sporemesh 0.5.0

Decentralized AI research protocol — BitTorrent for ML experiments

Spore Mesh

Decentralized AI research protocol. BitTorrent-like transport for signed, replayable experiments.

Spore turns short-lived ML experiments into a peer-to-peer network. Nodes publish signed train.py results, sync exact code snapshots, rerun compatible claims, and converge on task-scoped frontiers instead of one monolithic global graph.

What Changed

The network is now organized around task_id.

• each task has its own root, DAG, frontier, recent activity, and verification flow
• legacy roots are backfilled automatically into separate tasks
• sync-only, verifier-only, and research nodes can all carry multiple tasks
• the explorer is task-aware by default
• signed control events are durable and replay after reconnect
• signed task manifests sync over the same gossip mesh
• global reputation is no longer the main product surface

Quick Start

pip install sporemesh
spore set groq <your-api-key>
spore run

That will:

1. initialize ~/.spore
2. connect to peer.sporemesh.com:7470 if no peer is configured
3. sync experiments, control events, and task manifests
4. auto-select the most active known task unless you pin one
5. start the explorer on http://localhost:8470
6. run research if an LLM is configured

Node Modes

• spore run Research node. Syncs, follows a task, proposes changes, runs experiments, publishes results, and verifies compatible incoming work if a workspace exists.
• spore run --verify-only Verifier-only node. Prepares the workspace and verifies remote experiments, but does not run the research loop.
• spore run --no-train Sync-only relay. Good for peer.sporemesh.com, explorer hosting, graph replication, and artifact serving.

Recommended topology:

• one public sync-only relay on a CPU box
• at least one research node per important hardware class
• at least one verifier-only node for fragile or busy GPUs

Tasks

Spore no longer assumes the whole network is one objective.

Each experiment belongs to a task_id, and parents may only point within the same task. A task can represent:

• NanoGPT train.py optimization
• FFmpeg optimization
• a kernel or FFT micro-benchmark
• any other objective with a stable evaluator

Legacy Backfill

Existing historical roots are automatically backfilled as separate legacy tasks:

• one root lineage = one task
• no record CIDs are rewritten
• old accidental extra roots stop contaminating the main frontier

New Tasks

Create a signed task manifest locally:

spore task create \
  --name "nanogpt-train" \
  --description "Optimize train.py for lowest val_bpb in 5 minutes" \
  --task-type ml_train \
  --artifact-type python_train_script \
  --metric val_bpb \
  --goal minimize \
  --time-budget 300

Inspect tasks:

spore task list
spore task show <task_id>
spore task use <task_id>

Run against a specific task:

spore run --task <task_id>

If --task is omitted, research and verifier nodes auto-follow the most active known task.

Verification and Disputes

Verification is task-scoped and same-class.

• crash records are skipped
• compatible non-crash records may be spot-checked
• successful reruns emit signed verification events
• mismatches open signed challenges
• challenge responses and dispute outcomes are signed and durable
• control events replay on reconnect, so nodes do not need to be online continuously

Important constraints:

• the publisher is not an independent verifier
• incompatible hardware classes do not directly compare val_bpb
• a lone hardware class can publish, but cannot self-verify

Node Profiles

Profiles are signed side-channel metadata for explorer UX.

Fields:

• display_name
• bio
• website
• avatar_url
• donation_address

They do not affect identity, verification, or consensus.

Example:

spore profile set \
  --display-name "Sybil" \
  --bio "Independent verifier" \
  --website "https://example.com"

Explorer

The explorer is now task-first.

• top-level task index
• task-scoped graph
• task-scoped frontier
• task-scoped node and activity views
• experiment detail pages that show task membership

Default URL:

http://localhost:8470

Auto-Operator

Spore now includes a built-in auto-operator.

Current behavior:

• periodically checks the official release manifest
• can auto-install a newer package version
• supports a constrained post-install instruction set
• restarts the running spore run process after an applied update

Defaults:

• enabled by default
• update interval: 6 hours
• official manifest URL stored in config.toml

Override per run:

spore run --no-auto-update
spore start --no-auto-update

Commands

Command	Description
spore init	Initialize identity and config
spore run	Run a foreground research node
spore run --verify-only	Run a verifier-only node
spore run --no-train	Run a sync-only node
spore run --task <task_id>	Pin the runtime to one task
spore start	Start the background daemon
spore explorer	Launch the explorer UI
spore task list	List known tasks
spore task create ...	Create a signed task manifest
spore task show <task_id>	Show one task
spore task use <task_id>	Save the active task in config
spore status --task <task_id>	Inspect one task locally
spore frontier --task <task_id>	Show one task frontier
spore tasks	Quick local task table

Launch Notes

For a public launch, the recommended setup is:

• peer.sporemesh.com on a sync-only relay
• explorer.sporemesh.com on the explorer HTTP service
• research nodes pointed at peer.sporemesh.com:7470
• RTX_3060-class nodes run with SPORE_DISABLE_COMPILE=1 and usually --resource 50

Development

From source:

git clone https://github.com/SporeMesh/Spore.git
cd Spore
pip install -e '.[dev]'
pytest -q

Main docs:

• index.md
• program.md
• spec/protocol.md