GitHub - AutomationOptimization/tsce_demo: This repo is for the demonstration of TSCE principles.

2 min read Original article ↗

TSCE Paper Companion Repository

This repository is a self-contained companion artifact for the TSCE paper. It includes the paper source and PDF, the paper-facing test sets, curated result artifacts that back the paper tables and figures, and the installable TSCE wrapper package. It intentionally excludes caches, virtual environments, generated bytecode, dashboards, model checkpoints, broad exploratory runs, and unrelated training outputs.

Layout

  • paper/: final think_before_you_speak_v2 PDF/TEX, mechanism addendum, diffusion framework, support notes, referenced figures, and artifact-local figure-generation code.
  • test_sets/: HDA task-suite eval/smoke/baseline-failure sets, GSM/GSM-Hard references, synthetic format-control eval/smoke sets, manifests, and HDA candidate files used by the format sweep.
  • results/: curated paper-supporting result folders plus docs/tsce_results_inventory.csv and docs/tsce_results_inventory_summary.json as full-corpus provenance.
  • package/: minimal installable TSCE wrapper package, examples, and wrapper tests.
  • scripts/: artifact-local validation and figure-regeneration entry points.
  • MANIFEST.md: source-to-destination provenance for every copied or generated artifact except the checksum file.
  • checksums.sha256: SHA-256 hashes for the exported files.
  • SECURITY_REVIEW.md: credential/key scan summary for publication.

Quick Validation

From the repository root:

python scripts/validate_artifact.py
shasum -a 256 -c checksums.sha256

The validation script checks for banned artifacts, required paper/package/result files, checksum integrity, paper figure references, artifact-local executable paths, absence of local absolute paths, strict credential signatures, and package importability.

Wrapper Install

cd package
python -m pip install -e .
python -c "from tsce import TSCEChat, TSCEClient, TSCEReply"

Wrapper tests from package/:

python -m pytest tests/test_tsce_wrapper.py tests/test_tsce_chat_smoke.py

If pytest, numpy, or matplotlib are missing, install optional local reproduction dependencies from the repo root:

python -m pip install -r requirements-reproduce.txt

Paper Build

The checked-in PDF is paper/think_before_you_speak_v2.pdf. The TEX source uses figure paths relative to paper/, so a LaTeX engine can build it from that directory. For example, with Tectonic installed:

cd paper
tectonic --outdir /tmp/tsce-paper-build think_before_you_speak_v2.tex

Regenerate Figures

The figure-generation script has been relocated to use this repository layout only. It reads the curated results under results/ and writes to paper/figures/tsce/:

python scripts/build_figures.py

Result Mapping

  • Fixed-policy Gemma neutral validation: results/tsce_agent_demo/results/gemma_fixed_latent_bias_compact_heldout_n100_seed260426/.
  • Format-robustness sweep on 69 baseline failures: results/tsce_agent_demo/results/hda_anchor_format_full_sweep_seed260501_u69_k16_combined/.
  • Controlled greenhouse proxy-family and token-causality run: results/tsce_agent_demo/results/story_subject_greenhouse_v1/.
  • Naturalistic forgiveness-theme token-causality run: results/tsce_agent_demo/results/story_theme_forgiveness_v1/.
  • Activation-space diagnostic used by figure 6: results/tsce_agent_demo/results/gemma_latent_zone_v1_20260424_progress2/.
  • Cross-run black-box figure inputs: curated archive/2025-05-30_*, results-100-*, results-300-*, tsce_diffusion_2025-12-03_*, tsce_eval, and tsce_eval_deterministic folders under results/.

Integrity

checksums.sha256 covers every tracked artifact file except itself. Re-run the checksum command after any intentional edit, then regenerate the checksum file before publishing.