# Known limits and honest problems

AetherForge Linux (forge-mesh) is built for **machines you own and authorize**. This document records deliberate gaps and platform realities—not a bug list to fix silently.

## Deployment and spread

- **No worm-style LAN spread.** WinRM/GPO/BITS-style push lanes have no Linux equivalent in scope. Authorized summon (`curl | bash` with pin/campaign), SSH keys, Ansible, and enrolled sibling spread replace blind propagation.
- **Triple onion is enrolled-only.** The 14-tier Linux deploy chain runs on hosts you have already authorized—not arbitrary internet targets.
- **Earn-before-burn is phenotype-scoped.** Autospread to siblings only after local hashrate exceeds threshold for N minutes, and only within the same enrolled phenotype group.

## Tunneling and networking

- **cloudflared is sidecar-only on Linux.** The deck does not embed cloudflared in-process. Use `LAUNCH.sh` or your own unit to run `cloudflared tunnel run --token …` and set `AF_TUNNEL_EXTERNAL=1`. Agents must be configured to reach the tunnel hostname, not assume in-process tunnel state.
- **WireGuard is operator-managed.** The plan includes WireGuard mesh (tier T9) and API hooks, but Forge Mesh does not auto-provision kernel WireGuard configs, key rotation, or NAT traversal. You install `wireguard`, distribute keys, and point agents at the deck overlay IP.
- **Subnet mapping is opt-in.** ARP/ping sweeps run only on **operator-declared CIDRs** from the dashboard. There is no blind or internet-wide probing.

## Mining and hardware

- **GPU tiers need drivers first.** lolMiner/Rigel KawPoW paths assume `nvidia-smi` or `rocm-smi` already work. The summon installer does not install proprietary GPU stacks.
- **Huge pages and cgroups vary by host.** Agents attempt tuning (`vm.nr_hugepages`, cgroup v2 caps) but success depends on capabilities, sysctl policy, and distros.
- **Stratum upstream is yours.** The deck proxies to pool URLs you configure; wallet policy pins one address but pool connectivity and ban risk remain operator concerns.

## Intelligence and Court

- **Court / Seer needs an LLM you operate.** Every exhaust tick that invokes prosecutor/defender/judge requires local Ollama or a configured OpenAI-compatible API. There is no bundled model.
- **eBPF telemetry is optional.** Build tag `ebpf` enables deeper signals; default builds may lack thermal/steal-time probes Court prompts expect.
- **Clearance gates are policy, not crypto.** L0–L4 gates remote pause, reboot, screenshot, shell, and verdict dispatch—but compromise of deck credentials bypasses the model.

## Packaging and install

- **Signature verify is a placeholder until W8 forge signing ships.** `install.sh.tpl` documents ed25519 verification; dev builds may skip verify with a warning.
- **USB portable deck is not hardened live media.** `pack-usb.sh` bundles config templates and optional binaries; operators must rotate passwords, fleet secrets, and tunnel tokens before production use.
- **Screenshot on headless hosts is best-effort.** `grim`/`scrot` or stubs may return placeholders without a display server.

## Persistence and scale

- **SQLite is single-node default.** Multi-deck HA and Postgres are optional future paths; do not expect active-active fleet state on SQLite alone.
- **Erasure and torrent tiers need seeded artifacts.** T12/T13 assume shards or seeds exist and are reachable; cold start without forge output will fail those tiers.

## What we intentionally do not promise

- Windows parity for every persistence or evasion mechanism.
- Mining profitability or algo selection optimality—only tiered fallback to one wallet.
- Legal compliance in your jurisdiction—self-hosted fleet ops are your responsibility.

For architecture and workstream ownership, see the project plan at `.cursor/plans/aetherforge_linux_edition_71dd0fbb.plan.md` (or `docs/` once mirrored in-repo).