TheoLeeCJ/SemIf-OpenJev

Review signals: 1 connection review, 30 hardcoded secret, 6 known vulnerability. 3 static matches retained as context.

Reviewed as expected for a test run: resolver address lookups that send no data.

Repository processes attempted these non-loopback connections; proxies can hide the final hostname. 47 distinct locations. 11 attributed processes.

Gitleaks found a possible hardcoded secret; value withheld. 7 distinct locations. 18 findings.

OSV reports a known vulnerability in this dependency version. 3 distinct locations. 6 findings.

2 encoded blob matches retained as context; full list in the evidence.

1 encoded execution matches retained as context; full list in the evidence.

OSV scanned supported lockfiles; unpinned dependencies and unsupported formats are outside this check.

A short run cannot establish that code is safe.

Connections show IP addresses; proxies can hide the final hostname.

Tracing covers dispatched commands and their children, not the whole OS; guest root can interfere with tracing.

Full safety evidence →

• Each result contains typed option scores, timing, the exact model revision, and a prompt hash.: Real input produced the expected result: "prompt_sha256": "7ac35785358f0c656eeb741ca0a51f03e0133753a71e2879a18ad0ab56d0c024"

  Source: Each result contains typed option scores, timing, the exact model revision, and a prompt hash.

  Input: examples/decisions.jsonl scored with the llama.cpp CPU backend

  Expected: each row has a prompt hash and the pinned model revision

  Command: <guest-path> --mode direct --backend llamacpp --gguf <guest-path> --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a --input examples/decisions.jsonl --output <guest-path>

  Observed: "prompt_sha256": "7ac35785358f0c656eeb741ca0a51f03e0133753a71e2879a18ad0ab56d0c024", "prompt_version": "direct-options-v1", "model": {"source": "Qwen/Qwen3.5-4B", "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a", "backend": "llamacpp", "dtype": "gguf-quantized", "gguf": {"file": "Qwen_Qwen3.5-4B-Q4_K_M.gguf", "bytes": 3013027808, "sha256": "13c16f426047e2de38cd075bdade4a7bcbc8c774384876f677740cda65f8a983"}, "vocab_size": 248320, "threads": 2, "n_gpu_layers": 0, "max_prompt_tokens": 4096, "context_tokens": 4352, "decode_chunk": 512, "llama_cpp_python_version": "0.3.35", "transformers_vers

• CPU only: the llama.cpp backend scores the same prompts from a local GGUF checkpoint with no CUDA device.: Real input produced the expected result: "backend": "llamacpp"

  Source: **CPU only:** the llama.cpp backend scores the same prompts from a local GGUF

  Input: examples/decisions.jsonl on a guest with no CUDA device, using the local Q4_K_M GGUF

  Expected: output rows record backend llamacpp

  Command: <guest-path> --mode direct --backend llamacpp --gguf <guest-path> --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a --input examples/decisions.jsonl --output <guest-path>

  Observed: "backend": "llamacpp", "dtype": "gguf-quantized", "gguf": {"file": "Qwen_Qwen3.5-4B-Q4_K_M.gguf", "bytes": 3013027808, "sha256": "13c16f426047e2de38cd075bdade4a7bcbc8c774384876f677740cda65f8a983"}, "vocab_size": 248320, "threads": 2, "n_gpu_layers": 0, "max_prompt_tokens": 4096, "context_tokens": 4352, "decode_chunk": 512, "llama_cpp_python_version": "0.3.35", "transformers_version": "5.17.0", "serving_config": "llamacpp-direct-v1"}, "readout": "quantized last-position logits restricted to declared answer slots; no generated tokens", "probability_status": "conditional option score over quantiz

Examples

• Score the owned decision examples on CPU with the llama.cpp GGUF backend: command completed successfully

  Source: README.md Quick start, CPU only llama.cpp path

  Input: examples/decisions.jsonl and the local Q4_K_M GGUF

  Expected: three scored rows with a prompt hash

  Command: <guest-path> --mode direct --backend llamacpp --gguf <guest-path> --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a --input examples/decisions.jsonl --output <guest-path>

  Observed: 56": "13c16f426047e2de38cd075bdade4a7bcbc8c774384876f677740cda65f8a983"}, "vocab_size": 248320, "threads": 2, "n_gpu_layers": 0, "max_prompt_tokens": 4096, "context_tokens": 4352, "decode_chunk": 512, "llama_cpp_python_version": "0.3.35", "transformers_version": "5.17.0", "serving_config": "llamacpp-direct-v1"}, "readout": "quantized last-position logits restricted to declared answer slots; no generated tokens", "probability_status": "conditional option score over quantized weights; uncalibrated as decision confidence", "forward_seconds": 4.295587840999815, "total_seconds": 4.298743025000022}

• Offline test suite: command completed successfully

  Source: docs/REPRODUCE.md, pytest -q

  Input: repository tests

  Expected: all offline tests pass

  Command: <guest-path> -q

  Observed: .....................................................s..........s.... [100%] 67 passed, 3 skipped in 5.84s

Local or hosted: both

Score README decision examples (examples/decisions.jsonl) with the semif-score CLI, using the documented CPU-only llama.cpp backend and the pinned Qwen3.5-4B revision

cd <guest-path> && pwd && ls -la && id -u && cat /etc/os-release | head -3 && python3 --version && (git status 2>&1 | head -3)
cd <guest-path> && cat README.md 2>/dev/null | head -300
cd <guest-path> && cat AGENTS.md; cat pyproject.toml; cat requirements.txt; ls .coldpass examples docs tests src; nvidia-smi 2>&1 | head -3; nproc; free -g | head -2; df -h /tmp | tail -1; env | grep -i -E 'proxy|hf_|cuda' ; sudo -n true && echo SUDO_OK
cd <guest-path> && cat docs/REPRODUCE.md | head -150; echo ----; cat examples/* | head -40; echo ----; ls -la /inputs 2>&1; ls <guest-path> 2>&1
cd /tmp; for u in https://pypi.org/simple/torch/ https://huggingface.co/api/models/Qwen/Qwen3.5-4B https://huggingface.co/bartowski/Qwen_Qwen3.5-4B-GGUF https://github.com; do echo -n "$u -> "; curl -s -o /dev/null -m 20 -w "%{http_code}\n" "$u" || echo fail; done; python3 -m venv --help >/dev/null 2>&1 && echo VENV_OK || echo NO_VENV; pip --version; ls /usr/bin/python3*
sed -n 1,200p <guest-path> | head -120
cd <guest-path> && cat .gitignore && python3 -m venv .venv && . .venv/bin/activate && python --version && pip install -q --upgrade pip 2>&1 | tail -3; pip install -e '.[test]' > <guest-path> 2>&1; echo EXIT=$?; tail -15 <guest-path>
cd <guest-path> && rm -rf .venv && (sudo apt-get update -qq && sudo DEBIAN_FRONTEND=noninteractive apt-get install -y -qq python3.12-venv python3-pip) > <guest-path> 2>&1; echo APT=$?; tail -3 <guest-path>; python3 -m venv .venv && . .venv/bin/activate && pip --version && pip install -q --upgrade pip 2>&1 | tail -2; echo PIPUP=$?
cd <guest-path> && . .venv/bin/activate && timeout 3000 pip install -e '.[test]' > <guest-path> 2>&1; echo EXIT=$?; tail -8 <guest-path>
cd <guest-path> && . .venv/bin/activate && python -c "import torch;print(torch.__version__, torch.cuda.is_available())"; ls src/semif_phase1; grep -n -E "add_argument|device|cuda" src/semif_phase1/cli.py | head -60; cat manifests/models.json 2>/dev/null | head -30
cd <guest-path> && . .venv/bin/activate && timeout 900 pytest -q 2>&1 | tail -15; echo EXIT=${PIPESTATUS[0]}
cd <guest-path> && sed -n 34,110p src/semif_phase1/cli.py; grep -n -E "def load_causal_model|cuda|device" src/semif_phase1/core.py | head -30
cd <guest-path> && . .venv/bin/activate && mkdir -p <guest-path> && CUDA_VISIBLE_DEVICES=0 timeout 600 semif-score --mode direct --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a --input examples/decisions.jsonl --output <guest-path> 2>&1 | tail -8; echo EXIT=${PIPESTATUS[0]}; ls <guest-path>; sed -n 99,110p src/semif_phase1/core.py
cd <guest-path> && sed -n 1,60p src/semif_phase1/llamacpp_backend.py; grep -n -i "gguf" manifests/models.json docs/*.md | head -20; which cmake g++ gcc make; cat /proc/cpuinfo | grep -m1 flags | tr ' ' '\n' | grep -E "avx|fma" | tr '\n' ' '
cd <guest-path> && (sudo DEBIAN_FRONTEND=noninteractive apt-get install -y -qq cmake build-essential > <guest-path> 2>&1; echo APT=$? >> <guest-path>) ; tail -1 <guest-path>; cmake --version | head -1; . .venv/bin/activate && nohup sh -c "pip install -e '.[test,llamacpp]' > <guest-path> 2>&1; echo PIPEXIT=\$? >> <guest-path>" >/dev/null 2>&1 &
mkdir -p <guest-path> && cd <guest-path> && nohup sh -c ". <guest-path>; export HF_HOME=<guest-path>; hf download bartowski/Qwen_Qwen3.5-4B-GGUF Qwen_Qwen3.5-4B-Q4_K_M.gguf --local-dir <guest-path> > <guest-path> 2>&1; echo DLEXIT=\$? >> <guest-path>" >/dev/null 2>&1 &
sleep 5; echo started; cat <guest-path> 2>/dev/null | tail -2
sleep 240; tail -2 <guest-path>; echo ---; tail -c 400 <guest-path>; ls -la <guest-path>; du -sh <guest-path> 2>/dev/null
for i in $(seq 1 28); do if grep -q PIPEXIT <guest-path>; then break; fi; sleep 30; done; tail -4 <guest-path> | cut -c1-600
cd <guest-path> && . .venv/bin/activate && ls <guest-path>; time (timeout 1500 semif-score --mode direct --backend llamacpp --gguf <guest-path> --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a --input examples/decisions.jsonl --output <guest-path> > <guest-path> 2>&1); echo EXIT=$?; tail -15 <guest-path> | cut -c1-500; ls -la <guest-path>
cd <guest-path> && wc -l results-direct.jsonl && python3 -c "
import json
for l in open('results-direct.jsonl'):
    r=json.loads(l)
    print({k:(v if not isinstance(v,(dict,list)) or len(json.dumps(v))<200 else '...') for k,v in r.items()})
" | cut -c1-900
cd <guest-path> && head -1 <guest-path> | python3 -c "import json,sys; r=json.load(sys.stdin); print(r['model'])"; grep -n -i -E "revision" <guest-path> | head -c 300; echo; grep -rn -i -E "schedul|cron|webhook|\-\-once|\-\-now|while True|loop" src/ README.md docs/ 2>/dev/null | head; echo "grep-exit=$?"; grep -n -o '"prompt_sha256": "[0-9a-f]*"' <guest-path> | head -1; echo; (cd results/raw && sha256sum -c SHA256SUMS 2>&1 | tail -2)
cd <guest-path> && grep -F -c "Each result contains typed option scores, timing, the exact model revision, and a prompt hash." README.md; grep -F -c "**CPU only:** the llama.cpp backend scores the same prompts from a local GGUF" README.md; rm -f <guest-path>; mkdir -p .coldpass && cd .coldpass && cat > probe.json <<'EOF'
{"kind":"cli","argv":["<guest-path>","--mode","direct","--backend","llamacpp","--gguf","<guest-path>","--model","Qwen/Qwen3.5-4B","--revision","851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a","--input","examples/decisions.jsonl","--output","<guest-path>"],"cwd":"<guest-path>","expected_file":"<guest-path>","expected_text":"\"prompt_sha256\": \"7ac35785358f0c656eeb741ca0a51f03e0133753a71e2879a18ad0ab56d0c024\""}
EOF
cat > claims.json <<'EOF'
[{"kind":"claim","claim":"Each result contains typed option scores, timing, the exact model revision, and a prompt hash.","source":"Each result contains typed option scores, timing, the exact model revision, and a prompt hash.","input":"examples/decisions.jsonl scored with the llama.cpp CPU backend","expected":"each row has a prompt hash and the pinned model revision","probe_kind":"cli","argv":["<guest-path>","--mode","direct","--backend","llamacpp","--gguf","<guest-path>","--model","Qwen/Qwen3.5-4B","--revision","851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a","--input","examples/decisions.jsonl","--output","<guest-path>"],"cwd":"<guest-path>","expected_file":"<guest-path>","expected_text":"\"prompt_sha256\": \"7ac35785358f0c656eeb741ca0a51f03e0133753a71e2879a18ad0ab56d0c024\""},
{"kind":"claim","claim":"CPU only: the llama.cpp backend scores the same prompts from a local GGUF checkpoint with no CUDA device.","source":"**CPU only:** the llama.cpp backend scores the same prompts from a local GGUF","input":"examples/decisions.jsonl on a guest with no CUDA device, using the local Q4_K_M GGUF","expected":"output rows record backend llamacpp","probe_kind":"cli","argv":["<guest-path>","--mode","direct","--backend","llamacpp","--gguf","<guest-path>","--model","Qwen/Qwen3.5-4B","--revision","851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a","--input","examples/decisions.jsonl","--output","<guest-path>"],"cwd":"<guest-path>","expected_file":"<guest-path>","expected_text":"\"backend\": \"llamacpp\""}]
EOF
cat > examples.json <<'EOF'
[{"kind":"example","claim":"Score the owned decision examples on CPU with the llama.cpp GGUF backend","source":"README.md Quick start, CPU only llama.cpp path","input":"examples/decisions.jsonl and the local Q4_K_M GGUF","expected":"three scored rows with a prompt hash","argv":["<guest-path>","--mode","direct","--backend","llamacpp","--gguf","<guest-path>","--model","Qwen/Qwen3.5-4B","--revision","851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a","--input","examples/decisions.jsonl","--output","<guest-path>"],"cwd":"<guest-path>","env":{},"expected_file":"<guest-path>","expected_text":"\"id\": \"policy-1\""},
{"kind":"example","claim":"Offline test suite","source":"docs/REPRODUCE.md, pytest -q","input":"repository tests","expected":"all offline tests pass","argv":["<guest-path>","-q"],"cwd":"<guest-path>","env":{},"expected_text":"67 passed"}]
EOF
cat > cycles.json <<'EOF'
{"detected":false,"reason":"Entry points inspected: pyproject.toml defines only the one-shot console script semif-score (semif_phase1.cli:main). It scores the input file once and exits. A grep of src/, README.md and docs/ found no cron, scheduler, webhook, --once/--now, or polling loop. The only loop-like matches are GPU looped-mode code in shared.py and the README phrase 'decoding loop', which is a negation. Nothing runs repeatedly, so no product cycle exists to trigger.","checks":[]}
EOF
for f in probe claims examples cycles; do python3 -m json.tool $f.json >/dev/null && echo "$f OK"; done; ls -la <guest-path>
cd <guest-path> && cat > report.json <<'EOF'
{"outcome":"working","fault_owner":null,"operation":"Score README decision examples (examples/decisions.jsonl) with the semif-score CLI, using the documented CPU-only llama.cpp backend and the pinned Qwen3.5-4B revision","operation_scope":"CPU-only path on Ubuntu 24.04, 2 vCPU, 3 GB RAM, no NVIDIA GPU. Uses the Q4_K_M GGUF (3.01 GB) from bartowski/Qwen_Qwen3.5-4B-GGUF. The primary CUDA BF16 path in the README was not exercised because this guest has no GPU.","steps":["sudo apt-get install python3.12-venv python3-pip (python3 -m venv .venv failed with ensurepip missing)","python3 -m venv .venv && pip install --upgrade pip","pip install -e '.[test]'","pytest -q (67 passed, 3 skipped)","CUDA_VISIBLE_DEVICES=0 semif-score --mode direct --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a --input examples/decisions.jsonl --output <new> (failed: MPS is unavailable; no CUDA device)","sudo apt-get install cmake build-essential; pip install -e '.[test,llamacpp]' (llama-cpp-python 0.3.35 built from source)","hf download bartowski/Qwen_Qwen3.5-4B-GGUF Qwen_Qwen3.5-4B-Q4_K_M.gguf (3.01 GB)","semif-score --mode direct --backend llamacpp --gguf <Q4_K_M.gguf> --model Qwen/Qwen3.5-4B --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a --input examples/decisions.jsonl --output <new> (exit 0, 3 rows scored in about 45 s)"],"deviations":["Ubuntu venv lacked ensurepip; installed python3.12-venv via apt (setup, not a product fault).","GPU path from README Quick start failed with 'MPS is unavailable' because the guest has no CUDA device. Substituted the README's documented CPU-only llama.cpp path.","Downloaded the GGUF with the hf CLI from the README-cited repo; the GGUF was not in the workspace."],"stuck_step":null,"main_blocker":null,"plain_failure":null,"suggested_fix":null}
EOF
python3 -m json.tool report.json >/dev/null && echo REPORT_OK; cd <guest-path> && <env-var>=<guest-path> python3 -I <guest-path> --validate --replay 2>&1 | tail -40; echo CHECK_EXIT=${PIPESTATUS[0]}