NorviTech/armiger-ultra

Armiger Ultra v3.0 (v3.0-gold-r5) — Full Multi-Task Neural & Plugin Suite

Version: v3.0-gold-r5 · Weights: 32.7 MB FP32 ONNX / PyTorch + 8.3 MB INT8 ONNX · Parameters: 8,295,882 · License: Apache-2.0

Armiger Ultra v3 is a 9-head byte-level RoPE Transformer (d=256, 6 layers, 8 heads) with a 6-plugin pipeline for raw BitTorrent/DHT metadata (name, file-tree paths + sizes, optional tracker/DHT counters). Batched GPU latency is 0.21 ms/item (0.29 ms on M4 Pro MPS). All 6 plugins ship in this repo via ArmigerUltraPipeline.

Architecture & Bundled Capabilities

• 9 simultaneous heads (armiger_ultra.fp32.onnx & armiger.int8.onnx): category (11-class), adult, junk/malware, anime, english_subs, spoken lang (3-class + 8-class audio guard), title_language (english_title vs non_english_title), owner_scope policy gate, and per-byte span_bio (O, B-TITLE, I-TITLE, B-YEAR, I-YEAR, B-EP, I-EP) via a constrained 7-tag Viterbi mask (0 illegal transitions).
• Bundled 6-plugin suite (ArmigerUltraPipeline):
  1. SwarmSurvivalNet (swarm_survival.py + swarm_adapter.pt): 60-feature log1p-normalized residual MLP for swarm liveliness (92.42% AUROC on Public Tracker N=757; 96.83% AUROC on Multi-Source N=10,278).
  2. SwarmPriors (swarm_priors.py): 4-pathway Empirical Bayes tracker/DHT peer-seeder priors; 50% modality dropout for zero-seed resilience.
  3. ZeroCueCatalog (zero_cue_plugin.py + validated_zero_cue_catalog.json): 929,400-title BLAKE2b-64 offline catalog (26.0 MB) for zero-cue disambiguation (+23.86 pp on N=345 zero-cue titles).
  4. AudioLanguageGuard (audio_language_guard.py): Audio/subtitle-trap guard (VOSTFR/VOSE/Eng.Sub foreign audio vs. true English dubs).
  5. CalibrationAndDHT (calibration_and_dht.py + calibration.json): Per-head temperature scaling (ECE = 1.93% title language, 1.09% owner scope); Mondrian conformal gates (adult >= 0.5451, junk >= 0.1010, owner_scope >= 0.4770).
  6. ArmigerDHTDBSync (armiger_dht_db_sync.py + csam_guard.py): Opt-in BEP-44/46 DHT sync and pre-ingestion CSAM filter (opt_in_sharing=False by default).

Benchmark Verification (Suite B LOTO, W5A-Test, & R5-Test)

• Suite B LOTO N=1,025 — Non-Circular Leave-One-Teacher-Out: Scored against held-out 2-of-3 teacher consensus (each LLM vs. the other teachers). Armiger Ultra v3: 96.86% 6-dimension composite (95% CI [96.35%, 97.38%]; INT8 96.85%), ahead of Gemini Flash (94.56%), OpenAI GPT Fast Tier (93.49%, 5-dim; Junk N/A (Unprompted)), Claude Haiku 4.5 (92.88%), and Local 8B Instruct (84.79%). Junk/Malware (N=977, 65 positives): 99.23% Balanced Accuracy (64/65 caught).
• W5A-Test N=703 — Rule-3 Gold holdout: 97.01% 11-class accuracy (95% Wilson CI [95.47%, 98.04%], 93.06% Balanced Accuracy); 96.87% 6-core-head exact match (681/703).
• R5-Test N=996 — 7-key zero-leakage holdout (snap_wave13_r5_20261021): No cross-split leakage on infohash, norm_name, work_id, imdb_id, CleanTitle+Year, FileTreeSig, BytePrefix64. FP32 91.27% / INT8 91.16% 9-class accuracy (86.27% BalAcc), 99.22% Title Language (n=511), 99.09% English Subtitles (n=110), 98.90% Owner Scope (n=996, 99.13% BalAcc), 95.82% Anime (n=239, 86.84% BalAcc), 84.49% 8-class Audio (n=303), 90.34% BIO Balanced Byte Accuracy (n=882, 59.18% exact span, 0 illegal transitions).

Quickstart (Unified 6-Plugin Pipeline)

from armiger_ultra_pipeline import ArmigerUltraPipeline

pipeline = ArmigerUltraPipeline.from_directory(".", variant="ultra")
res = pipeline.classify({
    "name": "Tears.of.Steel.2012.2160p.UHD.BluRay.REMUX.HEVC.TrueHD.7.1.Atmos-OPEN",
    "size": 68_400_000_000,
    "files": [{"path": "Tears.of.Steel.2012.mkv", "size": 68_400_000_000}],
    "swarm": {"tracker_known": True, "tracker_seeders": 142, "tracker_leechers": 19, "tracker_completed": 810},
})
print(res["category"], res["mondrian_gates"]["passed_all"], res["swarm_viability"]["viability_tier"])

Connectors & Integrations

• Lightweight CPU model (8.3 MB INT8 only): NorviTech/armiger
• Core source & evaluation harness: spencernorton/armiger · License: Apache-2.0
• Standalone sidecars for unmodified indexers (spencernorton/Armiger-Connectors):
  • Stock Bitmagnet (bitmagnet-io/bitmagnet): bitmagnet/bitmagnet_connector.py — polls untagged torrents via GraphQL, runs vectorized Armiger inference, prunes .exe/.scr/.lnk malware and honeypot traps, writes armiger-* tags.
  • Stock Rats Search (librats/rats-search): rats-search/rats_search_connector.py — drop-in REST/SQLite sidecar classifier and safety filter.
• Turnkey indexer, Torznab & *Arr stacks (examples/ at v2.1.3 · Integration Guide):
  • DH-Waypoint processorapi/v1 plugin (spencernorton/DH-Waypoint): examples/dh-waypoint/ (armiger_waypoint_processor.py + 07-armiger-ultra.plugin.yml + docker-compose.yml) — admission (/v1/admit), batch (/v1/classify), 4-tier feedback (/v1/feedback).
  • Bitmagnet pre-ingest & GraphQL bridge: examples/bitmagnet/ (bitmagnet_armiger_bridge.py + docker-compose.yml).
  • Torznab pre-grab proxy (Prowlarr/Sonarr/Radarr/Lidarr/Readarr): armiger_torznab_proxy.py — filters malware lures, dead swarms, and unrequested foreign-audio releases from Torznab XML before grab.
  • Post-import *Arr/qBittorrent webhook: armiger_arr_webhook_receiver.py — ffprobe ground truth for local conformal calibration and opt-in anonymous BEP-44/46 updates.