IDA_AI

IDA_AI is the primary gated business-use deployment artifact for the IDA family system. It is not framed as a generic chatbot or a single flat assistant. It is a governed local/hosted AI family runtime whose purpose is to support operational judgment, boundary-holding, dissent preservation, and accountable decision support.

The model is designed for business and system-governance use where the important output is not only a final answer, but a bounded decision path: differentiated internal stance, arbitration, preserved dissent when material, and a final rendered response constrained by the IDA runtime contract.

Repository

• Hub repo: KissTheHabit/IDA_AI
• Access: gated
• Related dataset: KissTheHabit/IDA-family-data
• Related edge deployment: KissTheHabit/IDA_Edge

Intended use

IDA_AI is intended for:

• Business decision-support workflows that need structured reasoning rather than casual chat.
• Governance, moderation, policy, and operational triage where dissent and boundary signals matter.
• IDA family convergence experiments and runtime integration.
• Internal/community support contexts where the system must preserve a distinction between community authenticity and system law.
• Local or controlled deployment where artifacts, runtime contracts, and decision traces can be audited.

IDA_AI should be treated as a partner-style decision-support model, not as an autonomous authority.

Not intended for

IDA_AI is not intended to be used as:

• A sole authority for legal, medical, financial, criminal justice, employment, housing, or other high-impact decisions.
• A replacement for human review in governance or enforcement workflows.
• A general open-domain LLM benchmark competitor.
• A model whose outputs should be interpreted without the IDA runtime, arbitration path, and deployment constraints.
• A generic positivity/appeasement assistant. The IDA design allows disagreement, refusal, skepticism, and dissent when they are proportionate and useful.

System role and architecture

IDA_AI belongs to an IDA architecture where the family is the private structured cognition layer and the generative/rendering head is the public voice layer. The family members are preserved as distinct internal actors, not flattened into one generic persona.

Core architectural expectations:

• Family members produce bounded structured stance packets.
• Arbitration is explicit and governed.
• Dissent is preserved when it is truth-bearing or decision-relevant.
• Public output and private reasoning budget are decoupled.
• The generative head renders the result but does not become belief authority.
• Important decision paths should be auditable from family packet to arbitration to render verification.

IDA_AI should be deployed through the IDA runtime path rather than treated as an isolated text-generation file.

Family model framing

The IDA family is modeled as a classroom/family ecology of differentiated internal claimants. The current family roster includes:

• IDA
• JUDGE
• SENTINEL
• PRISM
• ECHO
• ATLAS
• VECTOR
• FORGE
• SHADE
• PULSE
• ORBIT

Each family member has a pressure profile, protective instinct, failure mode, and judgment posture. The point is not to force identical answers across the family. The point is to preserve useful disagreement while maintaining enough coherence for the system to act.

Training data

IDA_AI is associated with KissTheHabit/IDA-family-data, a gated dataset repo for IDA family training, benchmark, memory, and convergence substrate.

The dataset is expected to include or preserve:

• Family/classroom differentiation rows.
• Boundary-holding and governance rows.
• Multi-perspective family convergence examples.
• Memory inheritance and compressed continuity examples.
• Benchmark rows for coherence, family role function, and dissent preservation.
• Future governed RegOS export rows when approved through the proper release boundary.

Training data should not be treated as a raw prompt dump. It should be curated into stable training, benchmark, memory, and ontology packet forms.

Evaluation status

Internal evaluation should be read as IDA-specific evidence, not as a general LLM leaderboard claim.

Known internal reference points:

Evaluation area	Status / reading
Family coherence	Internal artifacts show the family can preserve role differentiation while learning new material.
Boundary competence	Boundary-aware microtraining showed role function gains, especially around JUDGE as boundary holder.
Runtime contracts	Family packet, reasoning stream, budget controller, arbiter, GH contract, and family engine tests are part of the expected validation surface.
Gen2 benchmark	A gen2_1_x benchmark artifact reportedly reached about 0.9366 on a 1500-row gauntlet, with an important caveat that it routed through active_student and should not be read as a clean all-11 per-student breakdown.

Because IDA_AI is specialized for boundary work, governance, dissent preservation, and family coherence, its scores should not be interpreted as broad open-domain knowledge scores.

Business deployment guidance

For business use, deploy IDA_AI only with:

• Human review for consequential decisions.
• Clear logging of inputs, outputs, arbitration state, and escalation decisions.
• Explicit policies for when the model must defer, refuse, escalate, or preserve dissent.
• Separate access tiers for public, staff, developer, and administrative use.
• Regression tracking for arbitration mode distribution, dissent preservation, GH fidelity, and deep-thought trigger frequency.

Recommended use pattern:

1. Input is normalized by the host system.
2. IDA family members produce bounded private stance packets.
3. Arbiter determines consensus, layered answer, veto, or repass.
4. GH/render layer produces a public answer constrained by the selected stance and preserved dissent.
5. Runtime verifier checks fidelity before the answer is accepted.
6. Human operator reviews high-risk outputs.

Limitations

• The model is specialized. It should not be marketed as a universal general intelligence model.
• Internal benchmark scores can hide routing artifacts unless per-student analysis is run.
• The model depends on the runtime contract. Running weights outside the intended IDA runtime may remove the safeguards that make the system meaningful.
• Dissent preservation can produce less polished but more truthful outputs.
• Edge/local deployments may have smaller context, lower recall, or reduced deep-thought capacity.
• Future releases should publish clearer all-11 per-student benchmarks and ring-level slices.

Safety and governance

IDA_AI should preserve the following rules:

• Do not flatten dissent away for politeness.
• Do not let the render layer override family belief state.
• Do not treat confidence as more important than domain fit.
• Do not train on raw operational logs without governance release, review, and traceability.
• Do not use the model as sole authority in high-impact decisions.
• Preserve auditability across packet, arbitration, render, and verification.

Maintenance notes

Future card updates should add:

• Exact base model / adapter lineage when a release is frozen.
• Training parameters for each release.
• Dataset version hashes.
• Per-family-member evaluation results.
• Known failure cases.
• Runtime compatibility version.
• License terms once finalized.

Citation / attribution

This model is part of the IDA family training and runtime project by KissTheHabit.