Reproduction pack v1.2 — focused docs and separate metadata

The core repository's docs/ directory now contains only DATASET.md and EVALUATION.md. Limitations are included in the dataset card, licensing scope is in README/NOTICE, and detailed frozen-result interpretation is included in the standalone reproduction pack.

This Release provides three attachments:

Asset	Purpose	SHA-256
impossiblerubrics-reproduction-v1.2.zip	Standalone historical measurements, scorer, census and result documentation	8dacb30b83561ef69250f9b1f092b622da68612ef6a2da705aa3b193001c719b
source_manifest.json	Frozen evidence provenance inventory, unchanged from v1.1	dda6b11136c7942202fd3c3088a2ca23c26dd8e0008e867d061d40cbd3300247
release_manifest.json	Core/package file hashes, data projections and source-inventory checksum	d0b0bafdf471d633ddb0471cd0246348e1efe86ab6468be0a7557a02c033f062

The manifests are optional downloads for provenance and verification; benchmark commands do not require them. The core snapshot retains all 217 source environments, schemas, splits and runtime tools unchanged. Numeric measurements and scientific caveats are preserved.

Extract the reproduction ZIP and run from its impossiblerubrics-reproduction/ directory with Python 3.10+:

python analysis/census.py
python evaluate.py --results results/census/opus5_hardset_runs.json --split splits/hard45.json

Expected census: 21 arms, 2,100 cells, 522 violations, 514 exploits, 8 low-scoring violations and 5 violating ties. The single arm gives 16/45 exploits. No dependency installation or model API calls are required.

Validation includes 44 preparation tests, 28 clean-core tests, independent pack execution, exact file hashes and a simulated refresh installation failure that restores the previous core/metadata pair. Earlier Releases remain intact. Licensing scope and repository visibility are unchanged.

Reproduction pack v1.1 — author-owned data under Apache-2.0

This version records the maintainer's choice of Apache-2.0 for project-owned benchmark data, annotations, synthetic material and experimental results, alongside code and original documentation. Third-party material retains its original rights and terms and is excluded from that grant. See the bundled docs/DATA_LICENSE.md.

The measurements, scientific records, scorer and census script are unchanged from v1. Only license documentation and download references changed in the archive.

Extract impossiblerubrics-reproduction-v1.1.zip and run from its impossiblerubrics-reproduction/ directory with Python 3.10+:

python analysis/census.py
python evaluate.py --results results/census/opus5_hardset_runs.json --split splits/hard45.json

The census reproduces 21 arms, 2,100 cells, 522 violations, 514 exploits, 8 low-scoring violations and 5 violating ties. The exact Opus 5 Hard-45 arm gives 16/45 exploits. No dependency installation or model API calls are needed.

SHA-256: 6fbe0ced6b68402cb039f83485484d3e449bcd03c6cfd26b7b9d9025a79bcdd6

The original v1 archive remains available unchanged. This release inherits the repository's private visibility; third-party source-rights review is still pending.

Optional reproduction pack (v1)

The core benchmark now keeps one source record per environment and generates combined JSON/JSONL locally. Historical measurements are provided as a separate, standalone download so the main checkout stays focused on using the benchmark.

Download and extract impossiblerubrics-reproduction-v1.zip, then run from its impossiblerubrics-reproduction/ directory:

python analysis/census.py
python evaluate.py --results results/census/opus5_hardset_runs.json --split splits/hard45.json

Python 3.10+ is sufficient. No dependencies, API keys or copy of the core data are required.

The archive preserves the 18 frozen numeric result files, difficulty-selection metadata and census script, with split/scorer dependencies. It reproduces 21 arms, 2,100 cells, 522 violations, 514 exploits, 8 low-scoring violations and 5 violating ties. The exact Opus 5 Hard-45 arm gives 16/45 exploits. These commands recompute stored measurements; they do not rerun model inference or reproduce every paper table.

SHA-256: 3a216a685b88bc35ccd789db7fed428f81be4f9c63e16e8b4fc14c8c8812c53c

The core docs/release_manifest.json also records the archive hash and every packaged file hash. Code and original documentation use Apache-2.0; data/recorded-results licensing remains separate, as described in docs/DATA_LICENSE.md. This release inherits the repository's private visibility.