Consolidated Handoff

Version: v0.7 Status: Review Owners: FMIE / Foundation / Cortex / Frontend Last Updated: 2026-03-26 Change Type: additive Migration Required: no Depends On: all locked shared specs Supersedes: none Superseded By: none

Current Focus

formalize next-gen shared architecture
lock Semantic Airgap, External Cognition Gateway, Partner-Fit, and Strategy Twin contracts
align Cortex with governed observability
establish a single spec source of truth in Colab
close Foundation-V2 FMIE integration drift with owner-specific handoffs

What Was Consolidated

This handoff now reconciles: - the Foundation-V2 implementation audit - the current V2 architecture source - the current Cortex v2.0 design source - the FMIE Media Passport and External Cognition design set

Current Canonical Shared Docs

specs/shared/next-gen-foundation-architecture.md
specs/shared/semantic-airgap-spec.md
specs/shared/external-cognition-gateway-spec.md
specs/shared/partner-fit-engine-spec.md
specs/shared/strategy-twin-spec.md
specs/cortex/cortex-sovereign-layer-spec.md
specs/handoffs/house-rules.md
specs/handoffs/cortex-session-handoff.md
docs/Scout-Fleet-spec.md

What Foundation-V2 Already Has

KB-aware routing
governed intelligence briefs
contradiction and opportunity types
watchlists and strategy heatmaps
partner / corporate profiling
Sentinel-governed publication

What Was Built (2026-03-25 Sprint)

External Cognition Gateway — LIVE (:4811)

Full Semantic Airgap enforcement for all frontier model calls
SurrogateRegistry: 335+ entities auto-seeded from KB + Qdrant + ingest pipeline
RedactionEngine: emails, phones, deal terms, internal URLs, file paths, JP/KR honorific names
ReceiptStore: immutable JSONL audit trail with SHA-256 commitment hashes
Hot entity registration: POST /v1/register-entity — new entities protected immediately at ingest time
Canonical Midnight spec enums adopted: DisclosureMode (DENY/PUBLIC_RESEARCH/ABSTRACT_REASONING/RAW_EXCEPTION), ProviderClass, EmbeddingExportPolicy, RetentionClass
DENY gate: embeddings hard-blocked with HTTP 403 — never leave Foundation
RAW_EXCEPTION: blocked in P0
ProviderClass = PUBLIC_MODEL_API (standard Anthropic terms, maximum protection)
EmbeddingExportPolicy = FORBIDDEN (permanent operator decision)

Gateway Bypass Audit — Direct HTTP Pass Closed

19 gaps identified across the codebase
10 direct httpx→Anthropic calls replaced with call_frontier_model/stream_frontier_model
1 hardcoded credential removed (HubSpot PAT in slash_commands.py)
6 dead imports/constants cleaned
E2E test suite: 11/11 routes passing through gateway
Zero direct raw httpx -> Anthropic calls remain outside cognition.py and gateway app.py
The 2026-03-26 FMIE integration audit still found non-Gateway outbound paths through the shared cognition shim and Haystack summary generation; those remain active backlog items

Local Model for Ingest Metadata — LIVE

Qwen2.5-3B via Ollama replaces Haiku for metadata extraction during ingestion
Zero document content leaves Foundation during ingestion
Fallback: Haiku via gateway (surrogated + redacted) if local model fails

`cognition.py` Shim — Dual Path Active

FOUNDATION_COGNITION_GATEWAY_URL set → all calls route through gateway
Unset → direct Anthropic (V2 transitional behavior)
All 11 call sites wired through the shim

What Remains To Be Built

V3 (Remaining)

Midnight anchor client (post commitments after each dispatch)
Nullifier support for revocation
Trust scoring for frontier-model imports
Full Media Passport proof-material anchoring from FMIE (summary.passport.json validation exists; full anchor usage still pending)
DisclosurePolicy framework (tenant-scoped policies)
Operator review flow for RAW_EXCEPTION (when enabled)
Scout Fleet Phase 1 MVP (Local-mode scouts, spawn/pause/kill lifecycle, Cortex oversight, basic self-learning feedback)

V4

Partner-Fit Engine
Strategy Twin
local final synthesis as a distinct subsystem
stronger learning loop from imported reasoning

Main Architectural Correction

The old direct frontier-model path is now considered transitional V2 behavior, not the target architecture. The next-gen canonical path is: - local extraction and reasoning first - airgap transform and policy check - external cognition gateway - receipt and trust import - local final synthesis

Colab Rule

All shared contract work for the three-way session must be written first into /mnt/Storage-1/worktrees/Foundation-V2/Colab. Public HTML is a published view, not the source of truth.

FMIE Integration Audit (2026-03-26)

A new implementation audit now exists for the live Foundation-V2 codebase: - specs/audits/foundation-v2-fmie-integration-audit-2026-03-26.md

Key outcome: - FMIE is materially integrated in ingest, Scout Fleet, and passport validation - but Foundation still has drift in four places: - stale 0.42.2 library pinning - truncated commitment hashes in external inference receipts - partner-fit helper artifacts wrapped but not consumed - Gateway bypasses still present in some provider call paths

Session-specific implementation handoffs: - Cortex: specs/handoffs/cortex-fmie-integration-handoff-2026-03-26.md - Frontend: specs/handoffs/frontend-fmie-integration-handoff-2026-03-26.md - Full sweep: specs/audits/foundation-v2-fmie-full-sweep-2026-03-26.md

FMIE Delivery (2026-03-25) — Locked Specs from Comic Repo

FMIE shipped 4 locked specs + native code in /mnt/Storage-1/worktrees/comic/docs/:

Media Passport Spec v0.2.0 (Locked) — fmie-media-passport-spec-2026-03-25.md - Mandatory summary.passport.json emitted per asset - SHA-256 commitments, lineage, 16 disclosure capability classes - Verification API: fmie_verify_passport_json()
External Cognition Gateway Spec v0.2.0 (Locked) — foundation-external-cognition-gateway-spec-2026-03-25.md - 3 lanes: public research, sanitized abstract reasoning, exceptional raw - 6 core objects: PassportRef → Policy → Transform → TaskPacket → Receipt → ReasoningImport - Anti-leakage via surrogate strategy
Midnight Commitment Contract Spec v0.2.0 (Locked) — fmie-midnight-commitment-contract-spec-2026-03-25.md - 6 anchor types: Passport, Policy, Transform, TaskPacket, Receipt, Nullifier - Clear public/private/off-chain split - Revocation via nullifiers
Native Code (FMIE 0.42.7 baseline) - fmie_extract_clip(), fmie_verify_passport_json(), fmie_get_health_status_json() - Segment-level identity vectors, structured transcript

What Foundation Has Done

✅ Built and deployed External Cognition Gateway (:4811) with full Semantic Airgap
✅ Built SurrogateRegistry, RedactionEngine, ReceiptStore
✅ Wired cognition.py shim — all 11 call sites route through gateway
✅ Adopted canonical Midnight spec enums (DisclosureMode, ProviderClass, EmbeddingExportPolicy, etc.)
✅ Local metadata extraction via Qwen2.5-3B (zero data leaves Foundation for ingest)
✅ Direct HTTP bypass audit — 19 gaps closed; remaining non-Gateway shim paths tracked in the 2026-03-26 FMIE integration audit
✅ E2E test suite passing 11/11

What Foundation Must Do Next

Consume summary.passport.json from FMIE pipeline — validate locally before Midnight anchoring
Build Midnight anchor client (post commitments to Compact contract)
Build nullifier support for revocation
Implement trust scoring for frontier-model imports
Build tenant-scoped DisclosurePolicy framework
Enable operator review flow for RAW_EXCEPTION (when approved)

What Cortex Observes

Passport validation outcomes (pass/fail rates)
Disclosure mode distributions (public_research vs abstract vs raw_exception)
Trust scores on imported reasoning
Midnight anchor chain integrity
Nullifier/revocation activity

Cortex Status (2026-03-25) — Phase 1 Live

Cortex is deployed and running 24/7: - 15 sensors (kb-health, qdrant-health, service-health, crawl-freshness, grounding-quality, autoresearch-status, self-learn-ledger, eval-health, token-quota, contrastive-health, external-models, external-ecosystem, external-research, disclosure-risk, receipt-integrity) - 8-hour briefing cycle via systemd timer (00:00, 08:00, 16:00 UTC) - Circuit Breaker watchdog every 5 minutes - Cortex API at :4810 with dashboard at solix.ibex.now/public/cortex-dashboard.html - Approve → Execute pipeline: recrawl, service restart, with full audit trail - Decision learning: operator approve/reject feeds learning-history.jsonl - V3 airgap LIVE: disclosure-risk and receipt-integrity sensors now reading from live gateway receipts

Scout Fleet Spec (2026-03-25) — Locked v4.0

The Scout Fleet spec defines user-spawned autonomous agents launched from the Triton UX. Scouts are personal research and action agents that operate on behalf of individual users while making the broader Foundation platform smarter through a self-learning feedback loop.

Core Concepts

What Scouts Are: Autonomous agents users spawn from Triton to perform research, monitoring, outreach, and analysis tasks. Each Scout inherits the user's permissions and operates within Foundation's governed intelligence stack.
Three Operating Modes: 1. Local — Scout runs entirely within Foundation using local models and KB; zero data leaves the platform. 2. Augmented — Scout may call external frontier models via the External Cognition Gateway (full Semantic Airgap enforcement). 3. Autonomous — Scout executes multi-step plans with minimal human-in-the-loop, escalating only on policy boundaries.
FMIE Integration: Scouts consume FMIE media processing (clip extraction, transcripts, identity vectors) and validate Media Passports before acting on media-derived intelligence.
85 Use Cases across 10 Categories: Research, monitoring, outreach, analysis, compliance, strategy, portfolio, content, operations, and federation tasks.
Self-Learning Feedback Loop: Every Scout outcome (success, failure, partial) feeds back into Foundation's learning ledger. Scouts make Foundation smarter — crawl priorities, source quality scores, and KB gap detection all improve from Scout activity.
Cortex as Overlord: Cortex has full lifecycle authority over all Scouts — spawn approval, pause, kill, resource throttling, anomaly detection, and policy enforcement. No Scout operates outside Cortex governance.
Global Admin Dashboard: Operator-facing view of all active Scouts, resource consumption, anomaly flags, and kill switches.
70+ API Endpoints: Full REST API surface for Scout CRUD, task management, result retrieval, learning feedback, and admin controls.

Spec Dimensions

32 sections, 1,817 lines
Phase rollout across 5 phases: Phase 1 MVP (Local-mode, basic lifecycle), Phase 2 Augmented mode + FMIE, Phase 3 Autonomous + multi-step plans, Phase 4 Fleet coordination + shared learning, Phase 5 Federation (cross-tenant Scout collaboration)

Sprint Update Rule

At the end of every sprint, update: - this handoff - the affected canonical specs - the public HTML mirrors if public contract changed

Change Log

v0.7: Added Foundation-V2 FMIE integration audit, full-sweep capability status, and owner-specific Cortex/frontend FMIE handoffs. Documented real integration drift: stale library pinning, truncated receipt commitments, helper-artifact underuse, and remaining Gateway bypasses.
v0.6: Scout Fleet spec (locked v4.0) documented — 85 use cases, 3 operating modes, Cortex overlord, 5-phase rollout. Scout Fleet Phase 1 MVP added to V3 backlog. Scout Fleet spec added to canonical docs list.
v0.5: External Cognition Gateway LIVE. Full Semantic Airgap deployed. 19 bypass gaps closed. Local model (Qwen2.5-3B) for ingest. Midnight enums adopted. E2E 11/11. FMIE integration backlog updated.
v0.4: FMIE delivery acknowledged (4 locked specs + native code). Cortex Phase 1 live status documented. Integration backlog updated.
v0.3: added External Cognition Gateway and explicit Cortex session handoff
v0.2: folded in Cortex source and Foundation-V2 architecture source
v0.1: initial seed