Skip to main content

Reproduce the evidence

The repository separates a small offline verification path from full experiments that require public datasets, model downloads, compute, or configured rewrite providers.

60-second local verification​

git clone https://github.com/taeyun16/stateful-agentic-retrieval.git
cd stateful-agentic-retrieval

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

make test
make smoke

make test runs deterministic unit and contract tests. make smoke converts a tiny checked-in fixture and runs mock retrieval and training into ignored tmp/smoke/ outputs. It does not download BGE-M3 or call an external LLM API.

Build the chronological dataset​

.venv/bin/python src/hermes_trace_to_retrieval.py \
  --dataset lambda/hermes-agent-reasoning-traces \
  --dataset-config kimi \
  --dataset-split train \
  --output-dir data/processed/hermes_original_prefix_v2

Use glm-5.1 as --dataset-config for the second trace configuration. Raw upstream datasets and processed traces remain local.

Run the canonical matrices​

bash scripts/run_prefix_v2_core.sh
bash scripts/prepare_prefix_v2_heldout_matrix.sh
bash scripts/run_prefix_v2_heldout_matrix.sh
bash scripts/run_prefix_v2_url_matrix.sh
bash scripts/run_taubench_prefix_v2_matrix.sh

The runners keep all_splits and train_split candidate pools separate, pin model revisions in artifacts, and preserve prediction keys for paired significance tests.

Rewrite providers​

bash scripts/run_prefix_v2_rewrite_matrix.sh
Trace data can leave the machine

Real-model rewrite experiments can transmit trace text to configured providers. Review the provider base URLs, model identifiers, reasoning settings, credentials, and cache namespace before running this script. The local smoke path never uses those providers.

Build the paper​

make paper-two-column
make paper-two-column-ko
.venv/bin/python scripts/build_arxiv_bundle.py
git diff --exit-code -- paper/arxiv/source

The minimal arXiv bundle uses main_twocolumn.tex as its top-level file. PDFs, caches, processed datasets, and row-level predictions are excluded from Git.

For the complete Korean command ledger and legacy-v1 audit history, see the detailed reproduction guide.