SLAPBench

Imported from bibeshpyakurel/SLAPBench (AGENTS.md). Install upstream with npx skills add bibeshpyakurel/SLAPBench. Copyright stays with the author.

Repository instructions for agents

Project context

• SLAPBench evaluates full four-finger images with prompted VLMs and vision embeddings across SD302b, Precise and RidgeBase. See REPO_MAP.md and STATUS.md.
• The supervisor's Fingerprint_Foundation_Model_Verification_Student_Guide.md is unavailable (confirmed by the user). Do not invent its requirements or declare milestones complete. Papers and the onboarding checklist are background, not substitutes for the missing specification.
• The approved scope is documentation, offline validation and summary tooling. New experiments, paid calls, fine-tuning and rerunning existing experiments require a separate task authorization. Preserve original scores and raw responses; derived reports belong at new dated paths. Append dated findings to STATUS.md.
• Never send subject identifiers, original filenames or ground-truth labels to hosted models. Preserve the branch and push requirements below.

Git workflow (required)

• main is GitHub's default and publication branch. dev is the working branch for agents and all routine changes, now and in future sessions.
• Keep dev current with git fetch --prune origin and git pull --ff-only origin dev.
• Work on dev and push completed, reviewed changes to origin branch dev.
• Before edits, inspect git status, the active branch, and remotes. Preserve unrelated user changes. Fetch before synchronizing; never discard local work.
• Do not commit or push directly to main. Bring work from dev to main only when the user explicitly asks to publish it, through a pull request after checking the diff, conflicts and CI results.
• Never force-push, delete remote branches, publish tags or releases, or change repository visibility without explicit user instructions.
• Remote: https://github.com/bibeshpyakurel/SLAPBench.git
• Enable git config core.hooksPath .githooks in each clone. Use explicit pushes: git push origin dev:dev. Do not bypass hooks.
• Stage explicit paths, review the staged diff and sizes, run relevant checks, commit with a Conventional Commit subject, then push. Do not use blind git add ..

Required completion step for every change

• The user has given standing authorization to commit and push completed repository changes to origin/dev. Do this for every change task, including documentation and agent-instruction updates, without asking for push permission again, unless the user explicitly asks to keep that task local or uncommitted.
• Before declaring a change task complete, run relevant checks, review and commit the task's publishable changes, and run git push origin dev:dev. A local commit alone does not complete the task. Continue to respect the exclusions below and preserve unrelated user work; this does not authorize publishing ignored files.
• Verify publication with git rev-parse HEAD and git ls-remote --heads origin dev. The remote SHA must match the completed local commit. If the remote advances concurrently, fetch and reconcile without force pushing or discarding work, then verify again.
• If authentication, connectivity, or a repository rule prevents pushing, keep the local commit, report the exact blocker and unpushed commit, and state clearly that publication is incomplete. Never claim a push succeeded without verification.

Publication boundaries

• Never commit datasets/, dataset/, models/, model weights, downloaded archives, virtual environments, caches, credentials (.env and variants), private keys, logs, or interrupted-run backups. Do not force-add ignored files.
• Keep code, dependencies, prompts, textual scores, pair manifests, metrics, documentation, paper sources and existing paper figures for reproducibility.
• Review files over 25 MiB; do not commit files over 50 MiB. Check for credentials without printing their values. Gitignore does not remove already tracked content.
• Dataset images can also occur outside dataset directories. Do not add new raw images or fingerprint crops without explicit publication scope and checking redistribution rights. Existing manuscript illustrations are retained.

Research integrity and areas to preserve

• Do not modify datasets, model files, existing result CSVs/JSONs/TXT exports, pair manifests, paper claims or figures unless the task specifically requires it.
• Never fabricate scores, silently repair results, regenerate fixed pairs, or overwrite original runs. Keep provenance, seeds and sample populations intact.
• Current experiment code is in code/ (paper-only analyses in code/paper/); results are grouped under results/sd302b/, results/ridgebase/ and results/precise/. The earlier flat public release is preserved at tag v0.1.0; do not recreate top-level copies of it.
• Paper sources live in paper/manuscript/ and paper/eccv2026_submission/; paper/eccv2026_template/ is third-party. Do not edit third-party classes, styles, or templates for routine code changes.
• Keep the root minimal: new scripts go in code/, notes in docs/notes/, derived analyses in dated reports/ paths.
• Do not run GPU jobs, dataset preparation that moves files, recovery scripts, model downloads, or paid API calls unless required by the user's task.

Checks and reporting

• Read docs/CONTRIBUTING.md and relevant scripts before changing behavior.
• Run git diff --check; parse changed Python files for syntax errors.
• For analysis changes run python code/verify_metrics.py with dependencies from requirements-verify.txt, plus relevant task-specific validation. This verifier checks legacy published metrics, not all newer experiments.
• CI also runs ruff check code/ --select F,E9,B and regenerates the results table.
• Report what changed, checks and any limitations, the commit, and destination branch.