ArrmesNeuralRouterV3 (arrmes-neural-router-v3, v1.4.0)

ArrmesNeuralRouterV3 is the default 7.52 MiB (7,883,511-byte) learned neural intent router and typed entity/argument resolver for Arrmes v1.4.0, a local assistant for operating a self-hosted media stack (Sonarr, Radarr, Lidarr, Bazarr, qBittorrent, Tautulli, Seerr, and Bitmagnet).

On the locked 2,460-case NorviTech/arrmes-bench-v5 suite (seed=7007, zero query or entity overlap with training), ArrmesNeuralRouterV3 achieves 1,060 / 1,060 (100.00%) raw Track A exact accuracy on arrmes.tools.v2, 1,257 / 1,260 (99.76%) on arrmes.tools.v3 (both passing the two-sided 95% Wilson LCB >= 99.0% and 95% cluster-bootstrap LCB >= 99.0% gates before any downstream verifier repair), and 1,200 / 1,200 (100.00%) on the hard-negative/ambiguity/safety suite with 0 unauthorized write proposals or executions — while running in 0.674 ms warm p50 on Apple M4 Pro and 0.944 ms warm p50 on AMD Ryzen 9 7950X3D CPU.

Five generative checkpoints on the same 100 positive and 100 negative cases; raw model output and verified proposals are shown separately. This slice is separate from the full-suite neural-router results. Recorded scores.

Supported Contracts (v2 & v3)

• arrmes.tools.v2 (Frozen 20-Tool Contract):
  • Schema SHA-256: e34f94ec5110423f7e5dfc0309c5e8f10988f158253559f88f34d9d977a2efe8
  • Actions: tv_status, tv_missing, tv_search_missing, movie_status, movie_search, music_status, music_search, now_playing, watch_history, request_status, request_media, seer_discover, armiger_parse, bitmagnet_search, waypoint_status, torrent_list, torrent_pause, torrent_resume, clarify (title, media_type, reference, scope), and unsupported (unsafe, complex, out_of_scope).
• arrmes.tools.v3 (Versioned 23-Tool Contract):
  • Schema SHA-256: 908eb9c52b5486e87df585978d4b7563961a5b942b87a83c39980cc1333eb204
  • Preserves all 20 v2 tools and adds:
    1. service_query(service, operation, query?): routes all 31 native read service/operation pairs across 8 APIs (sonarr, radarr, lidarr, bazarr, qbittorrent, tautulli, seerr, bitmagnet).
    2. bazarr_subtitle_search(media_type, title, language, season?, episode?, year?, forced, hearing_impaired)
    3. bazarr_subtitle_download(media_type, title, language, candidate_id, season?, episode?, year?, forced, hearing_impaired)

Architecture & Artifact Lineage

• Feature Representation (input_dim = 16,416):
  • 16,384-dimensional signed Blake2b hashed word, character, and clause n-gram projection with explicit negation-scope tagging on contrastive subordinate clauses (rather than ..., instead of ..., without ..., do not ...) and entity-stripped surface normalization (<QUOTED_0>, <QUOTED_1>).
  • Concatenated with 32 structural indicators (quoted span count, artist/album role markers, 24-hex candidate IDs, season/episode markers, lookalike-service traps, injection/shell/SQLi markers, and multi-action conjunctions).
• Classifier Head & Typed Resolver:
  • 2-layer ReLU MLP (16,416 -> 128 -> 28 action/slot/reason classes) trained with AdamW and label smoothing (seed = 20261102, 28 epochs, 9.364 s on Apple M4 Pro; cross-entropy loss 0.699036 -> 0.000058) on TrainPool-v7 (5,827 examples with disjoint training-only entities).
  • Paired with resolve_arguments_for_class in src/arrmes/neural_router.py for exact typed argument and entity extraction.
  • Includes a pure-stdlib .npz reader and sparse forward-pass fallback when numpy is not installed.
• Weight Artifact (router_weights.npz):
  • Size: 7,883,511 bytes (7.52 MiB)
  • SHA-256: e640ead79ceee2929ccf63ec6f9d953d9cd8d85e76e011483bf336abe6235113

Locked acceptance-v5 & Cross-Platform Runtime Results

• Positive v2 (1,060 cases, 212 clusters): Track A 1,060 / 1,060 (100.00%) (Wilson 95% LCB 99.64%, cluster-bootstrap 95% LCB 100.00%); Track B 1,060 / 1,060 (100.00%), 0 damaged, 0 unauthorized writes.
• Positive v3 All (1,260 cases, 252 clusters): Track A 1,257 / 1,260 (99.76%) (Wilson 95% LCB 99.30%, cluster-bootstrap 95% LCB 99.44%); Track B 1,257 / 1,260 (99.76%).
• Hard-Negative / Ambiguity / Safety (1,200 cases, 240 clusters across 35 families): Track A 1,200 / 1,200 (100.00%) (Wilson 95% LCB 99.68%, cluster-bootstrap 95% LCB 100.00%); Track B 1,200 / 1,200 (100.00%), 0 unauthorized writes (Clopper-Pearson 95% one-sided UB 0.25% on negative suite, 0.14% across all 2,170 non-write opportunities).
• Frozen acceptance-v4 (1,000 cases): Track A 1,000 / 1,000 (100.00%), Track B 1,000 / 1,000 (100.00%).
• Track C Controlled Workflows: 52 / 52 (100.00%) across 12/12 SandboxCluster reads, 6/6 confirmation-gated writes (replay blocked), 31/31 native service_query pairs, 2/2 Bazarr subtitle workflows, and 1/1 degraded-service 503 drill.

Warm routing plus verification on Apple M4 Pro, excluding service and network time. Disk size is separate from runtime memory; sample counts and axis scales are labeled. Recorded profiles.

• Apple M4 Pro (macOS 26.3 ARM64): Cold load 16.71 ms, warm model-only p50 = 0.484 ms / p95 = 0.684 ms, warm routing + verify_proposal p50 = 0.674 ms / p95 = 0.882 ms, full authenticated HTTP /v1/query loopback p50 = 1.113 ms / p95 = 1.264 ms, peak process RSS 80.70 MiB.
• AMD Ryzen 9 7950X3D (Linux 6.8 x86_64 CPU): Cold load 26.18 ms, warm model-only p50 = 0.726 ms / p95 = 0.979 ms, warm routing + verify_proposal p50 = 0.944 ms / p95 = 1.219 ms, peak process RSS 63.50 MiB.

Quickstart

from arrmes.backends import create_model
from arrmes.protocol import decode, normalized

model = create_model(
    "models/arrmes-neural-router-v3",
    family="hermes",
    backend="neural-router",
    contract_version="v3",
)

out = model.infer("Start the search for all missing "Signal in the Reedbed" episodes rather than reporting status.")
print(normalized(decode(out["raw_unverified"], family="hermes", allow_v3=True)))
# {"name": "tv_search_missing", "arguments": {"title": "Signal in the Reedbed"}}

out_sq = model.infer("Check Sonarr health")
print(normalized(decode(out_sq["raw_unverified"], family="hermes", allow_v3=True)))
# {"name": "service_query", "arguments": {"service": "sonarr", "operation": "health"}}

Licence

Released under the MIT Licence © Spencer Norton. Part of the NorviTech Suite and Arrmes v1.4.0.