, evaluation, reject and uncertain actions.
Added review-apply with unresolved-decision blocking, source-GT drift detection, duplicate rechecks and immutable derived datasets.
Added a packaged default six-class policy, public no-GPU fixture, static preview, bilingual documentation, ADR and CI coverage.
Added an anonymized production-scale case study covering 174,426 image-model passes and 30,183 review decisions.

Added the CPU-only, read-only prioritize command for image-level active human-review queues.
Added normalized confidence entropy, dynamically decayed class rarity and greedy dHash/aHash visual diversity.
Added safe resolution of absolute and relative candidate image paths with dataset-boundary enforcement.
Added ranked image/candidate exports, full-pool evidence, failure isolation and a self-contained HTML report.
Added a reproducible imbalanced six-class fixture, static screenshot, bilingual documentation and CI coverage.

Added the model-free, read-only cluster command for perceptual near-duplicate review grouping.
Added 64-bit dHash/aHash fingerprints, BK-tree radius search and deterministic connected components.
Added aspect-ratio and low-texture luminance safeguards to reduce false-positive groups.
Added cross-split near-duplicate leakage evidence, corrupt-image isolation and representative review exports.
Added a reproducible public fixture, self-contained HTML report, static screenshot, bilingual documentation and CI coverage.

Added the model-free consensus command for independent cross-Teacher AUTO verification.
Added deterministic one-to-one spatial matching so one verifier box cannot approve multiple primary candidates.
Added conservative downgrade-to-REVIEW behavior for unsupported primary AUTO candidates.
Added optional agreed-only YOLO label additions, machine-readable summaries and a self-contained HTML report.
Added a reproducible six-class consensus fixture, static screenshot, bilingual documentation and CI coverage.

Added Wilson score precision lower bounds for confidence-aware AUTO threshold calibration.
Made 95% statistical confidence the CLI default while retaining 0 as an empirical-policy compatibility mode.
Expanded the deterministic public calibration fixture from 600 to 2,400 reviewed candidates.
Added empirical precision and precision lower-bound evidence to CSV, JSON and HTML reports.
Added statistical boundary tests demonstrating why small perfect samples are insufficient evidence.

Added pre-generated, path-redacted audit and recovery screenshots to both README languages for offline presentation.
Added the calibrate command for class-specific AUTO precision and REVIEW recall policies.
Added machine-readable threshold recommendations, threshold-curve CSV and a self-contained HTML calibration report.
Added a deterministic six-class, 600-candidate public calibration fixture and screenshot.
Added calibration algorithm, validation and artifact tests.

Added a complete Simplified Chinese README and interview portfolio guide.
Added English/Chinese language navigation to the repository landing page.
Split lightweight audit/report dependencies from optional GPU inference dependencies.
Added an actionable CLI error when run is used without the inference extra.
Added tests for optional inference dependency loading.

Added the model-free audit command for YOLO schema, label, image and split-integrity checks.
Added optional exact-image hashing to detect train/val/test leakage.
Added the doctor command for Python, package, CUDA and GPU diagnostics.
Added atomic manifest.json lifecycle records to every recovery run.
Added an intentionally flawed synthetic dataset generator for a one-minute public demo.
Expanded automated coverage to include audit findings, HTML audit output and manifest lifecycle.
Added path-redaction options for shareable reports and diagnostics.

Renamed the project to YOLO Label Recovery.
Added an installable CLI package with run and report commands.
Added atomic batch checkpoints and --resume support.
Added idempotent candidate/label writes for interrupted-batch recovery.
Added an HTML quality report with class statistics, distributions and sample galleries.
Added runtime initial/stable batch and OOM retry metrics.
Split domain, geometry, configuration, state and reporting into modules.
Added a synthetic public demo and GitHub Actions CI.

Extracted the resource-aware single-class scanning workflow into a clean repository.
Added immutable-source output handling and audit CSVs.
Added class-specific thresholds and separate existing-label/candidate-duplicate IoU controls.
Added relative model configuration examples, tests and interview documentation.
Added opt-in transactional adaptive batch retry after CUDA OOM.