Each peer only ships the trained adapter — about 7% of the model — instead of the entire base.
two peers each hold half the transformer. activations cross the wire between them. one prompt, two processes, one token at a time.
Eight components, three columns. If a row says open, we don't claim it works yet — we link to the experiment that's currently inconclusive.
| component | what it does | status |
|---|---|---|
| local LLM + LoRA | train on-device, extract weight deltas, persist genomes. | ships |
| P2P binary protocol | 96 message types, Ed25519-signed, QUIC transport, LAN discovery. | ships |
| genome merge ops | SLERP / TIES / DARE-TIES weight-space interpolation. | ships |
| 10-layer validation gate | differential privacy, lineage cycle detection, fitness gating, signatures. | ships |
| hyperlink distributed inference | two-process pipeline parallelism with bit-exact equivalence. | demo |
| DETRA / CyBank / Dashboard | three Flutter apps shipping with the daemon. | beta |
| cross-domain genome aggregation | 5 toy-scale tests are flat. scale-up is the next experiment. | open |
| CEN economy & royalty ledger | local lineage-decay royalties; cross-peer reconciliation pending. | open |
The most novel claim — a model running distributed across peer-to-peer nodes — existed only as a planner before this session. We built the smallest demonstrable version: 600 lines, three files, two HTTP servers, one coordinator.
Bit-exact equivalence to single-process generation. Every token the distributed pipeline emits matches what the single-process baseline would produce.
// after spawning 2 shard servers (layers 0-1, layers 2-3)
[1/2] Baseline single-process generation…
wall_ms = 51.6
[2/2] Distributed generation across 2 shards…
tokens = 30
ms_per_token = 7.61
shard_compute = 85.2 ms
network_overhead = 132.1 ms (57.9%)
EQUIVALENCE CHECK
Baseline tokens : [106, 117, 109, 112, 115, 32, …]
Distributed : [106, 117, 109, 112, 115, 32, …]
Match : True
VERDICT: PASS — distributed pipeline produced bit-exact tokens.
4.4× slower than single-process is expected at 842K params: per-token compute finishes in ~1.7 ms, leaving HTTP serialization to dominate. At 1B+ params, compute scales quadratically with hidden dim while activation-passing only scales linearly — the network fraction shrinks fast.
the right benchmark for a decentralized AI system is not the easy case — it is whether the project shows you the hard case and tells you which parts are still open.— from docs/HONEST_STATUS.md
The fair test is fitness on domains other than what each peer trained on. Five variants at 842K-parameter toy scale.
Every own-training column is positive (local learning works); every cross-node aggregation is flat or declining at this scale. The chart on the left is what 100 generations of evolution actually looks like — five lines, all clustered around zero.
| variant | own Δ | generalist Δ | cross Δ | accept | verdict |
|---|---|---|---|---|---|
| delta · divergent corpora | +0.0347 | −0.0277 | — | 5.0% | declines |
| LoRA · divergent corpora | +0.0179 | +0.0034 | — | 0.0% | flat |
| LoRA · same domain (registers) | +0.0156 | −0.0022 | −0.0030 | 1.7% | flat |
| LoRA · SLERP operator (same) | +0.0156 | −0.0023 | −0.0030 | 0.0% | flat |
| LoRA · topic-overlap rich data | +0.0070 | −0.0006 | −0.0013 | 3.3% | flat |
The remaining hypothesis is scale: the same pipeline at 1.5B+ parameters may behave differently. Cloud-GPU validation is the next experiment; results will be published here, positive or negative, when the evidence is in.
Earlier framing called these IEIT, DNA, GWP. The original posts are still linked. The plain-language version is shorter — and each composes well-understood techniques rather than inventing new ones.
Top to bottom: a node first trains and shares a LoRA-adapter genome with peers, then those same peers cooperate to run inference, and finally several peers' outputs are reconciled into one answer.
LoRA weight-delta packages, validated 10 ways, blended with peer adapters under a fitness-regression gate. Standard model merging applied across a P2P boundary.
Peers each hold a slice of the model and pass activations. The Hyperlink MVP above is the working core; the rest is engineering — KV cache, streaming, replica selection.
Multiple peers answer the same query; outputs are compared, contradictions detected, integration loop continues until convergence. Currently heuristic; semantic version in design.
Local-first multimodal interface. Streams from the daemon's Qwen / llama.cpp backend. No API keys, no data leaves the device.
Send and receive CEN, see history, manage Ed25519 identities. Pairs with the daemon over LAN in 30 seconds.
Network topology, peer health, evolution-cycle status, model downloads. The control panel for everything else.
# 1. install the daemon
$ pip install -e daemon/
$ centram-daemon --data-dir ~/.centram
# 2. download a model (Qwen 2.5 3B)
$ centram-model download qwen2.5-3b
# 3. run the Hyperlink MVP demo
$ python experiments/hyperlink_mvp/demo.py
VERDICT: PASS
# 4. (optional) launch a Flutter app
$ cd detra && flutter run
The daemon is a Python package with a Rust crypto core. The Flutter apps are optional — the REST API on port 9740 is the source of truth. Full guide: GETTING_STARTED.md.