.gitignore

# Python
__pycache__/
*.py[cod]
*$py.class
*.egg-info/
.pytest_cache/

# Virtual environments
.venv/
venv/
env/

# OS
.DS_Store
Thumbs.db
.Trash-*/

# Editor
*.swp
*.swo
*~
.vscode/
.idea/

# Secrets
.env
.env.local
secrets.yml

# Repo-specific
/logs/
benchmarks/results/qwen2.5-7b-fp16/sanity/

# Backup files (any *.bak in repo)
*.bak

# Download infrastructure scripts (bootstrap-only, not part of research)
scripts/download_models.sh
scripts/download_watchdog.sh
scripts/download_progress.py

# Reference materials (papers, PDFs) — managed by Zotero, not tracked in repo
references/

# === EMBARGO: paper-bound benchmark results ===
# Until peer-reviewed publication, the following are kept
# local-only. Public release scheduled with paper publication.
# See benchmarks/results/README.md for full embargo policy.
benchmarks/results/scaling/*.json
benchmarks/results/scaling/*.csv
benchmarks/results/scaling/*.parquet
benchmarks/results/scaling/plots/
benchmarks/results/scaling/figures/
# Allow .gitkeep so the directory persists in git
!benchmarks/results/scaling/.gitkeep

# Phase 2 raw sweep data — EMBARGO until publication
# Aggregated SUMMARY.md, results_table.csv, and *.png plots are committed;
# raw per-run files are kept local-only.
benchmarks/results/*/thermal-runs/

# Run-3 orchestrator logs capture run_concurrent stdout (throughput numbers) —
# EMBARGO_paper_bound (§11.3), local-only until paper acceptance.
benchmarks/results/_run3_orchestrator_logs/

# Raw terminal logs from interactive sessions — local-only by default
logbook/raw_logs/

# Local-only session continuation notes (private)
SESSION_RESUME.md

# WIP manuscripts + pre-consent collaborator outreach — local-only, NEVER public.
# Paper drafts and outreach files name proposed co-authors / reviewers (with
# ORCIDs, affiliations, authorship-order) before their consent; they must not
# appear in the public repository. Manuscript work lives in the private vault.
paper/
eval-rag/

# Raw vLLM sanity server logs — engine-side throughput (Polish models §11.3-
# embargoed) lives in these; raw logs are local-only, while the curated .json
# envelope (§11.1: load / VRAM / KV / max_concurrency) stays public.
environment/sanity-tests/*.log

# Claude Code local-only files
.claude/settings.local.json
.claude/local/
.claude/cache/
.claude/worktrees/
.claude/agent-memory/

# Embargoed benchmark results — DO NOT COMMIT until paper acceptance
benchmarks/results/

# Embargoed planning doc — contains throughput refs
docs/sessions/PLAN-NEXT.md

# AWQ calibration corpus — corpus.jsonl + catalogue + extraction script ARE
# committed (see calibration/LICENSE — source-specific terms, NOT root CC-BY-4.0).
# Raw third-party SmPC PDFs and the text-only backup stay local-only:
calibration/clinical-pl/chpl_raw/
calibration/clinical-pl/corpus.text-only.bak.jsonl
# clinical-EN mirror: raw English EMA PI PDFs + build venv stay local-only
# (corpus.jsonl, scripts, manifest.json, fetch_log.json ARE committed):
calibration/clinical-en/epar_raw/
calibration/clinical-en/.venv/
# fetch_log.json = build artifact (per-PDF URL+bytes+sha256); SHA256 hashes trip
# detect-secrets (false positive). Provenance preserved via corpus.jsonl source_url
# + RUNBOOK 61-URL table; per-PDF integrity stays local-only.
calibration/clinical-en/fetch_log.json

# Source literature (PDFs of papers the project builds on) — local-only
literature/

# Claude.ai conversation history backup (local-only)
history/

# Obsidian vault artifacts — navimed-umb is also a Vaults-main subfolder.
# Folder notes + Claude import metadata live on disk for Obsidian navigation
# but must NOT pollute the paper/code git repo.
_index.md
*_index.md
_claude_ai_metadata.md
.obsidian/

# local-only dashboard preview (never deploy)
site/_preview_*
site/_preview_build.py

# Local-only full results (embargoed §11.2/§11.3 — never publish)
site/_local-*.html