Next-Gen Foundation Architecture

Version: v0.2 Status: Review Owners: FMIE / Foundation / Cortex / Frontend Last Updated: 2026-03-25 Change Type: additive Migration Required: yes Depends On: FMIE Media Passport, External Cognition Gateway, Semantic Airgap Supersedes: none Superseded By: none

Purpose

Define the canonical next-generation architecture after reconciling the implemented Foundation-V2 system, the locked FMIE Media Passport architecture, and the current Cortex design direction.

Source Inputs

• Foundation-V2 implementation audit in /mnt/Storage-1/worktrees/Foundation-V2
• /mnt/Storage-1/worktrees/Foundation-V2/foundation-v2-architecture.html
• /mnt/Storage-1/worktrees/Foundation-V2/docs/Cortex-spec.md (canonical Cortex spec v2.1)

Design Principle

Use frontier reasoning without exporting raw institutional memory by default.

Current V2 Baseline

Foundation-V2 already has these real layers: 1. Interaction Plane - Webchat UI - Dashboard UI - Discord bridge 2. Foundation API - FastAPI gateway - turn routing, provenance, intelligence workflows 3. Knowledge & Reasoning Plane - Triton router - BORG KB client - Answerer - Aristotle / Deepthink analysis - Provenance store 4. Data Stores - PostgreSQL - Qdrant - materialized KB - turn history and runtime artifacts 5. Ingestion & Control Plane - Morpheus collector / registry / learning - extractors and writeback - Haystack structured extraction 6. Governance & Orchestration - Atlas - Sentinel - Proteus - IAM control

V2 Constraint

Foundation-V2 can already route, retrieve, analyze, publish, and govern briefs, but its external model use is still direct. Prompt routing, answer synthesis, and structured analysis can send raw prompts and rich internal evidence to Anthropic without a true disclosure-transform boundary.

Canonical Target Layers

1. FMIE - multimodal perception runtime - Media Passport emission - provenance, authenticity, transform-ready outputs
2. Foundation Core - KB, orchestration, governed intelligence briefs - partner-fit engine - strategy twin - local final synthesis
3. Semantic Airgap - disclosure policy engine - transform compiler - surrogate-ID packet builder - outbound prompt firewall
4. External Cognition Gateway - public research lane - abstract reasoning lane - approved raw exception lane - receipts and trust import
5. Midnight Commitment Layer - commitments - revocation / nullifiers - selective disclosure proofs
6. Cortex - sovereign meta-layer for health, optimization, research, and fleet intelligence

Canonical Boundaries

FMIE owns

• multimodal perception
• provenance and authenticity primitives
• Media Passport outputs
• disclosure-ready derived views

Foundation owns

• reasoning over internal evidence
• strategy twin
• partner-fit engine
• disclosure policy
• external cognition orchestration
• local final synthesis

Cortex owns

• governed observability
• optimization and recommendation loops
• world model and fleet intelligence
• autonomous action policy within guardrails

Triton Front-end owns

• operator UI contracts
• trust / provenance rendering
• disclosure exception approvals
• progress and state visualization

New Mandatory Components

Semantic Airgap

Every outbound frontier-model call must pass through: - policy resolution - transform compilation - surrogate replacement where required - outbound validation - receipt generation

Partner-Fit Engine

Use Hermophylax-style corp profiles, startup capability signals, opportunity candidates, contradictions, and watchlists to generate private partner rankings.

Strategy Twin

Persist strategic state over time with: - beliefs - contradictions - negative evidence - confidence decay - open questions - change events

Midnight Layer

Anchor commitments for: - source passport - transform - task packet - external receipt - later dispute or revocation

Cortex Placement

Cortex sits above the stack but does not bypass the stack. It can observe health, telemetry, receipts, and safe summaries, but not unrestricted tenant raw data.

Key Changes From V2

• direct Anthropic calls are no longer canonical
• all frontier-model egress must pass through the airgap and gateway
• partner-fit and strategy-twin become first-class Foundation systems
• Cortex observes governed views, not unrestricted tenant raw data
• Media Passport becomes the provenance bridge between FMIE and Foundation

Migration Path

V2 -> V3

• add Semantic Airgap
• add External Cognition Gateway
• add receipt and trust scoring
• keep existing briefs, routing, and governance

V3 -> V4

• add Partner-Fit Engine
• add Strategy Twin
• add stronger local final synthesis
• promote external reasoning imports into persistent strategic state

Open Questions

• whether embeddings can ever leave Foundation under abstract_reasoning_only
• what qualifies for approved_raw_exception by modality and tenant
• how much of partner-fit can be externalized without leaking strategic joins

Change Log

• v0.2: reconciled current Foundation-V2 architecture and Cortex design into the canonical next-gen stack
• v0.1: initial seed