forgo.cloud
Sign in
Repo workspace

forkjoin-ai/gnosis

Teleport limitations (read this before you believe the hype)

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

Teleport limitations (read this before you believe the hype)

This work is genuinely cool and the proofs are real. It is also bounded in specific, important ways. This file is the honest ledger. If a claim is not here and not in a certificate, treat it as unproven.

What it is NOT

  • It is not faster-than-light. Causality is intact. geodesicLength:0 means the cached result does not traverse compute space — it was precomputed and is replayed locally. The cache key still travels at the speed of the channel. The win is avoiding the round trip, not beating physics.
  • "Survives Mars" is an Eb/N0-equivalence, not a distance claim. We cannot radiate to Mars with bench gear (see hardware below). The link-budget physics is real; the bench reproduces the Eb/N0 a given distance/dish/power would deliver and shows the FEC closes it. Every certificate states this substitution.
  • It is not a model-accuracy or end-to-end-throughput claim. The benchmarks are component figures (RF SNR, FEC A/B, link margin, replay latency). They do not claim tokens/sec for a full model under load.

The core functional limit: a hit requires a precomputed volume

The teleport serves a cached result. The flow only pays off on a cache hit:

  • A miss (a prefix the cache has never frozen) falls back to the full slow path — at Mars distance that is the entire round trip. Teleport amortizes repeated/known work; it does not conjure new compute.
  • Keying is exact-prefix: (prefix_hash, prefix_len, layer_lo, layer_hi). Any change to the token stream changes the hash and forces a miss. This is the soundness guarantee (you cannot replay the wrong volume), but it means near-misses do not hit.
  • The cache is small by design (DEFAULT_MAX_VOLUMES = 8, linear scan) with liquid-memory eviction. It is a hot-prefix accelerator, not a general result store.

Witness limits (Skymesh)

  • The spectral witness needs a strong, anchored carrier. Measured here: a random/empty FM slot gives noise (2/6 shared peaks); the 433 MHz ISM band was dead (rtl_433 decoded nothing in 12 s). It works when anchored on a genuinely strong carrier (102.49 MHz local FM here); scan for one per venue (rtl_power).
  • Gain invariance was a real bug, now fixed. Two physically different receivers differ in absolute energy (~2 log2 units here). The witness keys on the gain-invariant top-band SET, not absolute levels. Coarse (16) bands under-discriminate; fine (128) split the carrier; 64 is the measured sweet spot for these radios — it is not guaranteed optimal elsewhere.
  • A one-band witness perturbation is tolerated by design (resilience). Therefore a witness bit-flip is NOT the tamper test; tamper is the protocol69 forged-projection gate. Do not mistake witness agreement for the cryptographic gate — they are separate.
  • Witness entropy is modest. Top-3 of 64 bands distinguishes "the chosen carrier present" from "absent/different" well; it is not a high-entropy identifier. A decoded field (RDS station id, ADS-B ICAO) would raise entropy and legibility — not yet wired.

Hardware limits (local)

  • RTL-SDRs are receive-only. No transmit. The only TX is the ADALM-Pluto, and TX is operator-gated. The Skymesh demo is RX-only by design (no TX license needed) — a feature, but it means the "broadcast" in that demo is the ambient sky, not us.
  • The Pluto tops out at 6 GHz. It cannot reach the real Mars DSN bands (X-band 7.1/8.4 GHz, Ka 34 GHz). The Mars-channel work runs at 433.92 MHz / 2.4 MS/s as an Eb/N0 stand-in.
  • One process per RTL-SDR device index at a time. The two dongles are independent (no contention between them), but you cannot run two captures on the same index simultaneously.
  • libiio is a from-source build under /tmp (ephemeral). It must be rebuilt after a reboot; pluto-tx.ts fails closed with a clear message if the framework is missing.
  • Sample rate: the Pluto minimum (~2.083 MS/s) forces the modem to 2.4 MS/s; do not assume 2.048 MS/s end to end.
  • R1 RTL-SDR over USB-OTG is not qualified. On-R1 reception is aspirational; the proven path uses a host-attached dongle. The R1s are proven as /chain mesh nodes, not as SDR front-ends.

Implementation gaps (documented, not hidden)

  • Replay returns the frozen volume, not yet a rendered token end-to-end. The demo surfaces a cached answer ("Paris") via the orchestrator; wiring the replayed tail through the lm-head to a live token on every surface is a documented next step.
  • The BWAH capture wire carries tail_residual but not kv_slab. The capture route stores an empty kv_slab; the geodesicLength:0 verdict depends on the cache HIT, not kv contents, so this is correct for the teleport flow but the wire does not yet transport a full KV slab.
  • monster-guard / monster-swarm use a minimal inline protocol69 check, not the full primitive. They are lean crates that hold no amplituhedron cache, so a valid envelope is a local miss and they fall through. Only fat-station / moonshine / gnosis-uring carry a real cache + replay.
  • gnosis-uring uses a local protocol69 mirror, not the distributed_inference crate dep (to avoid dragging the whole inference stack into a lean transport). The wire bytes are 1:1 with the canonical contract; keep them in sync by hand.
  • aether wasm-simd teleport gate is opt-in. Default kernel behavior is unchanged until a caller binds a teleport cache.
  • The capture side of the cache is not yet wired into the mesh commit path in gnosis-uring (volumes are seeded via capture_teleport_volume); auto-freezing on heavy-compute completion is the natural follow-up.

Scope / proof discipline

  • Lean modules (SkymeshTeleport.lean, MarsDistanceLadder.lean) prove the admission algebra and the ladder monotonicity — finite arithmetic and logic, Init-only, no omega/Mathlib. They do not prove anything about the physical channel, the FFT, or the radios; that truth lives in the measured certificates, deliberately outside the kernel (the bridge rule).
  • The distance ladder uses representative planet distances and a single reference dish/power. Real missions vary widely; treat the ladder as a shape (monotone, with a break rung), not a mission plan.

If you find a claim in the prose that is not backed by a certificate, a passing test, or a Lean theorem, it is a bug in the docs — file it. The point of this project is the opposite of Sardis: the live witness under lake build / vitest / a real capture, not the label.