DendriSwarm v0.8.0 — CIFAR-100 Training Dashboard

DendriSwarm coordinates heterogeneous CPU machines to train and search a persistent Native10-derived Dendritron on real CIFAR-100 data. v0.8.0 keeps the full v0.7 campaign and adds a one-command local application for contributors and campaign operators.

New to DendriSwarm? Follow the first-time user guide for prerequisite installation, the correct coordinator/seed startup order, CIFAR-100 preparation, campaign launch, health checks, and common Windows fixes.

Open the app

Install the lightweight contributor package and run it with no subcommand:

python -m pip install "git+https://github.com/MMVFIRM/dendriswarm.git"
dendriswarm

dendriswarm app is equivalent. A browser opens to a token-protected dashboard on 127.0.0.1:8788.

A downloaded repository can be launched directly:

• Windows contributor: launch-dashboard.bat
• Windows campaign operator: launch-operator-dashboard.bat
• macOS/Linux contributor: ./launch-dashboard.sh
• macOS/Linux campaign operator: ./launch-operator-dashboard.sh

The launchers create .venv, install the appropriate package, and open the dashboard. Operator launchers install dendriswarm[coordinator]; ordinary contributors retain the smaller package.

What the dashboard controls

Contributor view

• Coordinator URL and optional out-of-band fingerprint.
• CPU and memory contribution percentages.
• Disk/cache budget and maximum task duration.
• Battery policy and system-load pause threshold.
• Start, pause, resume, and stop controls.
• Live task, outbox, credit, completion, capability, and log telemetry.

The worker continues to enforce the signed resource contract in an isolated subprocess. Dashboard settings update the same atomic, hot-reloaded SeedPolicy used by the CLI.

Campaign operator view

• Start and stop a local coordinator.
• Prepare and verify the official CIFAR-100 archive.
• Initialize Native10 topology or import an established checkpoint.
• Preview the next routing/model tournament.
• Configure independent search count, sample budget, optimizer steps, learning rate, and verifier quorum.
• Queue the next training round.
• Track canonical roots, routed accuracy, oracle routing gap, top-k category recall, promotions, worker-hours, and campaign history.
• Produce the final official-test report.

Mutating operator actions require a local coordinator-generated admin token. The token is never published through coordinator metadata.

Campaign topology

official CIFAR-100 images
        ↓
channel normalization + eight spatial 8×16 RGB patches
        ↓
8 trainable sensory field blocks
        ↓
96-wide shared representation
        ↓
1,000 trainable routing scouts
        ↓
top-4 routing with bounded low-margin expansion to top-8
        ↓
20 colonies aligned to CIFAR-100's 20 coarse categories
        ↓
5 fine classes per colony
        ↓
45 experts per colony; rotating 15-of-45 local updates
        ↓
4 nonlinear branches per expert + associative memory
        ↓
100 fine-class predictions

The exact model contains 4,898,812 trainable floating-point parameters, all reachable through bounded protocol operations.

Data and evidence boundary

The dataset is not redistributed. The operator supplies the official Python archive. Preparation verifies its published MD5, safely extracts it, preserves the official fine/coarse mapping, and creates:

Split	Rows	Use
Train	45,000	contributor training and public routing diagnostics
Selection bank	2,500	trainer-invisible candidate selection; one-shot folds
Replication bank	2,500	separate final replication; one-shot folds
Official test	10,000	final reporting only; rejected by the planner

Candidates promote only after independent search, hidden all-class selection, exact one-sided McNemar testing, familywise correction, effect and per-class harm gates, deterministic replay, and a separate one-shot replication quorum.

CLI remains available

Every dashboard action has a command-line equivalent:

dendriswarm doctor
dendriswarm seed --coordinator https://HOST --coordinator-fingerprint SHA256 --share 25
dendriswarm cifar100-prepare ./cifar-100-python.tar.gz --state ./state
dendriswarm cifar100-init --state ./state --checkpoint ./native10-checkpoint.json
dendriswarm cifar100-plan --state ./state
dendriswarm cifar100-queue-next --state ./state
dendriswarm cifar100-evaluate-test ./cifar100-test-report.json --state ./state

Dashboard security and package size

The dashboard uses the Python standard library HTTP server and the dependencies already needed by a contributor. It does not require Node.js, Electron, React, or a separate web server.

• Loopback binding only.
• Random 256-bit launch token.
• SameSite=Strict, HTTP-only browser cookie.
• Request-size ceiling.
• Local process supervision with PID creation-time checks.
• Separate coordinator admin token for mutating campaign actions.
• Bounded log tails; no packaged identities, databases, datasets, or secrets.

See docs/DASHBOARD_V08.md for the complete interface and persistence model.

Verification and claims

python -m pip install -e '.[dev]'
make test
make proof-v08

v0.8.0 adds dashboard/configuration tests and a dedicated usability/security proof while retaining every v0.2–v0.7 proof family. The official archive was not available in the packaging environment, so the package still does not fabricate a CIFAR-100 accuracy result. Competitive accuracy and distributed-compute leverage remain outputs of the real campaign.

See:

• docs/FIRST_TIME_USER.md
• docs/DASHBOARD_V08.md
• docs/CIFAR100_SWARM_V07.md
• docs/CLAIMS.md
• SECURITY.md
• DATA_LICENSES.md