Armiger 3.2.1

• Fix Torznab XML filtering when upstream indexers such as Prowlarr send numeric file counts in the files attribute (spencernorton/Armiger-Connectors#3, v1.1.1).
• Update prebuilt ghcr.io/spencernorton/armiger:3.2.1 images and armiger-bitmagnet-stack:3.2.1 / armiger-rats-search-stack:3.2.1 Compose packages.

Armiger & Armiger Ultra v2.1.0

Changelog

All notable changes to Armiger and Armiger Ultra are documented in this file.

[2.1.0] - 2026-10-18

Added

• Armiger v2.1 Base Model (armiger.int8.onnx, 8.3 MB):
  • 6.1M-parameter multi-modal Set-Transformer + Text-BiGRU + File-Tree Hierarchical Attention + 36-feature Structural MLP architecture with 8 simultaneous classification heads (category, adult, junk, anime, english_subs, lang, owner_scope, title_language).
  • Sub-millisecond CPU inference (~0.8 ms per item) via dynamic INT8 ONNX Runtime quantization.
• Armiger Ultra v2.1 (armiger_ultra.fp32.onnx, armiger_ultra.pt, 31 MB):
  • Full FP32 ONNX and PyTorch weights integrated with ArmigerUltraPipeline (armiger_ultra_pipeline.py) exposing 11 unified output heads (8 neural heads + swarm_viability, honeypot_trap, and csam_safety).
• 6 Official Production Plugins:
  1. Swarm Survival & Honeypot Defense (swarm_survival.py, swarm_adapter.pt): Orthogonal Weibull Accelerated Failure Time (AFT) survival curve, Beta-Binomial Empirical Bayes release-group reliability, and 5-feature malware/fake-seeder honeypot guard with guaranteed 0.00% category flip.
  2. Tracker Scrape & Bayesian Outcome Priors (swarm_priors.py): Four causally separated pathways (Way 1 active sampling weights, Way 2 14-d public tracker & release-group priors with modality dropout, Way 3 10-d MawaStealer/Vidar DLL-sideloading & .zip/.lnk trap detector, and Way 4 stage-2 grab viability ranking).
  3. Zero-Cue Catalog & Structural Disambiguator (zero_cue_plugin.py, validated_zero_cue_catalog.json): Keyed 64-bit BLAKE2b (person=b"armiger-zc-v2") zero-knowledge lookup catalog covering 929,400 entries (472,017 title+year, 257,016 anime, 200,367 unambiguous title-only) with zero plaintext titles redistributed.
  4. Cross-Head Audio Language & Anime Dub Guard (audio_language_guard.py): Joint title_language x lang x anime disambiguation resolving lowercase Romance .en. prepositions, <Foreign>.ENG.Sub false positives, regional dubs, and [Dual-Audio] releases.
  5. Mondrian Conformal Safety Gates & DHT RSI Engine (calibration_and_dht.py, armiger_dht_rsi.py, calibration.json): Temperature-calibrated (T = 0.72–0.92) Mondrian conformal safety gates (0.00% adult leakage, 0.21% junk false-positive rate) and zero-PII Ed25519-signed recursive self-improvement feedback ingestion.
  6. Modular CSAM-Protected Encrypted BEP-44/BEP-46 DHT DB Sync (armiger_dht_db_sync.py, csam_guard.py): Multi-stage Unicode/homoglyph/leetspeak pre-ingest CSAM & malware blocker, Ed25519-signed blinded SHA-256 tombstones, hybrid X25519 + ML-KEM-768 (FIPS 203) key encapsulation with ChaCha20-Poly1305 AEAD shard encryption, and hybrid Ed25519 + WOTS+-SHAKE256 K-of-N distinct-ASN quorum verification.
• Verification Suite:
  • 91 automated tests (test_armiger_ultra_pipeline.py, test_armiger_dht_rsi.py, test_dht_db_sync.py) using Creative Commons and public-domain fixtures.