Project hub

The operating system for the sovereign machine.

AetherOS is a capability-first personal OS research project: privacy by construction (VeilNet grants), universal execution via host AetherIR (hand JIT + CLIF paths), Liminal microkernel isolation under QEMU, and a local twin shell that leaves only under an explicit grant - never telemetry-first, never ambient authority. Formal claims climb a public verification ladder (T0-T1 active).

Phase Bootstrap+ | Cliffs Bridge
Demo gates D1-D327 green (lab)
Updated 2026-09-21 | D1-D327 lab
License Proprietary

This page is the living map of inception, architecture, and lab evidence. Green means QEMU serial and/or host lab observation - not an install CD and not a public source dump.

Planning horizon (~18 months)

Ambition bar for a personal-class OS path, not a Continuum gate counter. Lab gates can stack high while product cliffs (real Verus Tier 4, Fabric TX product, compositor, agents, multi-site WAN, kernel AEAD hop peel) still dominate the remaining road. Estimate only; not a ship date.

Now: D1-D327 lab depth Ambition ~32% | not done

Inception

Why this project exists as its own category.

AetherOS is a local-first, privacy-sovereign multi-layer personal computing system - a microkernel plus fabric plus swarm - aimed at people who want authority to be explicit, execution to be hostable, and intelligence to stay on-device.

HARDWARE IS DISPOSABLE. YOUR OS SHOULD MOVE WITH YOU ACROSS DEVICES.

Design lineage is honest about its teachers: seL4-style capabilities, Zircon-like bulk memory rights, Genode-like recursive worlds, Barrelfish-like explicit messaging later - always remixed into Liminal's own object model rather than a fork.

Private monorepo (offline git + survival mirrors). This site is the shareable surface for architecture and progress - the source remains proprietary R&D.

Principle enforcement

Manifesto constraints as an index into architecture, lab gates, and cliffs. Not a proof certificate and not marketing.

Constraint index over existing map chapters
Principle Layers Mechanism Gate / cliff Map
P1 Privacy is ontological L0 L3 L6 EgressGrant; cover bound; zero net at birth D3 D4 / N4 M1 Grants | Lab | Adversary
P2 Universal execution L2 AetherIR + triple path into capability host D2 / N1 M5 Runtime
P3 Formal foundations cross Ladder T0-T1 active; T2-T5 calibrated climb D* -> T0-T1 / N3 M2 Verify
P4 Generative continuity L4 L1 Local UI/twin; no ambient phone-home stub / M3 M4 Arch L4
P5 Hardware disposable L5 L6 Devices as projection; swarm continuity deferred / N2 M6 Cliffs
P6 Temporal sovereignty L6 Fork/discard/merge + grant world hygiene D5 / M6 Grants
P7 Cognitive sovereignty L6 L1 Twin egress only under explicit grant D5 / M3 M6 Grants
Manifesto one-liners
  1. Privacy is ontological - no ambient network path exists to forget to close.
  2. Universal execution - foreign binaries compile into a capability host via AetherIR.
  3. Formal foundations - security climbs a verification ladder; it is not asserted in marketing.
  4. Generative continuity - the UI is alive and local; the twin never phones home by default.
  5. Hardware is disposable - devices are projection surfaces for a personal swarm.
  6. Temporal sovereignty - fork timelines to experiment; discard or merge with grant hygiene.
  7. Cognitive sovereignty - the encrypted twin leaves the swarm only under an explicit grant.

Full architectural plan

Seven layers. Layer 0 is the root of authority; upper layers never invent ambient rights. Expand a layer for depth. Scan the authority flow first.

Live / lab green Design / stub / scaffold Deferred

Proven vs plan vocabulary

Collapsed layer lines state what is proven. Expanded depth may describe architectural plan. Green tags never mean full plan complete.

Do not read as now Allowed now Plan / cliff
Full ELF/PE/Mach-O product loaders IR + JIT + CLIF host demos (D2) N1 cranelift-jit; M5 personalities
Personal mix ring as finished product Phase0 grants + multi-hop host lab Full mix; kernel NIC (M1)
Living UI weaver as shipping UI Stub / minifb compositor M4 compositor beyond stub
Twin co-pilot product Twin shell surface + grant egress Cognitive companion depth
Formally verified kernel Verification ladder T0-T1 Active T2-T5 climb (scaffold / planned)

Cross-layer authority flow

Human is root of trust. L0 holds pure capability authority. L3 may only birth packets under EgressGrant. L6 twin and timeline paths do not invent ambient egress; they reuse grant hygiene. L1/L4/L5 never mint network rights on their own.

Authority edges (never ambient)
Cross-layer authority flow from human root through L0 to L1 L2 L3 L6
Diagram source (ASCII)
Human root of trust
 -> L0 Liminal microkernel (Live under QEMU)
 -> L1 agents (Design + stubs): only via held caps
 -> L2 runtime fabric (IR + JIT host): capability host rights only
 -> L3 VeilNet (Phase0 + multi-hop lab): EgressGrant packet birth only
 -> L6 swarm + timeline (host paths): twin egress only under explicit grant
L4 compositor (stub) presents; does not invent rights
L5 hardware synthesis (deferred): no rights invented yet
L6 --grant hygiene--> L3
L3 --no grant--> quiet (no packet)

L0 capability object model

Liminal does not grant ambient network or file rights at birth. Authority is carried in capability objects. Design lineage includes seL4-style caps, remixed into Liminal's own model - not an seL4 Isabelle import.

CSpace / caps

Capability table entries. Rights are explicit slots, not ambient process privileges. SerialWrite isolation demo: A ok, B denied.

Frame

Physical memory rights mapped into a VSpace. No ambient mapping of foreign frames without a held cap.

VSpace

Per-task address space with private CR3. Ring-3 enter via iretq; isolation proven under QEMU serial.

Endpoint

Explicit IPC path. Messaging is capability-gated, not ambient sockets.

SchedContext

Scheduling authority for preemptive RR and timer IRQ frame steal across dual VSpaces (lab green).

No ambient net/file

Construction rule: a fresh domain starts with zero network and zero ambient file authority. Upper layers may not invent those rights.

L0 objects and isolation path (public model)
L0 capability object model CSpace frames VSpaces endpoints SchedContext
Diagram source (ASCII)
CSpace/caps -> Frame
CSpace/caps -> VSpace -> Ring-3 tasks
CSpace/caps -> Endpoint -> Explicit IPC
CSpace/caps -> SchedContext -> Preempt RR / IRQ steal
No ambient net or file rights -.-> CSpace
QEMU: boot -> ring3+VSpace OK -> dual VSpace RR -> IRQ steal OK

L2 Runtime Fabric Codex

How a foreign binary is intended to enter AetherOS: lift into AetherIR, then execute only under capability-host rights. This is not Wine-as-architecture and not a claim that every loader format is product-complete.

Host lab green (D2): hand JIT + CLIF interpret + CLIF native emit Plan / cliffs: multi-format loaders, personalities, HostSys depth; cranelift-jit optional feature wired

Universal execution (P2) means foreign code becomes a guest of the capability host via AetherIR. Proven today on the host fabric: hand JIT + CLIF interpret + CLIF native emit, plus optional real Cranelift JIT behind feature cranelift-jit (quad parity on pure arith demos). Full ELF/PE/Mach-O product loaders and multi-personality HostSys remain calibrated cliffs.

Pipeline objects

Foreign binary surface

Architectural intent: ELF/PE/Mach-O intake. Not every format is a finished product loader on the public map.

AetherIR

Intermediate representation for host execution paths. Lift/decode feeds the IR; rights stay outside the IR itself.

Triple execution path

Hand JIT, CLIF interpret, and CLIF native emit with parity demos under D2 on the host lab.

Capability host

Execution only with held caps. No ambient net or file rights are invented by the runtime fabric.

Foreign binary pipeline (public model)
L2 AetherIR pipeline from foreign binary to capability host with D2 triple path
Diagram source (ASCII)
Foreign binary (plan: ELF/PE/Mach-O)
 -> Lift / decode -> AetherIR
 -> Hand JIT (D2 green)
 -> CLIF interpret (D2 green)
 -> CLIF native emit (D2 green)
 -> Capability host rights only (no ambient net/file)
 -.-> Personalities / HostSys (M5 mid cliff)
 -.-> cranelift-jit feature (N1 lab green, optional deps)
Not Wine-as-architecture. Host lab != guest kernel universal exec claim.

Proven vs design

Item Status Proof surface Note
Hand JIT Active (host) D2 host demos / tests Part of triple parity
CLIF interpret Active (host) D2 host demos / tests Oracle / parity path
CLIF native emit Active (host) D2 host demos / tests native_loop demos
Full multi-format loaders Design / partial Architectural intent Not product-complete on map
Personalities / HostSys depth Mid cliff M5 Host design surface Grant-only HostSys demos later
cranelift-jit crate Lab green (feature) Optional feature on Real Cranelift JIT; default build stays light

Principle map

Principle Mechanism Gate / cliff
P2 Universal execution AetherIR + triple path into capability host D2 / N1 / M5
P1 Privacy ontological Hosted code gets no ambient net; egress still needs grants D3 + L3
P3 Formal foundations D2 is host behavioral + tests (T0), not translation proofs #verify

Layer map (progressive disclosure)

Collapsed row: id, title, status. Expand for purpose, evidence, non-goals. Status tags stay calibrated.

L0 Liminal Microkernel Live under QEMU

Pure capability authority (seL4 lineage), frames/VSpaces, endpoints, SchedContext. No ambient network or file rights.

Evidence: freestanding boot, cap table isolation, private VSpace CR3, ring-3, Yield-RR, IRQ steal, virtio rng/console/blk under QEMU. Object model above. Related gate: D1.

L0 object model

L1 Aether Agent Swarm Design + early host stubs

Sandboxed local AI operators: binary archaeologist, API cartographer, security guardian, cognitive companion, UI weaver, driver synthesizer.

Non-goal today: shipping autonomous product agents. Agents act only through held L0 caps; no ambient network from any agent role.

L2 Universal Runtime Fabric IR + JIT live on host

Proven on host (D2): hand JIT + CLIF interpret + CLIF native emit (triple parity demos). Foreign binaries are intended to enter a capability host, not Wine-as-architecture.

Plan (not all green): ELF/PE/Mach-O loaders, NBTE AOT/JIT cache, multi-personality HostSys (M5). N1 real Cranelift is optional feature lab green. IR + JIT host is the live tag.

Runtime Fabric Codex

L3 VeilNet Privacy Fabric Phase0 + multi-hop swarm

Proven on host (D3/D4): EgressGrant packet birth, multi-hop path grants, cover bound to live grants. Construction property: no external packet without a live grant.

Plan (not all green): personal mix ring, cover morph, full mix profiles, kernel NIC TX. Tag remains Phase0 + multi-hop lab. See Grant continuity.

Grant lifecycle | Adversary ledger

L4 Generative Spatial Compositor Stub compositor

Proven today: stub compositor / minifb shell only. Generative spatial compositor is design language, not a shipping product surface.

Plan (not green): living UI weaver, multi-modal input (touch, voice, gesture). BCI remains deferred. Tag stays Stub compositor.

L5 Hardware Synthesis Deferred

On-demand driver synthesis under formal constraints; device fingerprint; verified Linux driver shim path (later).

No rights invented while deferred. When active, drivers must still enter through capability constraints.

L6 Personal Swarm & Temporal Continuity Mesh + timeline host paths

Owned-device mesh, encrypted digital twin, timeline fork engine for experiment and rollback. Hardware is disposable; the man is root of trust.

Host timeline fork/discard/merge green (D5). Twin egress only under explicit grant (P7). Cross-device durability is a mid cliff, not claimed finished.

Fork hygiene

Where we are

Honest status through Continuum D1-D327 (lab). Green means observed under test or QEMU serial, not marketing. D308-D327 is guest-directed AEAD hop peel: host ChaCha opens each layer, and the kernel refuses consume until the peel count matches. Next lab gate: D328 kernel-resident AEAD. Product cliffs remain (real Verus Tier 4, multi-machine WAN, ChaCha inside the no_std kernel).

Where D328 starts states the boundary. The gate is not green.

Showing the latest row only (most recent updates). Expand for D1 through earlier waves.

Grant continuity

How packets are born, how cover dies with grants, and how timeline forks keep experimental egress from leaking into the parent world.

Host fabric lab (D3-D5 green under lab / QEMU paths) Kernel NIC caps, full mix, COW kernel timelines: design or deferred

VeilNet is not a default-deny firewall bolted on later. It is a construction property: a fresh domain has zero network authority. External packets exist only while a live grant says they may. Cover traffic is charged to that grant and ceases when the grant dies. Timeline forks carry experimental grant worlds; discard revokes them; merge never auto-imports net grants without an explicit human gate.

Objects

EgressGrant

Capability object that authorizes packet birth for a subject domain. Scope, rights, budget, and cover binding are part of the object, not ambient policy files.

Path grant

Multi-hop authorization along an owned-device or lab path. Intermediate hops do not invent ambient clearnet rights. Observed on host multi-hop WAN TCP lab paths.

Cover cell

Grant-bound cover schedule. While the grant is live, cover may run; when the grant is revoked or expires, cover stops. Idle domains stay quiet by construction.

Experimental grant world

A timeline fork's tagged grant set used for experiment. Discard tears the world down. Merge into parent requires hygiene so residual egress cannot sneak through.

Three VeilNet laws

  1. Packet birth. External TX requires a live EgressGrant (or composed path grant). No grant means no packet.
  2. Cover bound to grant. Live grant implies cover charged to that grant's lifetime. Cover is not a free ambient presence channel.
  3. Trust is explicit. Peers are untrusted by default. Crypto and predicates, not ambient trust domains, authorize exchange.

Cover construction property

Cover is not ambient presence. It is a budgeted schedule charged to a live grant. While the grant lives, payload and/or cover cells may leave the domain. When the grant dies, cover ceases by construction so idle domains go quiet. That is a local construction property on host fabric paths, not a free win against a global passive adversary.

Cover lifetime equals grant lifetime
Cover construction property: grant live charges cover, REVOKE returns idle quiet
Diagram source (ASCII)
Fresh domain (zero net)
 -> No live grant -> wire: no domain-originated cells
 -> mint EgressGrant -> Grant live
 -> Cover charged to grant lifetime
 -> Payload and/or cover cells may leave
 -> REVOKE or expiry -> cover stops -> Idle quiet
Lab OPEN/TICK/REVOKE exercises this; not a consumer API.

Adversary observation ledger

Observer class: local underlay / link observer (A-ISP class). Explicitly not end-to-end global passive adversary (A-GPA) confirmation.

Domain state What may leave What observer may notice What stops / residual
No grant Nothing (no ambient TX, no cover beacons) Silence on that fabric path Other host/OS noise outside VeilNet may still exist
Live grant + cover Payload and/or grant-bound cover cells Presence of traffic; volume/timing shape of schedule Raises idle fingerprint cost; does not erase timing classes
REVOKE / expiry No new birth; cover scheduler stops Traffic for that grant ends toward quiet History already observed; other live grants may still emit

VeilNet Lab Console

Interactive browser model of OPEN / TICK / REVOKE and timeline FORK / DISCARD / MERGE. Demonstrates the construction property: no external packet without a live grant; cover dies with the grant.

Browser simulation Models D3/D5 public semantics Not a live mesh / not a consumer API

Domain + wire

Domain NoGrant
Cover IdleQuiet
Budget 0
Timeline Main

Fabric path (cells vs silence)

A-ISP class: silence on this fabric path (no ambient TX)

Event log

Try: OPEN then TICK several times, then REVOKE and watch silence. With this console focused, O, T, and R do the same. Or FORK, OPEN on the experimental world, DISCARD (or MERGE with HumanGate). TICK with NoGrant is always DENIED. Browser sketch, not a live mesh, not kernel AEAD.

Grant lifecycle

Lab control plane words OPEN / TICK / REVOKE (plus PING / HOST / SHUTDOWN on standing services) describe host fabric cover-daemon exercises. Not a public consumer API and not a production multi-host product.

EgressGrant lifecycle (public model)
EgressGrant lifecycle NoGrant GrantLive PacketBirth CoverBound Revoked Quiet
Diagram source (ASCII)
NoGrant --mint--> GrantLive --packet birth / cover--> GrantLive
GrantLive --REVOKE or expiry--> Revoked --> Quiet (idle = quiet)
NoGrant: no ambient egress path exists

Timeline fork hygiene

Demo gate D5 is green on host timeline engine paths: fork, discard, and merge of experimental continuity. This is not a claim of full kernel COW snapshots or cross-device timeline migration. Those remain later phases.

Fork, discard, merge with grant hygiene
Timeline fork discard merge with experimental grant hygiene
Diagram source (ASCII)
Main --fork--> Exp --mint scoped grants--> LabWorld
LabWorld --discard--> Sink (grants dead, no residual egress)
LabWorld --merge + HumanGate--> Main (no auto net grants)
Twin egress only under explicit grant

Principle map

Principle Object / mechanism Demo gate
P1 Privacy is ontological EgressGrant packet birth; cover bound to lifetime; idle quiet D3
P6 Temporal sovereignty Fork experimental grant worlds; discard revokes; merge with hygiene D5
P7 Cognitive sovereignty Encrypted twin leaves the swarm only under an explicit grant D5 / L6

Verification ladder

AetherOS does not claim seL4-complete proofs. The ladder is tracked in public language here so expectations stay calibrated. Security climbs; it is not asserted in marketing.

Demo gates D1-D5 green means lab observation under QEMU serial and host tests. That is not the same as model-checked or proof-discharged security. Tier 0-1 are the live floor today. Higher tiers are scaffold, planned, or design-compatible only.

Status vocabulary

Active

Runs in the lab now: rustc, cargo test, and authored cap property tests.

Scaffold

Shapes, hooks, or design docs exist. Not a green proof CI suite and not discharged proof functions.

Planned

Direction of climb only. Full Verus / Lean proofs are not claimed.

Design-compatible only

Lineage may be compatible with an seL4-style substrate. Not an Isabelle-complete import and not a forked verified kernel.

Ladder climb

Verification tiers (honest stop-lines)
Verification ladder tiers T0 through T5 with D1-D5 feeding T0-T1 only
Diagram source (ASCII)
T0 rustc + cargo test ACTIVE
T1 Cap property tests ACTIVE
T2 Kani SCAFFOLD
T3 Verus-style hooks SCAFFOLD
T4 Verus / Lean full proofs PLANNED (not claimed)
T5 seL4 substrate option DESIGN-COMPATIBLE ONLY
D1-D5 lab evidence feeds T0-T1 only

Demo gates to evidence class

Each green gate has a primary evidence class. Empty formal cells are intentional: gates do not invent Tier 4 proofs.

Gate Lab claim Evidence class Ladder touch
D1 Kernel isolation under QEMU Runtime observation (serial markers) T0 (+ tests where present)
D2 Hand JIT + CLIF interpret + native emit Host behavioral demo + unit tests T0
D3 Deny-by-default packet birth; cover bound Host lab / property intent T0-T1
D4 Triangle AEAD + multi-hop WAN TCP lab Host lab demo (not crypto formal proof) T0
D5 Fork / discard / merge host paths Host behavioral demo T0

Tier scan table

Tier Tool Status
0rustc + cargo testActive
1Cap property testsActive
2Kani model checkingScaffold
3Verus-style hooksScaffold
4Verus / Lean full proofsPlanned
5seL4 substrate optionDesign-compatible only

What each tier buys

T0 rustc + cargo test Active

Buys: typecheck and unit tests on paths that run in lab.

Does not buy: kernel isolation theorems or machine-checked refines-to-binary proofs.

T1 Cap property tests Active

Buys: authored property / invariant tests where present.

Does not buy: model-checked or proof-discharged caps.

T2 Kani model checking Scaffold

Buys: harness intent and cfg-shaped hooks.

Does not buy: a claimed green Kani CI suite.

T3 Verus-style hooks Scaffold

Buys: Verus-shaped hooks and a design invariant document.

Does not buy: real proof fn discharge on a Verus toolchain.

T4 Verus / Lean full proofs Planned

Buys: direction of climb only.

Does not buy: any complete kernel proof today.

T5 seL4 substrate option Design-compatible only

Buys: capability lineage compatible with seL4-style authority as a design option.

Does not buy: seL4 Isabelle completeness or a verified seL4 fork shipping as AetherOS.

What it can do now

Shipped slices you can exercise on the lab host today.

Liminal microkernel (QEMU/UEFI)

  • Freestanding Rust no_std boot via bootloader 0.11 + OVMF
  • Capability table + SerialWrite isolation demo (A ok, B denied)
  • Per-task VSpace, ring-3 enter, int 0x80 syscalls
  • Yield-RR across dual VSpaces + timer IRQ frame steal
  • Virtio modern: rng, multiport console CTRL+TX, blk R/W + MULTI+FLUSH, net TX+RX (DHCP/SLIRP lab)
  • Legacy virtio fallbacks when modern path unavailable

Host fabric crates

  • AetherIR -> hand JIT + CLIF interpret + CLIF native emit + optional real Cranelift JIT (feature)
  • VeilNet grants + multi-hop path grants + cover scheduler
  • WAN TCP multi-hop AEAD lab + multi-process cover daemon + standing process-per-host service lab (OPEN/TICK/REVOKE; grant survives reconnect)
  • Timeline engine + twin shell surface
  • Verification ladder + Verus-shaped invariant document
  • Stub compositor / minifb GUI shell
  • Guest-directed AEAD hop peel (D308-D327): host ChaCha opens each layer; kernel consume waits until peel notes match the hop count. ChaCha is not in the kernel image.
QEMU serial markers (subset)
Liminal freestanding boot OK
ring3+VSpace isolation OK
virtio-rng modern entropy OK
virtio-console multiport CTRL+TX OK
virtio-blk modern WRITE+READ OK
virtio-blk modern MULTI+FLUSH OK
ring3 IRQ steal OK
Aether: fs mount OK
Aether: fs write OK
Aether: fs read OK
Aether: elf load OK
Aether: elf exit=42
Aether: virtio-net modern TX OK
Aether: virtio-net modern RX OK
Aether: net app OK
Aether: grant-tx deny OK
Aether: grant-tx allow OK
Aether: grant-tx revoke OK
Aether: grant-tx sketch OK
Aether: verus lemma A1 discharged OK
Aether: verus lemma OK
Aether: multiport port1 data OK
Aether: console port1 OK
Aether: timeline cross-device OK
Aether: timeline grant hygiene OK
Aether: grant-tx cover quiet OK
Aether: grant-tx cover live OK
Aether: grant-tx cover OK
Aether: fabric-net deny OK
Aether: fabric-net allow OK
Aether: fabric-net OK
Aether: shell ready
Aether: shell session OK
boot OK

Where D328 starts

D308-D327 closed guest-directed hop peel in the host crate. D328, kernel-resident AEAD, is the next lab gate. It is not started and not green. The nav pill stays D327 | next D328.

Not started Host ChaCha Kernel counts notes
Host crate

ChaCha opens each layer here. That is the lab-green part of D308-D327. It is not the kernel image.

Wire

Each hop is a layer. A 3-hop path needs 3 peels before the payload is the payload. The outer layer is not the message.

Kernel notes

The kernel refuses consume until the peel count matches the hop count. Replay, revoke, a short wire, and an early consume still fail closed.

Kernel image

ChaCha is not in the no_std image. Putting AEAD there would be D328. That work has not started.

Closed peel versus the gate that has not started
Question D308-D327 D328
Where the AEAD runs Host crate Would have to be inside the kernel image. Not started.
What the kernel does Counts peel notes, then allows consume No kernel-resident AEAD yet
Status Lab green, 2026-09-21 Not started. Not green.

Still refused

This section does not claim multi-machine WAN, Verus Tier 4, a product appliance, or an ambition move. The bar stays about 35 percent of an 18 month personal-class estimate.

Fail-closed cases the peel lab already showed

Tamper, wrong key, truncated wire, empty path, oversize and 9-hop reject, replay, revoke, and consume before the note count matches. Those results belong to D308-D327. They are not D328.

The numbered next-gate line stays on the roadmap.

What is coming

True remaining cliffs after the modern MULTI+FLUSH / CLIF / multi-hop / Verus-hooks wave. The next numbered lab gate is D328 (kernel-resident AEAD). N1-N4 stay open surfaces and are not the next Continuum number. Mid is real architecture surface still under-sold by a flat four-item list. Never is deferred by design.

Proof surface: QEMU serial Proof surface: host lab Scaffold / design Deferred / never without ask

Planning horizon (~18 months) is at the top of the page, just under the hero. Jump to horizon - ambition bar, not a gate counter. Live data: /horizon.json.

Near (open surfaces)

Next

D328 kernel-resident AEAD

ChaCha still lives in the host crate. D328 is the next lab gate: AEAD inside the no_std kernel image. Not started. The seam section on this page states the boundary. Not multi-machine WAN. Not Verus Tier 4.

design / next lab gate

N1

cranelift-jit feature

Optional cranelift-jit feature wires real Cranelift (0.120). Quad parity lab green on pure arith demos. Default build stays light without the feature. Remaining: broader ops and product loaders.

host lab / feature on P2 universal execution

N2

Multi-host WAN standing service

Standing process-per-host cover service lab is green (shared grant across reconnect; two-host localhost). Remaining: production durability, supervision, routable WAN beyond lab. Not a consumer product.

host standing lab green P1 privacy | P5 swarm

N3

Verus toolchain port

Invariants documented; hooks scaffold. Next is real proof fn discharge on a Verus toolchain.

scaffold (T3) P3 formal foundations

N4

Virtio-net + port1 data

Virtio-net modern first path is lab green under QEMU (DRIVER_OK + TX + RX via DHCP Discover / SLIRP Offer). Remaining: multiport port1 data queues, IRQ-driven RX, multi-buffer. Not a guest Fabric TX stack.

QEMU net TX+RX lab green P1 no ambient NIC

Mid (architecture gaps, not polish)

These are real remaining surfaces already admitted by layer tags and lab boundaries. They are not secret green work and not claimed finished.

M1

Kernel-integrated VeilNet TX

Host Phase0 + multi-hop labs are green. Grant-bound egress inside the guest kernel / NIC path is not claimed.

design boundary P1 ontological privacy

M2

Kani runnable CI (T2)

Kani-shaped harness intent exists as scaffold. Runnable green Kani CI is not claimed.

scaffold P3 formal ladder

M3

L1 agent swarm crates

Design + early host stubs only. Not a shipping local AI operator product.

design + stubs P4 | P7

M4

L4 compositor beyond stub

minifb / stub compositor only. Living UI weaver is not a shipped product surface.

stub P4 generative continuity

M5

Personality / FS grant-only depth

IR + JIT live on host. Full multi-personality and HostSys FS grant demos remain deeper cliffs.

partial host P2 universal execution

M6

Cross-device timeline durability

Host fork / discard / merge paths are green. Cross-device timeline migrate is later phase, not claimed.

host paths only P5 | P6 | P7

Explicitly deferred / never without ask

...

Hard non-goals

Full Cocoa/Win32 as core UI | seL4-complete Isabelle | neuromorphic / BCI | bare metal without an explicit ask.

deferred by design

FAQ

What is AetherOS?

A ground-up capability-first personal OS research project: Liminal microkernel, VeilNet, AetherIR fabric, personal swarm and timeline continuity.

Is it open source?

No. Proprietary private R&D. This hub tracks vision and status; it is not a public source tree.

Is it a Linux distro?

No. Freestanding Rust no_std kernel plus host crates - not a distro or kernel fork.

What is VeilNet?

Deny-by-default network fabric. Packets require a live EgressGrant; cover is grant-bound so idle stays silent.

How is AetherOS layered?

Seven layers: capability microkernel through personal swarm and temporal continuity. Upper layers never invent ambient authority.

Can I run it?

Lab builds target QEMU/UEFI on a development host with Rust + QEMU. There is no public consumer installer yet.

How is an external packet born?

Only under a live EgressGrant (or composed path grant). A domain with no grant has no ambient egress path. See Grant continuity.

What happens to cover when a grant dies?

Cover is bound to grant lifetime. On REVOKE or expiry, cover ceases. Idle domains stay quiet by construction on host fabric lab paths (D3). See Cover construction and adversary ledger.

What does timeline discard do to grants?

Discard drops the experimental branch and must revoke experimental egress so no residual packet birth remains. Merge does not auto-import net grants without an explicit human gate (D5 host engine paths).

Can the digital twin phone home by default?

No. Cognitive sovereignty: the encrypted twin leaves the personal swarm only under an explicit grant. There is no ambient twin egress channel.

What is actually proven today?

QEMU serial markers and cargo tests (Tier 0-1 Active). Continuum D1-D327 are lab-green observations, not machine-checked full proofs. D1-D5 are what feed that T0-T1 floor.

Does Scaffold mean almost proven?

No. Scaffold means hooks, shapes, or design docs exist. It is not running proof CI and not discharged Verus or Kani suites.

Is AetherOS seL4-proven?

No. seL4-style capabilities are design lineage. Tier 5 is design-compatible only. Isabelle-complete seL4 is explicitly deferred.

Can outsiders audit formal proofs?

No public source tree. The monorepo is proprietary R&D. This hub publishes calibrated status, not proof artifacts.

Is AetherIR Wine or Rosetta?

No. Foreign binaries are intended to enter a capability host via AetherIR, not reimplement a foreign OS as architecture. See Runtime Fabric Codex.

What is proven on the runtime path today?

Host lab D2: hand JIT + CLIF interpret + CLIF native emit with parity demos. Full multi-format loaders and multi-personality HostSys are not claimed complete.

Is cranelift-jit shipping?

Not as a default dependency. Optional feature cranelift-jit wires real Cranelift with lab parity tests. Hand JIT + CLIF paths remain the default light stack.

Does cover defeat traffic analysis?

No free GPA defeat on public low-latency edges. Cover is grant-bound construction on host fabric paths: idle quiet when grants die. See the adversary observation ledger.

What is the next lab gate?

D328, kernel-resident AEAD. Lab depth is D1-D327. Ambition stays about 35 percent of an 18 month personal-class estimate. That is not a finished product.