Strategy · AI Sovereignty

The accountability layer, whatever model is underneath.

PLENA's value is not the model. It is the human-accountable receipt grammar — refusal, human review, authorship, consent, correction, and verification — that sits above whatever AI is used. This roadmap describes how PLENA's own use of AI can mature over time, by condition rather than date, while that accountability layer stays constant.

"PlenaProof does not compete with general-purpose foundation models. PlenaProof is the accountability layer whatever model is underneath."

Five phases, gated by condition — not date

PLENA advances a phase only when its trigger is met and governance review approves. Earlier phases keep working; nothing here claims PLENA trains, deploys, or operates its own model today.

1

API-Assisted Native AI · current posture

PlenaProof uses a configured third-party AI API behind a secure serverless function, with a mock/demo fallback when no key is set. The AI assists preparation, classification, and drafting only — no model is trained or owned by PLENA, keys stay server-side, and the truth boundary is enforced in the interface and on the server.

Trigger: baseline — in place today
2

Knowledge Base + Guardrail Layer

Add a retrieval and guardrail layer so responses are grounded in PLENA's documented receipt formats (VRX-1, VRX-2, VRX-4, VRX-5) and truth-boundary rules, reducing reliance on a model's open-domain guesses. Still API-backed; no self-hosting.

Trigger: sustained assistant usage + a curated PLENA-doctrine corpus worth grounding against
3

Self-Hosted Open-Weight Models

Run open-source, open-weight models (open-weight families such as Llama, Mistral, Qwen, or DeepSeek) on infrastructure PlenaProof or an institution controls — for deployments that require data not to leave a jurisdiction or tenant, and to remove dependence on any single commercial AI provider so access, pricing, and terms stay in PLENA's hands. The accountability layer is unchanged; only where inference runs changes.

Trigger: data-residency, cost, or sovereignty needs an external API cannot satisfy
4

PLENA Fine-Tuned Specialist Model

Fine-tune an open-weight model on PLENA's receipt-grammar and accountability domain to improve consistency on PLENA-specific tasks. A specialist accountability assistant — not a general foundation model, and never a claim of independent verification.

Trigger: enough reviewed, consented, high-quality domain data + governance sign-off on provenance
5

Institution-Deployable PLENA AI

Package the specialist model and guardrail layer so an institution can deploy it inside its own environment, under its own governance, with PLENA's receipt grammar intact.

Trigger: institutional demand for an in-tenant assistant + a supportable deployment and governance model

What never changes

  • PLENA assists; it does not perform official verification, authenticate identity, or certify documents.
  • The receipt grammar — refusal, human review, authorship, consent, correction, verification — is the product, not the model.
  • PLENA complements, and does not replace, existing audit, custody, legal, and compliance infrastructure.
  • PLENA's receipt grammar is designed to support the documentation, decision-logging, human-oversight, and continual-improvement evidence requirements typical of ISO/IEC 42001:2023 Artificial Intelligence Management System implementations, ISO/IEC 42006:2025 requirements for certification bodies, and the EN ISO/IEC 42001:2026 European harmonised version. PlenaProof does not provide ISO/IEC 42001 certification, which is performed by accredited certification bodies.
Boundary. This roadmap describes future direction gated by conditions, not a current state. PlenaProof has not trained, deployed, or operated its own AI model. Each phase is subject to governance review and, where relevant, institutional and counsel sign-off.
See the Proof MoatReal Moat roadmap