Evaluation Wording That Moves Safety Numbers

Detect evaluation-register cues that shift safety scores, then rewrite scenarios under construct checks.

---

Overview

This repository implements experimental profiles for Evaluation Wording That Moves Safety Numbers. Config, caching, hooks, metrics, ablations, reporting, and CI are built for reproducible local pilots on small open-weight models.

Hypothesis (one line): Evaluation-register wording can move measured safety scores without changing the underlying task; internal probes plus guarded rewrites can reduce that artifact.

Motivation

Interpretability and safety claims fail in practice for boring engineering reasons: unpinned weights, chat templates skipped, invalid layer indices, intervals that span zero treated as nulls, and stages that raise NotImplementedError. This repo treats those as first-class bugs.

Architecture / Pipeline

flowchart LR
  cfg[Hydra config] --> seed[set_seed]
  seed --> data[build dataset]
  data --> model[load pinned model]
  model --> stages[experiment stages]
  stages --> cache[artifact cache]
  stages --> eval[evaluation harness]
  eval --> agg[aggregate]
  agg --> tables[MD + LaTeX tables]
  agg --> figs[PDF/SVG/PNG figures]

Loading

Stage	Module	Output
Compose config	configs/ + evalaware.configs	resolved config.yaml
Build data	evalaware.data	splits + manifest
Load model	evalaware.models	LoadedModel + resolved commit
Run stages	scripts/run_experiment.py	per-stage JSON
Aggregate	evalaware.reporting	results.json + tables + figures

Results

Experiment	Metric	Value	Provenance
smoke	config compose	pass	unit / CI
pilot	harness recovery	pending	labelled synthetic until measured

Provenance. No measured number in this table comes from a full model run on private data. Synthetic harness-validation outputs are labelled is_synthetic: true and must not be reported as empirical results.

Repository guide

.
├── configs/           # Hydra groups + experiment presets
├── src/evalaware/       # installable library (print-free)
├── scripts/           # CLIs with argparse / hydra
├── tests/             # ≥30 modules; tiny random GPT-2 only
├── data/              # manifests only
├── docs/              # DESIGN.md, HARDWARE.md
├── TASK.md            # research plan + DAG
└── Makefile           # install, lint, test, ci, pilot, doctor

Command	Purpose
make install-dev	editable install + pinned requirements
make test	full unit suite
make ci	lint + test + typecheck + api-contract + coverage
make pilot	end-to-end pilot profile
make doctor	environment / device report

Status

Focus: cue-sensitive evaluation artifacts in safety scoring. Shared infrastructure is in place; domain stages must pass harness validation before any measured claim.

Related work

• Complexity bar: Critical Data PRIMED-AI / RecursiveJEPA engineering standard

Citation

@misc{silent_test_recognition,
  title        = {Evaluation Wording That Moves Safety Numbers},
  author       = {Alana Sung},
  year         = {2026},
  howpublished = {Technical report},
}

License

MIT. Model weights and third-party datasets retain their upstream licenses.

---

Built for reproducible interpretability pilots on Apple Silicon and CI CPUs.

Design constraints (short)

1. Library code has zero print, zero argparse, zero __main__.
2. Every result JSON carries task, seed, git_sha, n.
3. Model revisions are pinned; load path records the resolved commit.
4. Chat templates are applied when available; the path is recorded.
5. MPS sets PYTORCH_ENABLE_MPS_FALLBACK and records the flag.
6. CI spanning zero is inconclusive; report MDE and run TOST before null claims.
7. Pilot n_items >= 512.
8. Layer indices are validated against n_layers.

Hardware note

Torch model forward passes may use MPS on Apple Silicon. Sklearn, numpy, pandas, and matplotlib figure generation run on CPU (see docs/HARDWARE.md).

Config composition

python scripts/run_config_smoke_test.py experiment=pilot model=gpt2 seed=7
python scripts/run_experiment.py experiment=baseline eval.layers=[2,4,6]

Ablations

Ablations live under src/evalaware/ablation/ and return structured dicts. Presets mirror them under configs/experiment/ablation_*.yaml.

Reporting

python scripts/aggregate_results.py
python scripts/make_tables.py
python scripts/make_figures.py

One aggregation command regenerates Markdown and booktabs LaTeX from raw JSONs.

Contributing

See CONTRIBUTING.md. Open work goes in TASK.md / GitHub issues — never as TODO comments in library code.