CascadeMind

CascadeMind is the SemEval-2026 Task 4 system paper codebase for Narrative Story Similarity. Given an anchor story and two candidate stories, the system predicts which candidate is more narratively similar to the anchor across abstract theme, course of action, and outcome.

The public release is intentionally conservative: it exposes the cleaned Gemini runner, paper source, historical non-Claude experiment scripts, and reproducibility tooling, while keeping raw shared-task data, generated logs, review packages, and submission bundles out of git until redistribution rights are confirmed.

Paper

• Canonical source: paper/latest-paperfeb12026/semeval2026_final.tex
• Build notes: paper/README.md
• Public code URL for the paper: https://github.com/epoch-learn/CascadeMind

The camera-ready paper should treat the official shared-task submission as the main result:

Result type	Split	Accuracy	Notes
Official submission	Track A test	72.75%	Listed as rank 10 in the task overview table
Archived local post-hoc file	Released Track A test labels	73.0%	Diagnostic only, not a change to the official standing
Paper-era cascade diagnostics	Development subset	81.0%	Useful for routing analysis; denominator differs from full dev baselines

What To Run

Use a fresh, rotated API key. Do not reuse any key that has appeared in chat, logs, git history, or screenshots.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env
# Fill in only the keys you need, then load them into your shell.
set -a
source .env
set +a

Primary entrypoints:

python best.py
python baseline.py
python train_ensemble.py

Camera-ready experiment runner:

python scripts/run_camera_ready_experiments.py \
  --data data/dev_track_a.jsonl \
  --suite balanced \
  --max-examples 5

The scripts read credentials from environment variables directly; they do not auto-load .env.

Data Policy

The SemEval task data and archived submissions are not tracked in this repository. See data/README.md for expected filenames, row counts, and checksums from the previous LFS pointers.

Local generated outputs should live under ignored artifacts/runs/ directories. If you run experiments for the camera-ready paper, keep the generated manifest with the model IDs, parameters, token/call counts, row counts, and git commit.

Repo Map

Path	Purpose
best.py	Cleaned Gemini bidirectional evaluator
baseline.py	Minimal Gemini structured-output baseline
train_ensemble.py	Multi-signal symbolic ensemble trainer
experiments/	Historical variants and ablations; not all are canonical entrypoints
scripts/run_camera_ready_experiments.py	Manifested Gemini rerun harness
scripts/check_release.py	Public-release and camera-ready sanity checks
paper/	System paper source and style files

Security

This repository previously contained committed API credentials. Those literals have been removed from the working tree, but any exposed credential should be considered compromised and rotated. The Gemini key pasted during planning is also exposed and should not be used for experiments.

Before publishing, run:

python scripts/check_release.py

This public repository is intended to have clean reachable history. Historical secret-bearing branches should remain unreachable from the public remote.

Citation

@inproceedings{kawada-holyoak-2026-cascademind,
  title     = {CascadeMind at SemEval-2026 Task 4: A Hybrid Neuro-Symbolic Cascade for Narrative Similarity},
  author    = {Kawada, Sebastien and Holyoak, Dylan},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026)},
  year      = {2026},
  address   = {San Diego, CA, USA}
}