warp is the Layer 3 operating system layer of the Siliconcordat stack.
It isolates AI agent tools inside Linux sandboxes, verifies local model adapters before they touch GPU memory, and gives Linux hosts single-reboot disk recovery.
It is built for operators running local open-weight models and coding agents who want hard system limits instead of prompt-only rules.
| Layer | Repository | Role |
|---|---|---|
| L5 Improved MCP | siliconcordat/mde |
Structured tool protocol and code editing seams |
| L4 Harness | siliconcordat/exo |
Agent loop, memory log, and human approval gate |
| L3 OS | siliconcordat/warp |
Sandboxed execution, verified model loading, and OS recovery |
| L2 Model | siliconcordat/weft |
Open-weight adapters and deterministic adapter compilers |
| L1 Hardware | siliconcordat/loom |
Hardware architecture research and theory |
- Verified model gatekeeper (
warp-modeld): A zero-dependency Python 3 standard-library service sitting between L4exoand L2llama-server. It checks model fingerprints, lineage hashes, runtime flags, optional L5zone-manifest/v0.1bank-tag pins, exact integer math certificates (including nativesiliconcordat/weftmanifest ingestion), and system-owned canary tests before activating any adapter. - Instant safe rollback: Base weights stay on read-only storage. A kill switch sets adapter scales to zero, flushes every required slot state bank (
kv_cache,ssm_recurrent, andlattice_horizontal_streamsfor Latent Lattice architectures), quarantines the slot on incomplete flush (ERR_LATTICE_FLUSH_FAILED), and writes per-bank signed receipts to an append-only JSONL ledger. - Mandatory tool sandboxes: Declarative
Landlock,seccomp, andbubblewrapblueprints confine agent tools inside an empty network namespace with read-only access to verifier canaries and base weights. - Recoverable Linux images (
mkosi+btrfs+ UKI): Declarative Arch Linux profiles (vmtest,compute-rtx3080,client-omarchy) pair read-onlybtrfsroot snapshots with archived Unified Kernel Images (UKI) for single-reboot recovery.
| Path | Contents |
|---|---|
src/warp_modeld/ |
Four-stage fail-closed verifier (fingerprint.py, certificate.py, canary.py, llama_client.py, state_banks.py, ledger.py, verifier.py, server.py, cli.py) |
schemas/ |
Adapter bundle JSON schema (adapter-bundle.schema.json), verifier-owned canary suite (verifier-canaries.json), and daemon config template (modeld-config.example.json) |
blueprints/mkosi/ |
Base Arch btrfs + UKI config (zero sudo), post-install and ancestor mode-hardening scripts, and vmtest, compute-rtx3080, and client-omarchy profiles |
blueprints/systemd/ |
Hardened service units for llama-server (--lora-init-without-apply), warp-modeld, and pre-upgrade checkpoints |
blueprints/sandbox/ |
Landlock rules, default-deny seccomp filter, and fail-closed bubblewrap tool executor (WARP_TOOL_EXECUTOR_REQUIRED=1) |
blueprints/recovery/ |
btrfs snapshot and archived UKI checkpoint/rollback helper (warp-recovery.sh) |
tests/ |
Unit, end-to-end Unix socket daemon, blueprint validation, and adversarial tests against a local HTTP mock of llama-server |
Requires Python 3.11 or newer with no external runtime packages:
PYTHONPATH=src python3 -m unittest discover -s tests -vVerify a signed ledger file, archive and reseed a damaged ledger explicitly, run a division-free certificate check, probe an installed llama-server binary, or run the opt-in live GGUF Stage 1 to Stage 4 test:
PYTHONPATH=src python3 -m warp_modeld.cli verify-ledger --ledger <ledger.jsonl> --key-file <ledger.key>
PYTHONPATH=src python3 -m warp_modeld.cli repair-ledger --ledger <ledger.jsonl> --archive-dir <archive> --key-file <ledger.key> --confirm <phrase>
PYTHONPATH=src python3 -m warp_modeld.cli check-certificate --cert <cert.json> --bounds <bounds.json>
WARP_LIVE_LLAMA_TEST=1 PYTHONPATH=src python3 -m unittest tests/test_live_llama_smoke.py -v
WARP_LIVE_LLAMA_TEST=1 WARP_LIVE_BASE_GGUF=<model.gguf> WEFT_SRC=<weft/src> PYTHONPATH=src python3 -m unittest tests/test_live_llama_smoke.py -vwarp separates human control from experimental Linux compute so a failing driver or model test cannot break a primary workstation:
- Clean macOS Base Hub (64 GiB unified memory): Runs stock macOS with zero OS changes. Holds human keys, source policies, and approval prompts. Can run read-only user-space
llama-serveron demand for 27B model checks. - RTX 3080 Compute Host (8 GiB VRAM, 32 GiB RAM): Runs a reversible dual-boot
warpLinux image alongside Windows 11. Serves 4B to 8B open-weight models with multiple adapters and compiles adapters on CPU. - Omarchy Linux Client (lightweight x64 laptop): Runs the desktop shell, L4
exo, L5mde, and local sandboxes on a dedicated test laptop (or inside a VM on shared hardware), sending large model requests over a private link to the compute host.
- Verified offline now (53 passing unit, RPC, adversarial, and blueprint tests plus 1 opt-in live GGUF test): Covers all four
warp-modeldstages, Unix socket RPC (warp_infer,warp_complete), and CLI subprocesses. Adversarial cases verify forged certificates, inflated scales, stale lineage, flag and zone-manifest drift, forged canaries, ledger tampering, and blueprint invariants. - Verified against simulated multi-bank Latent Lattice backend:
tests/test_lattice_state_isolation.pyproves KV-only flush lets poisoned horizontal latents re-contaminate fresh KV writes, verifiesERR_LATTICE_FLUSH_FAILEDquarantine on KV-only or lying backends, and confirms atomic multi-bank (kv_cache+ssm_recurrent+lattice_horizontal_streams) rollback with per-bank signed receipts. - Verified live on
llama-server0.5.0 (build 11146, commit7fe450e19, CPU mode onQwen2.5-1.5B-Instruct): Full Stage 1 to Stage 4warp-modeldlifecycle over a0600Unix socket mounted a realweftrank-4 F16 GGUF v3 LoRA (84,384bytes onblk.27.ffn_up.weight, 0-byte re-derivation). Stage 2 ran in16.71 us, 5 live shadow canaries cleared in4.90 s, and kill-switch plusPOST /slots/0?action=eraseran in4.22 ms, restoring bit-exact baseW0output with verified HMAC receipts. - Requires Linux target hardware or real Lattice serving engine (not yet executed): Building and booting
mkosidisk images in QEMU, flashing dual-bootcompute-rtx3080orclient-omarchylaptops, live CUDA multi-adapter benchmarks on the 8 GiB RTX 3080 host, and live hardware execution on a real Latent Lattice checkpoint once a GGUF/C++ Lattice backend ships.
- ARCHITECTURE.md explains the zero-trust inference boundary, the four verification stages, the three-machine layout, and recovery rules.
- PLAN.md lists milestones, acceptance gates, and non-goals.
Code in this repository is licensed under the Apache License 2.0 (LICENSE). Documentation is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).