GitHub - ninjahawk/teleporty: Physics-grounded teleportation research

🌀 What Teleporty Is

A little pencil with a swirling green portal on his head, trying to figure out how to move things from one place to another without going through the space in between. We're doing the same thing, but with actual physics.

I have no funding, committee approvals, or vested interest in any particular answer. Just the equations and whatever they say.

🎯 Headline (current state)

End-to-end functional teleportation pipeline demonstrated in simulation for C. elegans and on real dense Drosophila connectome topology. Pool-stim scan-inverse recovers the C. elegans rate-model connectome at Pearson 0.99; the end-to-end pipeline (per-neuron v4 scan, r=0.92) passes all four behaviors under noise, EM error, and deployment stress.

Tests on real FlyWire Drosophila connectome topology first exposed a model limitation — real synapse weights span 3.4 orders of magnitude, and a uniform-parameter rate model can't keep weak and strong synapses both responsive. The fix: a heterogeneous-excitability model (per-neuron gain/threshold, homeostatically calibrated by response-matching). With it, the dense FlyWire whole-connectome subsets — where the uniform model failed catastrophically (behavioral divergence 0.68) — pass the five behavioral tests, but only with probe-doubling (the K·M ≥ 6N coverage rule):

• N=2000: baseline K=400 is a marginal FAIL (stim 1 div=0.0505); doubling to K=800 → all 5 PASS (max div 0.0484), 50/2000 fallback neurons.
• N=5000: K=1500 → all 5 PASS (div 0.028–0.032), Pearson 0.45, 201/5000 neurons on the degree-scaled fallback.
• N=9998 (whole-connectome top-10000): K=2000 → all 5 PASS (div 0.034–0.037), Pearson 0.53, 405/9998 (~4%) fallback. ~2.5 hr single-core (a real finding about compute cost at scale).

This is a real PASS on real connectome topology under a rate model — not a comfortable one, and not a simulation of the live animal. See model-mismatch → resolved, heterogeneous model, and probe-doubling. No physics barriers; the load-bearing unproven claim is the identity assumption (below).

• Per-person info budget: ~247 GB (bulk tissue dominates; brain functional spec ~42 KB — but see the extrapolation caveat below)
• Transmit at 1 Gbps fiber: ~33 minutes
• Fabricate (1 hour at 4 °C, hypothermic): requires 10¹⁰ cells/s and 10⁷ nozzles — no physics barrier, 10⁶× from current SOTA
• Scan trial budget (human): ~4×10⁶ trials with combinatorial driver lines (M=10⁵), ~35 days serial / ~1 hour at 1000× parallelism
• Marginal cost per teleport: ~$10K (illustrative — energy + bio-ink + amortized printer; not derived from a built system)
• R&D timeline: ~20 years, ~billions (illustrative)

The four quantum approaches surveyed (CM tunneling, Cheshire cat, Penrose-Diósi, quantum Darwinism) are all closed with negative verdicts. Direction 1 (classical-information functional teleportation) is the only viable path and is fully wired.

⚠️ Scope — what the simulations actually show

The validated results are simulations of a Wilson-Cowan rate model running on real connectome topology (C. elegans, FlyWire Drosophila). Two things this is not:

• Not a simulation of the live animal. The "behaviors" are stimulus→response patterns of the rate model, not measured worm/fly behavior. The rate model omits spike timing, dendritic computation, neuromodulation, and glial coupling.
• Not a human-scale result. Everything at human scale (42 KB brain, ~247 GB body, trial budget, cost, timeline) is extrapolation from these small organisms, not a demonstrated number.

The one load-bearing unproven assumption is that connectome-level + tissue-level fidelity yields a functionally equivalent person (identity question). It is a working hypothesis, not a result. If it is false, no classical approach works and the project terminates.

💀 Where the Old Approaches Died

These aren't "hard problems we haven't solved yet." These hit fundamental walls — the kind the universe built into itself.

🔬 The Five Alleyways

We went through all of them. Here's what happened.

All five directions worked. Four closed. One demonstrated.

📐 Direction 4: The Quantum Ceiling

From math/direction4_penrose_diosi_threshold.md

The Penrose-Diósi model predicts wavefunction collapse from gravitational self-energy:

τ_PD = ℏ / E_G     where     E_G = 6Gm² / (5R)     (fully separated superposition)

The human body emits ~2.6 × 10²² thermal photons per second at body temperature. Each one collapses a quantum superposition. At room temperature in vacuum, a human-scale quantum superposition cannot exist for even 10⁻²² seconds. No vacuum pump fixes this — the body itself is the thermal source.

🧮 Direction 1: The End-to-End Result

The information budget (revised)

The brain term is negligible. Bulk tissue dominates. Body fits on a consumer SSD; transmission is 1–2 hours over 1 Gbps fiber.

¹ The brain figure is in the middle of an 8-OOM extrapolation (range ~5 KB to ~250 KB depending on which Drosophila d_eff estimate and which scaling method); it doesn't change the total, which bulk tissue dominates either way.

The d_eff scaling fit (three organisms)

* The "~700" is a power-law-extrapolated full-spectrum estimate. Directly measured numbers: top-300 SVD participation ratio = 141 (captures 40.9% of variance), top-3000 SVD = 455 (75.7%). Use 141 and the ordering becomes monotonic (mouse 146 ≈ Drosophila 141); use 700 and Drosophila exceeds mouse.

Fit (using the extrapolated 700): d_eff = 1.85 × N^0.46. This is a 3-point fit with method-mixing across organisms, so the exponent carries real uncertainty (α ≈ 0.39–0.55 depending on which Drosophila estimate and which method).

⚠️ Extrapolating to human (8.6 × 10¹⁰ neurons) is an ~8-order-of-magnitude reach beyond the largest measured point. The result is a range, not a number: d_eff(human) ≈ 2 × 10⁴ – 10⁶, brain spec ≈ ~5 KB (consistent-method α=0.385) to ~250 KB (insect-biased fit). The "42 KB" used elsewhere is in the middle of this range, not a fixed value. See direction1_scaling_law.md, which labels the human extrapolation low confidence.

Scan: pool stimulation (the main result tonight)

Per-neuron stimulation works at C. elegans (r=0.81, 9000 trials) but doesn't scale. Random POOL stimulation (each trial activates M ~ 15 neurons simultaneously) gives higher fidelity at 3× fewer trials, and scales linearly with N.

That last row is the rate-distortion principle in action: many weight configurations produce equivalent behavior. The criterion that matters is behavioral equivalence, not bit-exact weight recovery.

Robustness: pool stim PASSES under

• Rate noise 0% – 5% (Ca²⁺ imaging floor is ~1%)
• EM segmentation errors up to 5% FP + 5% FN
• Scaling tested N = 300 → 1000 synthetic, r ≈ 0.99 throughout
• 5/5 random synthetic networks PASS
• Held-out behaviors (thermo, nociception) reconstruct correctly

Tissue distortion thresholds

Worst-case tissue determines the bit budget per voxel.

Cardiac propagates electrical waves and is most sensitive to coupling heterogeneity. Skeletal muscle is the opposite — parallel-fiber summation averages out per-fiber variance.

The fabricator (the actual bottleneck)

From math/direction1_fabricator.md and direction1_vascular_patency.md

To build a human in 1 hour at 4 °C (within the DHCA viability window):

• Throughput: 10¹⁰ cells/s (10⁴× current bioprinters)
• Print head: 10⁷ nozzles in ~1 m² (existing MEMS inkjet density)
• Resolution: 1 μm (neural), 5 μm (somatic)
• Energy: 175 kWh ($17.50)
• Vascular patency: 8× safety margin under three specs (bubble-free saline, ≥1.5 m reservoir head, ≤4 °C)

No physics barriers. ~10⁶× engineering gap from current SOTA. Manufacturing + economics problem.

How long to send it

For the brain alone: trivial. For the full body: cloud-backup scale, not real-time.

How much do the raw materials cost

A human being costs $42 in raw ingredients. The universe is doing something impressive with that $42.

The identity question (the load-bearing assumption)

Does connectome-level fidelity make a functionally equivalent person? Honestly: probably yes, but it's unresolved — and this is the single assumption the whole project rides on. It is philosophical, not physical, and not experimentally settled beyond C. elegans.

• Neurons replace their molecules every few weeks while identity persists — function > quantum state
• No known mechanism for single-particle quantum states to influence cognition
• Decoherence in warm wet neurons: ~10⁻¹³ s — orders of magnitude faster than any neural computation
• Counter-argument is philosophical (Penrose-Hameroff). Not experimentally supported.

We proceed with connectome-level + tissue-level = functional equivalence as a working hypothesis. If wrong, the project terminates — no classical approach works.

🧪 Simulation Results (current)

End-to-end pipeline (C. elegans, 300 neurons)

From simulation/run_teleportation_pipeline_v2.py

Scan → compress → transmit → reconstruct → verify, at 1% Ca²⁺ imaging noise:

Pearson r on weight matrix = 0.92 (this end-to-end run uses the per-neuron v4 scan; the higher r=0.99 is the separate pool-stim protocol — see the scan table below). Spec size 6.35 KB. Transmit 52 μs @ 1 Gbps. The two held-out behaviors prove the recovered connectome generalizes — it isn't curve-fit to probe-set behaviors. (Reproduced from scratch 2026-05-25: ‖dW‖/‖W‖=0.351, r=0.9157, all 4 behaviors PASS.)

Distortion sweep (foundational, simulation/run_distortion.py)

The brain's 30% distortion tolerance is confirmed within the rate model — i.e. the recovered weights can be off by ±30% and the model's behavioral output still tracks. This is a model-internal result, not a measured property of biological neurons.

Deployment stress test (most honest result)

From simulation/run_deployment_stress.py

Combine all biological deployment constraints simultaneously: cell-type-driver pools + 1% rate noise + 5% FP + 5% FN EM errors. Pure type-only pools FAIL on the tap circuit. Hybrid pools (type drivers + ~50 multi-type combinations) PASS all 4 behaviors. Pearson r=0.43 on weights, but behavioral divergence < 3%. Rate-distortion principle in action.

Tissue D-thresholds

From run_tissue_distortion.py, run_muscle_distortion.py

• Cardiac (2D Aliev-Panfilov, 60×60 sheet): D-threshold = 0.05
• Skeletal muscle (parallel-fiber bundle): D-threshold = 1.0 (20× more tolerant than cardiac)

The CLT damps per-fiber variance in muscle; cardiac propagating waves don't have this advantage.

⚠️ These are toy-model thresholds (a 60×60 Aliev-Panfilov sheet, a parallel-fiber bundle), not measured biological constants. Real cardiac electrophysiology is far more complex; the 0.05 figure is a simulation artifact useful for setting a relative per-tissue bit budget, not a clinical tolerance.

🗺️ Where We Are

Phase 1: Survey existing approaches          ✅ all dead or lead to Direction 1
Phase 2: Work all five alleyways             ✅ 4 closed, 1 demonstrated
Phase 3: Full math for Direction 1           ✅ info budget, energy, scanner
Phase 4: Rate-distortion lower bound         ✅ brain 42 KB, body ~247 GB (7-tissue stratified)
Phase 5: Scanner technology roadmap          ✅ pool stim; ~10⁸ trials @ human scale (coverage-limited)
Phase 6: Reconstruction system design        ✅ fabricator math + vascular patency
Phase 7: C. elegans testable simulation      ✅ d_eff, distortion, R-D
Phase 8: Scan inverse problem solved         ✅ pool stim, r=0.99 at 1% noise
Phase 9: Full end-to-end pipeline test       ✅ PASS at C. elegans + deployment stress
Phase 10: Generalization beyond C. elegans   ✅ 5/5 synthetic, scaling N=300→1000
Phase 11: Real dense connectome (FlyWire)    ✅ N=2000/5000 PASS (hetero model, probe-doubled)
Phase 12: Scale to N≈10000 real topology     ✅ N=9998 PASS (div 0.034-0.037, ~4% fallback)

Pipeline complete at small scale; the human-scale claims remain extrapolation. Remaining work:

• Push real-topology validation past N=5000 (N=10000 attempted, did not complete)
• Empirical d_eff per cell type in mammalian cortex (open — needs MICrONS data; would shrink the 8-OOM extrapolation)
• Body-scan compression bound (empirical, on Visible Human data)
• Whether a rate model on real topology faithfully represents the live circuit (open)
• Engineering scale-up (out of project scope)

⚡ Key Numbers

† Human-scale figure — extrapolated/illustrative, not a demonstrated measurement. Brain spec is the optimistic end of an 8-OOM range (~42–250 KB); cost is a rough projection from an unbuilt system.

📂 Files

Math & Derivations

Direction 1 (functional teleportation):

Directions 2–5 (all closed):

Background Research

Papers

Architecture

Simulation Code

Models:

Scan inverse (evolution — each supersedes the previous):

Pipeline:

Tissue:

Foundational (earlier phases):

🆕 What's Novel Here

Most of the framework is prior art: classical-information teleportation philosophy (Bostrom, Tipler, Parfit), rate-distortion theory (Shannon), connectome inference from activity (Pillow, Paninski, Linderman), single-cell optogenetics (Deisseroth, Boyden, Bargmann), and the wormhole/decoherence math (Visser, Ford-Roman, Penrose-Diósi) are all established.

What appears genuinely novel here (subject to a real literature search):

1. Pool stimulation > per-neuron stimulation for connectome inference, with the explicit empirical comparison (r=0.99 vs 0.81, 3× fewer trials, 2.5× more noise-tolerant). Contrary to naive intuition.
2. Tonic steady-state probes > pulsed probes because of τ/dt noise amplification in the linearization.
3. "Pearson r=0.43 with behavioral div=3%" — explicit demonstration that weight-matrix recovery is a misleading metric vs functional equivalence (rate-distortion theory in concrete instance).
4. Tissue-stratified body information budget with specific D-thresholds per tissue type and the resulting ~247 GB total.
5. Hybrid type-driver + combination pool design that PASSES under combined deployment stress when pure type pools FAIL.
6. Three-organism d_eff scaling fit (1.85 × N^0.46 from C. elegans, Mouse V1, Drosophila). A 3-point fit, not a "law" — the human extrapolation is an 8-OOM reach and is explicitly low-confidence.
7. End-to-end pipeline demonstration at C. elegans combining scan + compress + transmit + reconstruct + verify in one passing rate-model simulation on a real biological connectome's topology (not a simulation of the live animal).

None of this puts a person on a transporter pad. It shows the recipe has no obvious physics holes within the rate-model framing — not that it works on a real brain. The engineering gap is real and decades away, and the rate model is a coarse proxy for real neural dynamics. The scariest-looking obstacle along the way — reconstructing mega-hub neurons at scale — was diagnosed (it's an observability problem: hubs saturate under probing) and addressed (mixed-amplitude probe ladder + probe-doubling, the K·M ≥ 6N rule), pushing the validated range to N=5000 on real connectome topology (with ~4% of neurons on a degree-scaled fallback). The remaining gaps are scale-up and the open question of whether a rate model on real topology faithfully represents the real circuit.

🚫 What Physics Actually Forbids

Real, proven, not going anywhere:

• No-cloning theorem — cannot copy an unknown quantum state
• No-communication theorem — entanglement cannot transmit information FTL, ever
• Ford-Roman quantum inequalities — negative energy density is bounded by the vacuum itself
• Heisenberg — simultaneous precision measurement of conjugate variables is impossible

What GR does NOT forbid (mathematically valid, physically unbuilt):

• Traversable wormholes (Morris & Thorne 1988) — requires exotic matter we can't make
• Alcubierre warp drive (1994) — same problem
• Closed timelike curves — probably destroyed by quantum effects before usable (Hawking)

The gap between "solution to Einstein's equations" and "thing that exists" is where the interesting physics lives.

Standard Model + GR as baseline. All extensions labeled speculative. All numbers shown with derivations. If something is wrong, the calculation says so.