ATTACK-COST AUDITS (separate from protocol timings)

All supplied raw audit results concern n=6144,k=1965,w=1600. They are attack-model estimates, not independent certification of the modified polar family.

1. audit_isd.py uses CryptographicEstimators 2.1.1 SDEstimator (Prange, Stern, Dumer, May-Ozerov, BJMM), bit_complexities=1, memory_access=0, workfactor_accuracy=1. The target weight is EXACT w; expected random-code solution multiplicity is the library default. This is fixed-syndrome, not a DOOM, structural, multi-key or quantum analysis. The installed wheel hash and environment are in provenance/estimator_manifest.json.
   Official source: https://github.com/Crypto-TII/CryptographicEstimators

2. audit_doom.py calls the authors' dsDOOM at log2(M)=0,32,64,96,112,128. Fetch the pinned author source with fetch_doom.py first (network access is explicit). URL, commit and SHA256 are fixed in that script and provenance/doom_source_commit.json. No external doom.py is redistributed here. The official function includes expected exact-weight multiple solutions but excludes obtaining/hash-generating M target syndromes; its raw output must not be presented as a complete free-message forgery cost. Very large unbounded memory and random-code assumptions apply.
   Official source: https://github.com/Gabsadio/Multi-Instance-KEM
   May and Sa Diogo, ePrint 2026/517, DOI 10.1007/978-3-032-36264-3_24.

3. sendrier2011_crosscheck.py independently evaluates the collision-DOOM model in ePrint 2011/367 Eq.(2),(3),(5),(8) and Section5.1. It reproduces four published examples approximately and explicitly charges target hashing and target transformation. Its weight-at-most envelope/reserve is confined to this collision model and searched ranges; it does not certify all modern attacks or structural security.
   Primary source: https://eprint.iacr.org/2011/367

To reproduce (optional, potentially expensive; these tools are NOT run by the integrated benchmark):
python -m pip install -r environment/requirements-attacks.txt
python attacks/audit_isd.py
python attacks/fetch_doom.py
python attacks/audit_doom.py
python attacks/sendrier2011_crosscheck.py
python attacks/tail_factor_bound.py

Scripts differ from the historical run only in import/output paths and command-line output arguments. Exact historical results remain byte-for-byte intact. Original/packaged hashes and diffs are recorded in ../provenance/. Elapsed optimizer timings in those results are historical diagnostics, not protocol benchmarks.
