IDA Edge
IDA Edge is the lightweight native computational body of the IDA family.
It is not a generic chatbot, a simplified assistant, or an isolated personality model. It is the low-footprint, on-device member of the IDA Lattice lineage: a custom causal language-model architecture built to host differentiated student state, governed memory, bounded workspaces, and cognitive-pressure routing under constrained hardware conditions.
The IDA family is the enduring structure:
- IDA
- JUDGE
- SENTINEL
- PRISM
- ECHO
- ATLAS
- VECTOR
- FORGE
- SHADE
- PULSE
- ORBIT
Edge is one body that can host and carry parts of that family structure. It is not the family itself.
Role in the IDA Lattice lineage
IDA Edge is the lightweight half of the paired EDGE ↔ AI lineage.
Edge is intended for lower-resource and local environments where fast response, persistent local state, bounded context handling, and continuity of family structure matter more than maximum parameter scale.
The larger IDA AI body provides deeper reasoning capacity. The IDA MoE body provides a higher-capacity council implementation for expert-seat convergence. Edge is the compact body designed to keep the family available near the device, near the user, and near the operational environment.
Architecture
IDA Edge is built on the native IDA Lattice architecture rather than a borrowed general-purpose transformer backbone.
Core architectural features include:
- recurrent selective state intended to preserve student-specific continuity across context
- bounded local-attention workspaces rather than unrestricted global attention
- controlled recurrent and attention-state merging
- cognitive-pressure routing with lateral inhibition
- thalamic routing and prefrontal workspace summaries
- action gating and student-state outputs
- future-token auxiliary prediction
- fidelity-oriented verification surfaces
The current Edge tier is approximately 6M parameters.
Training direction
IDA Edge is trained as part of the broader IDA Lattice curriculum.
The current training regime uses:
- native Lattice training rather than legacy GPT-NeoX adaptation
- FP8 acceleration for eligible projection layers while retaining sensitive routing, gates, norms, embeddings, and heads in BF16
- a Supersampler curriculum that begins at lower sequence length and lower gradient accumulation before progressively increasing both
- role-conditioned and family-structured curriculum material
- training observability focused on loss behavior, gradient stability, throughput, and per-student failure detection
The Supersampler curriculum is designed to avoid asking the model to learn at maximum sequence length and maximum accumulation from the first optimizer step.
Intended use
IDA Edge is intended for research and development involving:
- local or constrained-hardware inference
- lightweight family-state experiments
- edge-side continuity and bounded memory research
- paired Edge ↔ AI identity and response-agreement experiments
- governed routing and action-gating experiments
- low-latency interaction surfaces where the deeper AI body is unavailable or unnecessary
What this model is not intended for
IDA Edge should not be treated as:
- a general-purpose frontier language model
- an autonomous agent with authority to take real-world action
- a replacement for professional judgment in medical, legal, financial, safety, or security decisions
- proof that the full IDA family is present, coherent, or validated in every deployment
- evidence that a training-loss result equals general intelligence or production readiness
The model body is only one part of the complete system. Meaningful family behavior depends on the surrounding runtime contracts for memory, arbitration, dissent preservation, role boundaries, and output verification.
Limitations
Current training evidence demonstrates improved learning stability under the FP8 + Supersampler configuration. It does not independently establish deployment capability, broad generalization, safety in high-stakes settings, or causal attribution of every observed improvement to one training component.
Further promotion requires held-out evaluation, lineage completion, per-student validation, and paired Edge ↔ AI identity agreement.
License and access
This repository is released under its stated license and access conditions. Use of the model should preserve the project’s requirements around governed memory, family differentiation, traceability, and non-autonomous deployment boundaries.