TheOrc v1.11.3
TheOrc v1.11.3
100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.
What's new in v1.11.3
Context Fabric's NoKvSlot mystery is finally closed — and it isn't TheOrc's bug. A months-old class of infrastructure crashes on Gemma-4-12B runs turned out, after direct A/B elimination of every leading theory (cross-conversation KV exhaustion, SWA cache sizing, force-recycle timing), to be a Gemma-4-specific limitation in upstream llama.cpp, not a defect in TheOrc's own runtime. Meta-Llama-3.1-8B and qwen2.5-coder-7b show zero NoKvSlot occurrences across 500+ combined benchmark questions on three independent machines — the crash is isolated to one model family, and Context Fabric's own retrieval/reduction mechanics are proven sound underneath it.
Retrieval quality climbs from 31% to a validated 48%+ pass rate through three deterministic fixes, not guesswork. Each tier targets a specific, measured failure mode found by categorizing every B3 miss against the evidence actually supplied to the model: Tier 1 fixes multi-word entity names dissolving into common-word noise (bag-of-words scoring couldn't tell "Station Alpha" from any card containing "station" and "alpha" separately); Tier 1.5 adds proximity-pair matching for paraphrased questions that invert entity word order; Tier 2 stops silently rejecting oversized evidence cards and truncates them instead. Cumulative effect on the 100-question suite: pass rate 31→45→56, pure retrieval misses 49→21→10, B3 now clearly beats a conventional top-k RAG baseline for the first time.
Tier 2.5 closes two more real gaps and finds the honest edge of what's left. Multi-hop chain questions were failing because the greedy evidence-fill spent its whole token budget on distractor segments that share an entity name with the question but belong to an unrelated fact chain — a reference-chasing pass now follows the chain's own shared identifiers into the linked segments, with 30% of the evidence budget reserved specifically so it always gets a turn. Measured live: full-retrieval MultiHop cases rose 6→9 out of 24. Separately, root-caused and fixed the long-standing boundary-stitch failure (Meta-Llama emits its linkedFacts field as JSON objects instead of strings — a tolerant parser now handles both) and closed a citation-precision shortfall, both now passing their gates and holding stable across five independent live validation runs. What's left is honestly reported, not hidden: the remaining MultiHop gap is a model-instruction-compliance ceiling (the model under-citing a multi-part chain answer), not a retrieval defect — logged with three scoped next options rather than claimed as solved.
Model Benchmark window (Phase 1) and safer benchmark tooling. A new read-only panel under Models → Model Benchmark… surfaces every model's CF-7 benchmark history — GO/NO-GO verdicts, question pass rate, citation precision, segment coverage — scanned automatically from .orc/adversarial/ artifacts. The benchmark CLI also gained --model/THEORC_CF_MODEL pinning so a shared model depot changing mid-run (a model toggled on or off by unrelated work) can no longer silently swap which model a benchmark run actually measures, plus verbose preflight diagnostics that print every candidate model's admission verdict and any disabled GGUFs on rejection.
Training Pit tab redesign. The dataset/training dashboard now scrolls as one page (a right-side scrollbar reaches every section regardless of window size), and the three pipeline stages — Generate Dataset, Orc Academy, The Foundry — are laid out side by side as numbered, color-coded cards instead of stacking as individually-expandable accordion sections. Inventory tiles (datasets/adapters/models) now read as stat cards with a colored accent bar and a large count.
Known issue: opening a workspace folder while the Training Pit tab is active can trigger a native stack-overflow crash under investigation — isolated reproduction attempts (real window, real repo data, multiple layout widths) have not yet caught it live; tracked for a follow-up patch.
🪟 Windows — one-click install (recommended)
- Download OrchestratorSetup.exe below
- Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
- Launch from the Desktop shortcut or Start Menu
Portable: download TheOrc-1.11.3-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.
🍎 macOS (Apple Silicon)
- Download OrchestratorSetup below (no extension)
- In Terminal:
chmod +x OrchestratorSetup && ./OrchestratorSetup - First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run
xattr -d com.apple.quarantine OrchestratorSetupfirst. - The wizard detects your hardware, downloads the right runtime and model, configures everything
Portable: download TheOrc-1.11.3-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.
Intel Macs (osx-x64) aren't built yet — only Apple Silicon.
🐝 Warband (headless HIVE node, no GUI)
Deploy theorc-warband on any machine you want to join your HIVE without opening the
full app — a cloud box, a home-lab server, a Raspberry Pi. Download the binary for your
OS below, chmod +x theorc-warband, then either run it directly or set it up as a
systemd service (Linux) / launchd agent (macOS) so it survives reboots. See
--show-identity and --pair --target <host> --expect-fingerprint "<phrase>" for
joining an existing HIVE headlessly.
🖥️ System requirements
| OS | Windows 10/11 (x64) · macOS (Apple Silicon) |
| .NET | Not required — fully self-contained |
| GPU | NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline) |
| RAM | 8 GB minimum · 16 GB recommended |
Full changelog: v1.11.2...v1.11.3
Built from v1.11.3 · .NET 10 · Source · Support the Orc ☕