Skip to content

GROK-ASCENT-STARLINK-ARCHITECTURE.md

Latest commit

 

History

History
464 lines (341 loc) · 20.9 KB

GROK-ASCENT-STARLINK-ARCHITECTURE.md

File metadata and controls

464 lines (341 loc) · 20.9 KB
Raw

Grok over ASCENT over Starlink

Title: Resilient terminal Grok via ASCENT on Starlink (and beyond)
Status: Research synthesis + implementable architecture (v0.2, SkyPulse)
Date: 2026-08-12
Scope: Public evidence only. Legal stock Starlink kit. No RF hacks.
Product hooks: Pitchfork-and-Torch/ascent SPEC 1.0.0-rc1, ascent-wire 2.1.0, SkyPulse PATHHINT, live Wire Lab 3.0 Crystal Wire.

ASCII hyphens only in this document.

Executive truth (read first)

Three layers people collapse into one slogan:

Layer What it actually is
Starlink Last-mile IP (dish <-> LEO <-> optional ISL mesh <-> ground gateway <-> PoP <-> public Internet)
ASCENT Wire encoding + agent session grammar + optional ASCENT-D outer ECC. Not a network.
Grok Frontier-scale cloud inference at xAI (api.x.ai / Colossus-class terrestrial GPUs)

Achievable today: When your fiber/cell dies but the public Internet and xAI still live, Starlink restores the pipe; ASCENT structures flaky sessions; a local model keeps a brain when the API is gone.

Not achievable as a consumer product today: Full cloud Grok while "worldwide terrestrial Internet is completely unavailable." That scenario almost certainly kills gateways, PoPs, and xAI DCs. ISLs do not host Grok.

One sentence: Starlink keeps you on the public Internet when your last mile dies; ASCENT keeps sessions and payloads honest when the link is ugly; SkyPulse PATHHINT gives CCA/apps a fail-closed freeze/capacity seed so LEO flaps waste fewer retransmits; only local models keep a brain when the public Internet (and therefore xAI) is actually gone.

SkyPulse does not make Starlink faster at RF. It improves usable session bandwidth and session continuity. Not a bare "bandwidth upgrade."

0. SkyPulse (PATHHINT) on Starlink IP

Priority for this architecture: usable session bandwidth, not PHY Mbps slogans.

Piece Role
PATHHINT / SKYSTATE 0xC5 schema 0x01 Compact hint: path_id, next_capacity (predicted sender bottleneck bits/s, not RF PHY), freeze_until, confidence, ttl; optional obstruction/elev
Fail-closed CRC/short-RS fail, stale TTL, unknown version => erase; never apply a bad hint to CCA
ASCENT-E-LEO Usage profile of E: light CRC or short RS(255,239) I=1 (not stacked); no full P9 RS on interactive IP
ASCENT-D Keep for spool / deep-space / high-BER; RS(255,223) erase-on-fail
QUEUE Daemon mode when obstruction or RTT flaps; CLOUD/EDGE still name the brain honestly

LeoAware (OrbitStack) hybrid fuse: PATHHINT may seed freeze + next_capacity; Orb pathID still owns REPROBE when not suppressed; assist suppress ~2s / no hybrid util-MD; predictive pre-hop freeze is growth freeze only and must never gate ep:loss_burst. Absolute dual-gate remains a paid OrbitStack product fence. See docs/SKYPULSE.md and docs/ORBITSTACK-LEOAWARE-BRIDGE.md.

Canonical lab unit (50 Mbps is a hint, not a dish measurement):

c5 01 00 1a 0e 00 00 00 00 00 00 00 42 02 fa f0 80 ...

Python: encode_pathhint / canonical_pathhint_bytes in ref/ascent_codec.py.

1. Research dossier

1.1 ASCENT mapping (from SPEC 1.0.0-rc1 + local codec)

Plane Role Satellite / LLM use
P0 Sacred ASCII byte < 0x80 forever Greppable logs, prompts, shell-safe transcripts over any path
P1 Fixed control fences Stream structure survives partial readers
P5 Agent frames 0x9A..0x9B + 0xC1 opcodes ROLE, TOOL, THINK, HANDOFF, CAP, SAFETY, STOP for Grok turns
P6 Multimodal 0x9D 0x4D Prefer REF over INLINE on flaky links; hash-bound assets
P7 Quantum-safe control 0x9C Session keys/kids; keys stay OOB
P8 DEF self-maps Profile negotiation, ECC table, freeze mid-cruise
P9 / ASCENT-D Outer sync D5 E5 C0 DE + RS(255,223) GF(256) Erase-on-parity-fail; no mojibake agents; I=8 default

Profiles: ASCENT-7 (identity ASCII) | ASCENT-E (Earth) | ASCENT-E-LEO (usage of E: PATHHINT, light integrity) | ASCENT-D (deep outer) | ASCENT-A (archive).

Fail policy (D): parity/CRC fail => erase unit. Never soft-fill corrupt TOOL/SAFETY.

Sources: local SPEC.md C.1-C.5, ref/ascent_d.py, docs/AGENT-LOOP.md. Confidence: High (normative product).

1.2 Starlink / SpaceX state of the art (public, mid-2026)

Claim Confidence Implication
User path needs some Earth gateway + PoP for public Internet apps High No pure-space Grok for customers
Optical ISLs form a mesh; can route to distant gateways High Helps remote/ocean coverage; still needs exit
Typical land RTT ~25-60 ms; residential tens-hundreds Mbps down, ~10-40 Mbps up High Fine for chat tokens; budget uplink for large prompts
Dish local gRPC 192.168.100.1:9200 (community RE; status/history) High Diagnostics only; not official residential API
Enterprise Management/Telemetry APIs exist for fleets High Ops, not inference
Direct-to-Cell: SMS/select apps; not full HTTPS agent sessions High Emergency C2 only
V3 / orbital AI / Marslink: roadmap or early deploy, not customer Grok path Med Phase 3+ speculation
Power: Standard dish tens of watts average; needs sky + account High UPS is part of the system

Primary anchors: starlink.com/technology, starlink V3 updates, performance legal docs, RIPE Labs Starlink measurements, sparky8512/starlink-grpc-tools, T-Mobile T-Satellite, xAI api.x.ai.

1.3 DTN comparison

Dimension DTN (BPv7 + CLA + optional LTP) ASCENT-D
Problem Intermittent multi-hop store-forward, long delay Frame integrity + erase-on-fail for agent wire
Layer Overlay network Application / codec
Starlink chat default Poor (IP already exists) Good with QUIC/TCP
Deep space later Required for non-IP contacts Complementary E2E integrity

Stacking decision:

  • Today Starlink IP: ASCENT (+ optional ASCENT-D) over QUIC/TCP/HTTPS. No BP daemon.
  • True deep space later: ASCENT as bundle ADU (A) + ASCENT-D E2E (C) + LTP red-part / optional ECLSA + CCSDS link coding.
  • Reject: ASCENT as a custom convergence-layer codec (breaks ION/HDTN/uD3TN interop).

FEC note: RS(255,223) rate ~0.875 matches classical CCSDS outer RS (t=16 errors / up to 32 erasures). ASCENT-D does not replace TM LDPC/turbo or LTP ARQ. On Starlink IP, do not stack P9 RS on top of PHY+TLS (double-FEC tax). Soft-decision LDPC/turbo beat hard RS on AWGN; RS + erase-on-fail wins for fail-closed agent safety on short spool / deep-space ADUs. PATHHINT optional IEEE CRC-32 is integrity, not FEC.

Sources: RFC 9171 BPv7, RFC 5326 LTP, CCSDS 131.0-B TM coding, ION/HDTN/uD3TN public trees. Confidence: High for standards; Medium for ECLSA packaging.

1.4 Interfaces usable today (legal)

Interface Use
https://api.x.ai/v1 (Bearer) Cloud Grok
Normal IP over Starlink WAN Transport
Dish gRPC :9200 status/history (community tools) Link health gating
Local Ollama / edge OpenAI-compatible Offline brain
Enterprise Starlink APIs Multi-terminal ops (account-gated)

Do not invent: residential dish REST for routing; space-side Grok endpoint; "secret Musk mesh" SLA.

2. End-to-end architecture

+------------------+     +-------------------+     +------------------+
| Operator TUI     |     | ascent-satd       |     | Local edge LLM   |
| (readline/CLI)   | <-> | session + queue   | <-> | (Ollama etc.)    |
+------------------+     | mode: cloud|edge  |     +------------------+
                         | sacred-meter      |
                         +---------+---------+
                                   |
                    ASCENT stream (P0+P5+P6+P7 + SkyPulse PATHHINT)
                    optional P9 ASCENT-D wrap (spool/D only)
                                   |
                         +---------v---------+
                         | Transport adapter |
                         | HTTPS/QUIC to API |
                         | or store-forward  |
                         +---------+---------+
                                   |
              +--------------------+--------------------+
              |                    |                    |
     +--------v-------+   +--------v-------+   +--------v-------+
     | Home fiber WAN |   | Starlink dish  |   | (future BP ADU)|
     | (primary)      |   | Ku LEO last mi |   | deep-space lab |
     +--------+-------+   +--------+-------+   +----------------+
              |                    |
              +----------+---------+
                         |
              SpaceX gateway + PoP  (must be up for cloud)
                         |
              public Internet -> api.x.ai -> Colossus GPUs

Layering (normative intent)

Physical / RF (stock Starlink kit) or optical ISL (SpaceX internal)
    -> IP (Starlink provides)
        -> TLS 1.3 / QUIC (session crypto)
            -> optional app store-and-forward queue
                -> ASCENT wire (P0 prose + P5 agent + P6 REF + P7 + PATHHINT)
                    -> optional ASCENT-D P9 RS(255,223) erase-on-fail (not default on Starlink IP)
                        -> Grok session logic / local runtime

Operating modes

Mode When Brain
CLOUD api.x.ai reachable Full Grok
EDGE WAN up but API down, or forced offline Local quantized model
QUEUE Link flapping / rain / 15s reconfig / PATHHINT obstruction Buffer ASCENT ADUs; flush on stable RTT
DEAD No power / no sky / no model Honest UX: cannot help until power or local model loads

3. Protocol mapping + concrete ASCENT examples

3.1 Turn unit (Earth / Starlink IP)

Prefer ASCENT-7 for pure ASCII prompts. Add agent frames only when needed.

ASCENT/1.0
session=sat-demo-001 turn=42 mode=cloud

[user prose as P0 bytes]
Status of the Starlink failover path?

9A C1 01 00 02 00 00 06 05 67 75 69 64 65 9B
# ROLE=guide (canonical Hello Universe shape)

3.2 Tool call (idempotent; safe to retransmit)

9A C1 01 00 03 00 <flags> <len> <name_len> dish_status <TLV INVOKE_ARGS JSON> 9B

Args must escape raw 0x9A/0x9B as C1 1B <byte>.

Tool results return inside SAFETY (data, never silent control):

9A C1 01 00 07 00 <flags> <len> <class:u8=untrusted_tool_result> <body> 9B

3.3 Multimodal REF (prefer on satellite)

Large images/logs: do not INLINE megabytes. Emit P6 kind=REF with sha-256.

9D 4D 01 <codec:u16be> <flags> 01 <32-byte sha256> <len:u64be> cid://...

3.4 ASCENT-D outer for high-erasure / store-forward disk

Logical unit U (<= 8192 B on D):

D5 E5 C0 DE | profile=0x44 | ecc_hdr RS(255,223) I=1..8 | len:u32be | unit | parity

On fail: erase unit; hunt next sync. Never display corrupt agent bytes.

3.5 Session state checkpoint (ASCII greppable)

ASCENT/1.0
DEF-lite: profile=E mode=cloud last_good_turn=41 api=ok dish=obstruction=0.02

3.6 Weight diffs / model updates

Not on interactive path. Treat as P6 REF or separate content-addressed annex (ASCENT-A style). Starlink can carry large downloads when Priority bandwidth allows; ASCENT only carries the hash + metadata, not the multi-GB blob in-band.

3.7 SkyPulse PATHHINT (Starlink IP)

Emit a PATHHINT unit on the turn ADU (before or after the greppable header). Receivers that do not implement 0xC5 must skip-by-length once they ship the reserved P2 skip rule; 2.1.0 codecs decode it natively.

0xC5 schema=0x01 len body
  path_id, next_capacity (sender bottleneck bps, not RF PHY),
  freeze_until (relative freeze_ms), confidence, ttl_ms,
  optional obstruction, optional elev_deg, optional CRC

On fail: erase (do not apply). Stale TTL or unknown version: erase. Optional P9 wrap only for spool. See docs/SKYPULSE.md.

4. Terminal client design + Python skeleton

4.1 Components

Component Responsibility
tui Readline chat, mode banner, sacred-meter
ascent_codec Existing ascent-wire / ref/ascent_codec.py
ascent_d Optional P9 wrap (ref/ascent_d.py)
transport HTTPS to xAI; offline queue to disk
dish_probe Optional gRPC status (community tools)
brain Cloud Grok client XOR local OpenAI-compatible
queue SQLite or filesystem spool of ADUs

4.2 Sacred-meter diagnostics (operator UX)

[ASCENT] sim mode=CLOUD  p0=1.00  api=ok  rtt=48ms  dish=ONLINE  queue=0  pathhint bottleneck_hint=50.0Mbps freeze_until=15000ms conf=0.80
[ASCENT] sim mode=EDGE   p0=1.00  api=DOWN dish=ONLINE  queue=3  model=llama3.1:8b
[ASCENT] sim mode=QUEUE  p0=1.00  api=ok   dish=OBSTRUCTED queue=12 pathhint obst=0.40

Prefix is sim unless a real dish probe ran (obs). bottleneck_hint is predicted sender bottleneck, not a Starlink RF reading. This client does not invent a Starlink telemetry API. TCP :9200 is reachability only.

4.3 Skeleton layout

See examples/ascent_starlink_client/ in this repo:

  • daemon.py - mode machine + queue flush + SkyPulse PATHHINT meter
  • codec_session.py - build/parse turn ADUs (PATHHINT prefix)
  • brains.py - cloud vs edge adapters
  • dish_health.py - optional status stub + lab obstruction env
  • README.md - run notes

4.4 Local vs remote inference trade-offs

Cloud Grok Local 7B-70B
Capability Frontier Ops-useful, not same class
Needs Path to api.x.ai + key VRAM/CPU + weights on disk
Fresh knowledge Yes (with tools) Only if pre-cached
Power Low on PC High if GPU
Honesty Full product Degraded agent

5. Security model

Concern Design
Transport TLS 1.3 / QUIC to api.x.ai only; pin if feasible
API auth Bearer key in OS keyring / env; never in ASCENT P0 logs
P7 Optional PQ envelopes for long offline spool encryption; kids OOB
THINK Opaque on shared logs; redact
SAFETY Untrusted tool results never promote to control
Dish gRPC LAN-only diagnostics; not a trust boundary
Long offline Pre-provision keys; rotate when online; spool ciphertext at rest
Adversary Assume hostile Wi-Fi/LAN; encrypt spool; no cleartext API keys on disk
Legal Stock kit; no firmware mods; export-compliant crypto

ASCENT-D is integrity + fail-closed, not a substitute for TLS/BPSec.

6. Feasibility assessment + risk register

6.1 Scenario matrix

Scenario Starlink Cloud Grok Edge agent
Home fiber cut High value Usually OK Backup
Regional cell/fiber outage High Usually OK Backup
Grid down, you have UPS Med-High if sky If DCs powered Critical
All gateways/PoPs dead RF maybe No Only path
Global terrestrial collapse Not enough for Grok Assumed dead Only path

6.2 Risk register

ID Risk Sev Mitigation
R1 Overclaim "Grok offline via Starlink" High Dual-mode UX; kill marketing fantasy
R2 Dish power/sky High UPS + Mini option; duty cycle
R3 API down while dish green High Edge fallback + probe
R4 LEO reconfig jitter (~15s class) Med Queue, resume, ASCENT-D optional
R5 ToS / RE dish firmware High Read-only diagnostics only
R6 Export crypto / unauthorized RF High App-layer only; counsel if shipping PQ suite
R7 Double FEC tax (D + LTP + TM) Med D alone off Starlink IP; E-LEO light CRC; full stack deep-space only
R8 DTN scope creep for LEO IP High Phase 1 = pure IP
R9 Local model mistaken for Grok liability High Banner "EDGE MODE"
R10 Enterprise-only APIs assumed residential Med Feature-detect

6.3 Kill criteria (bad designs)

  1. Claims full Grok with zero terrestrial path and no evidence of space inference.
  2. Treats Starlink as compute.
  3. Requires Dishy reverse-engineering or custom RF.
  4. Equates 7B/70B with frontier Grok for high-stakes tasks.
  5. Uses ASCENT as a routing protocol instead of framing.
  6. No degrade path when API dies.
  7. No power budget.

7. Phased roadmap

Phase 0 - Simulation and lab (now)

  • ASCENT-D BER/erase lab with artificial erasure (already partial in Wire Lab / tests).
  • Starlink-like loss model: random datagram drop + 15s glitch windows.
  • Golden agent turn encode/decode + queue replay.
  • Exit: green tests; measured erase rate vs simulated BER.

Phase 1 - Near-term prototype (works on stock kit)

  • CLI/daemon: dual-mode cloud Grok + local Ollama.
  • ASCENT framing for turns (P0 + P5).
  • Failover WAN (fiber primary, Starlink secondary) optional at router.
  • Dish status optional via community tools when a real dish exists. This repo does not invent a Starlink telemetry API.
  • Store-and-forward spool when API/RTT bad.
  • Exit: survive home ISP outage with cloud Grok still answering; survive API outage with edge mode.

Phase 2 - DTN-enhanced hybrid + stronger edge

  • Optional BPv7 lab (uD3TN AAP or HDTN) carrying ASCENT ADUs for delayed peers.
  • Stronger edge models; RAG over local continuity.
  • Optional ASCENT-D on spool files.
  • Exit: multi-hour delayed round-trip demo without lying about interactive Grok.

Phase 3 - Deep-space ready / partnership

  • Requires SpaceX and/or xAI cooperation, or mission hardware: contact plans, LTP, CCSDS coding, possible enterprise dual-home of inference.
  • Lunar/Mars class delays: one ADU per RTT budget; no chatty tool loops.
  • Exit: only with external partners or flight program - not a garage-only claim.

8. Open questions, experiments, today vs later

Achievable today (no partnership)

  • ASCENT codec + agent frames on any IP path
  • ASCENT-D encode/decode erase-on-fail in software
  • Starlink as failover WAN for api.x.ai
  • Local edge LLM when API unreachable
  • App-level queue across link flaps
  • SkyPulse PATHHINT encode/decode + fail-closed skip
  • ASCENT-E-LEO vs D integrity helper (no double-FEC on Starlink IP)
  • QUEUE on obstruction / RTT flap in the Phase 1 client

Requires partnership or new public capability

  • Space-resident Grok inference
  • Documented sat-only path that never touches public Internet but still reaches xAI
  • Official residential dish developer control plane
  • Mission-class BP/LTP over Starlink as a product

Recommended experiments

  1. Failover drill: cut fiber; confirm Grok chat via Starlink; log RTT and failures.
  2. API kill drill: block api.x.ai; confirm EDGE banner + local answers.
  3. Glitch inject: drop 1-5% UDP or reset TCP mid-stream; measure queue recovery.
  4. ASCENT-D spool: corrupt random bytes on disk; confirm erase, never bad TOOL.
  5. Power budget: measure Watts for Mini vs Standard + PC for 2h and 8h UPS.
  6. DTN lab only: one delayed ADU via uD3TN loopback; do not put production chat on BP yet.

Source and confidence index (abbreviated)

Topic Confidence Notes
ASCENT SPEC planes / P9 RS High Product normative
Starlink needs ground exit for public APIs High Architecture consensus + measurements
ISLs extend gateway reach High SpaceX marketing + RIPE-style studies
Dish :9200 gRPC community tools High sparky8512/starlink-grpc-tools
BPv7 / LTP standards High RFC 9171 / 5326
ECLSA production packaging Medium Strong papers; verify fork
V3 / orbital AI timelines Medium Roadmap / early deploy
xAI dual-homed on Starlink for customers Low / none public Do not design on it

Appendix A - Army roles and synthesis

Role Outcome used
ASCENT Protocol Specialist SPEC plane map, stream examples, P9 fail policy
Starlink Systems Researcher Path model, interfaces, DTC, V2/V3, critical egress truth
DTN Expert A/C vs D stacking; FEC comparison; reject ASCENT-as-CLA
Terminal Designer Mode machine, sacred-meter, Python skeleton
Security TLS + P7 + SAFETY + spool encryption posture
Feasibility Critic Kill criteria, scenario matrix, honesty bar
Architecture Synthesizer This document + roadmap

Critique loop applied: every "space Grok forever" claim was reduced to dual-mode resilience that works on stock hardware.

Appendix B - Related paths

  • Spec: SPEC.md
  • SkyPulse: docs/SKYPULSE.md
  • OrbitStack / LeoAware: docs/ORBITSTACK-LEOAWARE-BRIDGE.md
  • Agent loop: docs/AGENT-LOOP.md
  • Codec: ref/ascent_codec.py, ref/ascent_d.py, ref/ascent_skypulse.py
  • Client skeleton: examples/ascent_starlink_client/
  • Skill: ~/.grok/skills/ascent-starlink-resilience/
  • DTN stacking skill: ~/.grok/skills/ascent-dtn-space-stacking/
  • Live lab: https://ascent.jonbailey.xyz/
  • Public repo: https://github.com/Pitchfork-and-Torch/ascent