forgo.cloud
Sign in
Repo workspace

forkjoin-ai/gnosis

Mars delay-tolerant cache-key teleport

distributed-inference/docs/MARS_DELAY_TOLERANT_TELEPORT.md
forkjoin-ai/gnosis
Repo overview Open in workbench

Mars delay-tolerant cache-key teleport

Sibling to R1_RESILIENT_MESH.md and SKYMESH_DEMO.md. Where R1_RESILIENT_MESH relays a residual through /chain over in-room WiFi, this relays a cache key across an interplanetary light-time gap so a remote station replays cached compute instead of paying the round-trip. SKYMESH_DEMO removes the link entirely: two air-gapped receive-only stations derive the same sky-borne witness and replay the same cached answer in lockstep.

The idea

A "Mars station" is a mesh peer separated from Earth by a one-way light-time delay (opposition ~3 min, mean ~12.5 min, conjunction ~22 min). Round-tripping an inference request is catastrophic at that latency. Instead the cache key is broadcast (protocol69: broadcast symbol 66 XOR local FOIL operand 7 = 69). If the projection validates, the station does a geodesicLength:0 replay of a frozen amplituhedron volume locally — the key teleports, the compute does not traverse space.

This is the latency-amortization the FOIL loser-policy "vacuum of the future" (lean/Lean/ForkRaceFoldTheorems/FoilLoserPolicy.lean) already formalizes for the local mesh, stretched to interplanetary distance.

What is real vs emulated

  • Real: protocol69 admission (live gnosis-foil-control, fail-closed on drift); the amplituhedron capture/replay (distributed-inference/src/amplituhedron.rs, already wired into the prefill hotpath in model.rs); the light-time physics.
  • Emulated: the Earth<->Mars distance. We never wait real minutes — the modeled one-way delay is applied only to the cold round-trip A/B baseline, scaled by --delay-scale, while the certificate records the true modeled seconds.

Demo

rtlsdr-mock-sim/mars-teleport-demo.mts:

  1. Admit the over-the-air-recovered cache key (broadcastSymbol 66) via the live foil node — and confirm a wrong symbol is rejected (fail-closed teleport).
  2. Local geodesicLength:0 replay (amplituhedron cache-hit class), timed.
  3. Latency A/B per ladder rung: local replay (constant) vs the cold round-trip (modeled light-time).
  4. Multi-hop relay chain (Earth -> Moon -> Mars -> Jupiter store-carry-forward over the @a0n/aeon DTN stack), reporting the round-trip it avoids.
  5. Emit gnosis.mars-teleport.v1 certificate.
# from open-source/gnosis, with gnosis-foil-control running on :8787
node_modules/.bin/tsx rtlsdr-mock-sim/mars-teleport-demo.mts \
  --foil-url http://127.0.0.1:8787/.aeon/protocol69 --recovered-symbol 66 \
  --cert-out .qa-artifacts/rtlsdr/mars-teleport-certificate.json

Measured (2026-05-24)

  • admittedByLiveFoil: true; wrongSymbolRejected: true; projection 66 XOR 7 = 69.
  • geodesicLength 0; local replay ~0.19 ms.
  • Latency A/B (round-trip avoided vs local replay): Moon 13,267x; Mars mean (1501 s round trip) 7,765,652x; Voyager 1 (164,114 s) ~849,000,000x.
  • Relay chain Earth->Moon->Mars->Jupiter avoids a 6,694 s round trip.

Production exposure (wired)

This is now live in src/bin/fat-station.rs: POST /.aeon/teleport-admit parses the protocol69 envelope, calls the shared teleport::teleport_admit_canonical primitive, and on accept returns a BWAH encode_replay_response (amplituhedron_wire.rs) with X-Geodesic-Length: 0 + X-Replay-Micros — fail-closed 403 on projection drift, 404 on cache miss. POST /.aeon/amplituhedron-capture seeds a volume. Verified end-to-end (200 + geodesic 0 on a hit, 403 on drift, 404 on miss). The same primitive is the admission path across moonshine, monster, gnosis-uring, and the aether wasm-simd edge — see TELEPORT.md.

Remaining next step: auto-freeze a volume on heavy-compute completion (the capture side) and render the replayed tail through the lm-head to a live token on every surface — see TELEPORT_LIMITATIONS.md.