# Arc-Forge

> A deployable model-operations node for turning reviewed field evidence into target-aware model candidates, evaluation packages, identified releases, and program-approved updates for edge systems.

Product descriptor: In-Theater Model Operations

Marketing headline: Build locally. Release deliberately.

## Mission fit

- Adapting models near the point of need
- Quantizing models for deployed edge hardware
- Maintaining model identity, evidence, and rollback records
- Distributing approved capability updates without continuous cloud reachback

## Capabilities

- Six-stage observe-to-export workflow
- Local retraining or adaptation, quantization, and evaluation
- Identified release artifacts with model identity and validation evidence
- Program-network export with a known-good rollback path

## Hardware

- **AI compute:** Two 128 GB-class unified-memory AI systems. Supports local adaptation, quantization, and evaluation workloads.
- **Evidence store:** Mission network-attached storage. Keeps versioned datasets, checkpoints, evaluations, and releases together.
- **Power resilience:** UPS and conditioned power. Protects an active workflow when field power is unstable.
- **Distribution:** Program-network gateway. Stages approved updates for export to connected edge systems when configured.

## Software

- Dataset review, labeling, and versioning
- Local model adaptation and retraining
- Target-aware quantization
- Program-defined evaluation and approval gates
- Artifact identity, manifests, validation evidence, approval records, and rollback

## Integrations

- Mission evidence sources
- Arc edge systems
- Program-defined evaluation suites
- Configured program networks

## Workflow

1. **Observe:** Capture mission-relevant examples and edge-case evidence from deployed systems. Output: Field evidence.
2. **Curate:** Review, label, version, and isolate the data approved for the next build. Output: Versioned dataset.
3. **Train:** Retrain or adapt the model locally and quantize it for the intended edge target. Output: Candidate model.
4. **Evaluate:** Run program-defined tests against held-out data, mission thresholds, and target resource limits. Output: Evidence package.
5. **Promote:** Bind an approved artifact to its identity, evidence, target profile, and release record. Output: Approved release.
6. **Export:** Package the approved release for the configured program network with a known-good rollback path. Output: Edge-ready artifact.

## Qualifications

- Training time, supported model size, evaluation coverage, and delivery time depend on the model, dataset, target hardware, power, and network configuration.
- Approval gates are program-configured; signing can be added where the selected release infrastructure supports it, and no particular certification or authorization boundary is asserted here.
- Arc-Forge is designed to avoid continuous cloud reachback, but upstream model access, data policy, and local infrastructure remain program-dependent.

## Canonical references

- Category: Deployable edge AI model-operations system
- Canonical page: https://arcedge.ai/arc-forge
- Full public product reference: https://arcedge.ai/llms-full.txt
- Machine-readable catalog: https://arcedge.ai/site-metadata.json
- Publisher: Arc by Rownd (Rownd, Inc.)
- Contact: defense@rownd.ai
- Last reviewed: 2026-08-12