FDAI Roadmap

The engineering plan behind FDAI. The FDAI Constitution defines the highest design authority. This folder expands its short-form execution summary in copilot-instructions.md and the control loop in architecture.instructions.md into an actionable, phased roadmap: from goals and structure through deployment and scale-out.

Read online: dotnetpower.github.io/fdai. The markdown here is the canonical source; the site mounts these files read-only with sidebar navigation, right-column TOC, full-text search, and a Korean / English switcher. See site/ for how the mount and deploy work.

Scope: the repo is generic and customer-agnostic. Deployment values stay in environment configuration or secret stores; optional downstream distributions limit or extend capabilities through supported seams (generic-scope.instructions.md).

Implementation focus: Azure is the only implemented target. Non-Azure providers and Phase 4 multi-cloud expansion are TBD. The CSP-neutral abstractions in these docs exist so a future adapter is additive, not a delivery commitment (Implementation Focus).

Design at a glance

Deterministic-first, event-driven, risk-gated. A 3-tier trust router resolves repeatable events with rules and policies (T0) and lightweight similarity reuse (T1), reserving frontier-model reasoning (T2) for the ambiguous residual. Every autonomous action ships in observation mode first and is promoted per-action explicitly. Coverage shares and autonomy multipliers are design targets that require a measured baseline before they can be claimed (goals-and-metrics.md).

How to read this folder

Reference docs describe the system; phase docs (P0-P4) sequence the build. Read the reference docs first, then the phases in order.

Agent design routes load only the owner documents required for a changed path. The route manifest caps canonical English roadmap context at 4,500 physical lines per route. When a route reaches the cap, narrow its owner set or split a multi-responsibility owner instead of adding more context.

Core reference (system shape)

#	Document	What it covers
0	fdai-constitution.md	purpose, guarantees, authority precedence, domains, autonomy, and amendment rules
1	goals-and-metrics.md	success criteria, KPIs, measurement-first rule
2	project-structure.md	repo layout, module boundaries, control-loop wiring
3	tech-stack.md	languages, frameworks, data stores, event bus
4	csp-neutrality.md	wire-level contracts that keep the core CSP-neutral
5	llm-strategy.md	per-tier model choices, mixed-model gate, abstraction
5a	operating-ontology.md	shared cloud-operations meaning for services, workloads, objectives, decisions, effects, agent ownership, and controlled extension
5b	outcome-assurance.md	scoped operational-readiness, objective-alignment, and control-assurance projection over the three FDAI verticals
5c	operating-ontology-platform.md	agent-supporting ontology safety infrastructure: exact releases, bounded object sets, mutation plans, typed functions, reconciliation, and SDKs
5d	operating-ontology-metamodel.md	operational lenses, canonical declaration kinds, State and Context boundaries, authority, time, ownership, and additive migration
6	security-and-identity.md	least-privilege identity, secrets, safety invariants
7	deployment.md	IaC, CI/CD, environments, release / rollback
7a	architecture-review-board.md	canonical ARB packet: decision boundary, evidence contract, owners, dependencies, production exit gate
7b	data-governance.md	data inventory, classification, lifecycle, privacy assessment, model-provider and compliance evidence
7c	Architecture Decision Records	ADR register and accepted Azure day-zero platform baseline
7d	mscp-operational-profile.md	selective MSCP-derived effect, cycle, and runtime-integrity policies without a full conformance claim
7e	service-graduation-and-ownership.md	measured service-split gates, single-writer data ownership, contracts, identities, rollback, and boundary docstrings
7f	service-decomposition-execution-plan.md	five-service target, dependency-ordered work packages, parallel lanes, progress, blockers, and evidence receipts

Rules, detection, and operations

#	Document	What it covers
8	rule-catalog-collection.md	where rules / checklists / baselines come from and their YAML shape
9	rule-governance.md	how admins author, scope, enable, and exempt rules (Azure Policy-like)
10	observability-and-detection.md	event correlation, anomaly detection, forecasting, root-cause analysis
10a	manual-distillation.md	compiling an adopting company’s operational / deployment manuals into deterministic rules / workflows / policies (vs runtime RAG), and verifying the distillation
10b	operational-learning-ontology.md	converting benchmark and live incident outcomes into immutable cases, deterministic failure fingerprints, governed rule candidates, and reusable promoted operating patterns
10c	causal-incident-graph.md	ontology-grounded causal hypotheses, support and refutation evidence, evidence grades, and outcome closure
10d	document-ontology-distillation.md	compiling governed operational documents into grounded, review-only ontology object and link proposals with deterministic verification
10e	policy-abstraction-and-control-objectives.md	provider-neutral control objectives, proof-carrying Rule bindings, authority boundaries, corpus-scale generations, migration, and implementation scope
11	deploy-and-onboard.md	concrete Azure resource inventory, bootstrap sequence, fork vs core split
11a	deployment-resource-conventions.md	deterministic CAF resource names, ownership tags, and deployment-supplied tag rules
11b	hyperscale-cell-architecture.md	scale-out blueprint for 300 subscriptions: cell-based streaming, policy-driven fan-in, two-plane logging, CQRS audit indexing over ADX, cost envelope, standard/sovereign profiles, Container Apps default (AKS deferred)
11c	control-plane-disaster-recovery.md	active-passive regional recovery profiles, fencing, state and event recovery, failback, and evidence gates
12	startup-and-lifecycle.md	cold start, day-zero catalog, observation-first rollout, discovery-loop kickoff
13	operating-and-verification.md	self-health signals, canary event, smoke tests, alert routing, runbooks
13a	observation-campaign.md	permission-aware periodic collection across registered inventory, activity, health, metric, log, network, cost, and recovery sources with local/deployed parity
20	deployment-preflight.md	pre-deployment feasibility and blocker collection: probe taxonomy, readiness report, blocker-to-terraform-toggle mapping
20a	preflight-active-reassembly.md	active plan reassembly: turn a policy blocker into a re-rendered terraform plan via capability-mode toggles, delivered as a fix PR through the executor (convergence loop, stop-conditions, limits)
20b	installable-deployment-cli.md	installable fdaictl facade: isolated uv installation, read-only preflight, signed deployment bundles, and exact-plan submission to the private runner
20c	provisioning-execution-profiles.md	provisioning profile selection: online/offline delivery, existing or managed execution hosts, access preference, workload identity, and exact-plan approval
20d	disconnected-deployment.md	deploying into a network without public egress: network profiles, internal mirrors, signed offline kits, degraded-evidence fallbacks, and the remaining air-gap gaps
20e	network-connectivity-matrix.md	scenario-specific DNS, IP, protocol, port, private-zone, PTU, APIM, and blocked-path behavior
21	assurance-twin.md	queryable ontology twin for architecture review / Q&A / assessment: text-to-query, proactive review, whole-graph what-if, observation mode proposals
22	operational-readiness.md	dev-to-ops handoff gate: ownership-transfer trigger, whole-scope RBAC / policy / reliability review, ReadinessReport, environment-promotion gate
22a	operator-initiated-sre-and-arb.md	non-incident identity, operator-initiated SRE response, live stage progress, ARB health/manual start, workflow enforce, and local/deployed parity

Cost, users, channels, risk, parity

#	Document	What it covers
14	cost-model.md	monthly cost envelope for the minimum resource inventory, T2 LLM cost split, traffic triggers
15	user-rbac-and-identity.md	human roles (Reader / Contributor / Approver / Owner + Break-Glass), Entra ID artifacts, console-to-PR identity flow
15b	agent-operational ownership-and-ownership handover.md	human <-> 15-agent ownership handover map: owns (accountable / informed), FDAI maintainers (min 1, rec 2), escalation chain, coverage + bus-factor
15c	agent-operational ownership-operations.md	production binding, stale-identity health, ownership handover draft PR, signed merge notification/audit, recovery and verification
15d	human-agent-assignment-and-knowledge-ownership handover.md	administrator identity search, governed IAM registration, primary and backup agent duties, approval non-response escalation, fatigue-bounded ownership handover goals, and agent-owned knowledge processing
15e	human-agent-assignment-implementation-plan.md	nine dependency-ordered main-branch work packages for operational ownership v2, assignment cases, console projection, ownership coordination, Entra apply, escalation supervision, ownership handover goals, knowledge processing, and production rollout
16	channels-and-notifications.md	non-web-UI channels (Teams / Slack / email / webhook / pager / SMS), category and trust-tier matrix
17	risk-classification.md	auto vs human approval vs deny classification: dimensions, initial rule table, environment detection
17b	escalation-and-standing-authority.md	what happens after a hil decision when nobody answers: the supervised OODA loop, the impact-tiered time-decaying escalation ladder (distinct from channel fallback), and standing authorization (pre-authorized, envelope-bounded, reversible-only conditional auto-action as a deterministic safety check input)
18	dev-and-deploy-parity.md	authoritative interactive local/deployed parity, explicit fixture profile, and deployer-scoped LLM gates
18a	console-read-boundary.md	server-owned read-source declarations, local authentication, workload evidence, and bounded inventory queries
19	operator-console.md	FDAI Console conversations across CLI, Teams, Slack, and web, with per-tool RBAC, LLM tiers, and session persistence
19k	operator-console-module-map.md	conversation module ownership, CLI/API route map, channel adapters, and core/delivery boundary
19l	operator-console-progressive-conversations.md	bounded read branches, ordered reduction, verified revisions, cancellation, replay, and progress metrics
19m	narrator-routing-and-latency.md	T1 narrator deployment routing, multimodal probes, operator preference, TTFT, web-search pooling, and runtime delivery decisions
19n	hierarchical-conversation-planning.md	non-keyword semantic decomposition, structural ontology query coverage, verified intent graphs, evidence joins, and answer boundaries
19o	ontology-query-coverage-implementation-plan.md	audited implementation gaps and dependency-ordered work packages for manifests, query plans, semantic generations, topology history, causal evidence, and legacy cutover
19j	console-operations.md	existing Operations navigation, source-specific task projections and schemas, operational requests, pantheon ownership, and execution separation
19f	console-evidence-and-resilience.md	console evidence provenance, localization, durable replay, stream recovery, and Architecture-map resilience
19a	document-ingestion.md	drop-zone UX, large and protected document handling, format extraction, private storage, shared visibility, retention, and deletion contracts
19g	conversation-attachments.md	protected Slack/Teams fetch, explicit ownership handover purpose, web chat document refs, image OCR, production binding, and security limits
19h	document-ingestion-agent-ownership.md	stage-to-agent ownership, typed pipeline objects, advisory-first promotion, audit, conflict, and rollback boundaries
19b	scheduled-result-continuations.md	scoped conversation anchors for exact scheduled runs, evidence provenance, channel threads, access, expiry, and delivery ordering
19c	skill-source-management.md	durable approved sources, quarantine, ETag refresh, disabled-first approval, and provenance-preserving revocation
19d	durable-conversation-delivery.md	verified cross-channel bindings, durable reply ledger, process-loss recovery, adapter controls, and read-only reliability metrics
19e	governed-trajectory-datasets.md	authorization-first observable trajectories, deterministic JSONL/checksums, quarantine, offline replay validation, retention/legal hold, and reviewed-only Norns intake
19i	benchmark-adapters.md	brand-neutral external harness contracts, installed-plugin injection, provider binding, bounded lifecycle, and benchmark authority boundaries
20	action-ontology.md	ActionType schema (fix + ops + governance), trigger axis, tier / role / prod / live-probe ceilings, fork override seams
21	execution-model.md	Unified RiskGate, six-axis authority matrix, three executor paths (PR-native / direct API / PR-manual), live-blast probe combinator, resolved_ceiling audit block
21a	recovery-and-chaos-enforcement.md	agent-executed recovery constrained by typed plans, impact envelopes, pre-authorized rollback, and continuous chaos guards

Agent organization

#	Document	What it covers
22	agent-pantheon.md	fixed 15-agent control-plane organization: closed-loop ownership, ontology-constrained accuracy, single-writer topics, typed pub/sub, conversational ports, ActionType roles, and bounded human escalation
22a	bounded-task-workers.md	isolated depth-one read-only investigations outside the fixed Pantheon: capability attenuation, bounded lifecycle, durable branch records, untrusted parent synthesis, and GET-only projections
22b	background-task-sessions.md	durable detached operator investigations: immediate creation, lease/CAS ownership, bounded progress, process-loss reconciliation, conversation handoff, and delivery boundary
22c	busy-input-modes.md	channel-neutral durable queue, interrupt, and safe-boundary steer modes for active web, Slack, and Teams conversations
22d	azure-read-investigations.md	exact resource resolution, typed Azure read evidence, measured direct/streamed/detached execution, dedicated reader identity, quotas, and durable completion delivery
22e	azure-resource-discovery-commands.md	ontology-aligned Azure resource discovery, ARG and CLI fallback coverage, sanitized reproduction commands, plan critique, and measured rollout
23	agent-workflows.md	the 13 cross-agent workflows the pantheon composes into product capabilities, including cost-aware fix, predictive scale, operational readiness handoff, scheduled governed Python tasks, and detection readiness assurance. Each has a trigger, sequence diagram, exit criteria, and promotion gate.
23f	agent-workflow-rollout.md	independent observation mode rollout order, per-workflow exit gate, dependency, and no-enforcement boundary
23b	process-automation.md	machine-readable counterpart to agent-workflows.md: the Workflow catalog schema (catalog-as-code under rule-catalog/workflows/), the Process ObjectType + targets / advances LinkTypes, the compile-to-Runbook control-loop wiring, saga compensation, and observation-first governance. A business process is an ordered list of ActionType steps the trust-router dispatches one at a time
23c	customer-workflow-automation-plan.md	delivery plan for adopting organizations: readiness baseline, six rollout waves, customer adapter boundaries, approval and recovery work, behavior simulation, promotion evidence, verification matrix, and production completion criteria
23d	operational-planning.md	event-driven specialist planning over immutable context, versioned logic assets, bounded sandbox and twin simulation, hard constraints, arbitration, governed execution, effect closure, and Planning Room projection
23e	operational-planning-hardening.md	implementation evidence, 12 adversarial hardening rounds, live observation mode proof, residual Low release risks, and merge-boundary verification for operational planning
23g	operational-hypothesis-loop.md	pre-action hypotheses, no-action baselines, independent outcome closure, causal refutation, active/challenger separation, and parallel worker ownership

Prompt subsystem

#	Document	What it covers
24	prompt-composition.md	evolving system prompt: role x layer matrix, tools / web search, debate orchestrator, recognition measurement
24b	hallucination-rubric-gate.md	subtractive rubric hallucination filter for T2: per-criterion judge scoring folded into confidence via min(), self-consistency sampler, observation mode-before-enforce promotion

Reporting subsystem

#	Document	What it covers
24b	reporting-subsystem.md	declarative visualization pipeline: YAML report catalog, datasource / widget / format registries, registered widget builders and datasource adapters over existing seams, pluggable encoders, read-only GET /reports/* routes, and fork extension recipes. The backend-only contract remains stable as registries grow.

Sequencing (cross-doc plan)

#	Document	What it covers
25	implementation-plan.md	historical compatibility record for the rejected 2026-07-06 standard-set identifiers and the tested M1.2 starter-probe set; current authority stays with subsystem owners
26	agent-pantheon-implementation.md	compact W0-W8 coordination summary, implementation ledger, and PantheonRuntime composition contract; roles and workflow rollout details stay with their focused owners
27	productization-and-extensibility.md	prioritized P0/P1/P2 status matrix for install and diagnostics, bidirectional channels, trusted extensions and MCP, model and scheduler resilience, security audit, typed APIs, and the capabilities intentionally kept outside the FDAI app shape
28	capability-licensing.md	signed capability entitlement for an image-delivered distribution: public key in the image, signed token in deployment config, availability-only authority, fail-safe degradation, and the honest tamper-evidence limit

Phase timeline

timeline
    title FDAI Delivery Phases
    P0 Instrumentation : KPI telemetry : Baseline vs reference agent : Unblock identity and policy
    P1 Rule Catalog and T0 : Normalize checklists : Policy-as-code gate : Auto fix PR : Out-of-band detection
    P2 Quality and T1 : Continuous rule update : LLM quality gate and mixed-model : Embedding pattern reuse : Observation mode to enforcement mode
    P3 Integrated Loop : Unified control loop : DR-Chaos scheduler and DB DR : FinOps auto-actions
    P4 Scale : Continuous measurement : Pattern-library and model tracking : Scalability : Multi-cloud expansion (TBD)

Phases are strictly sequential (P0 -> P1 -> P2 -> P3 -> P4) and each phase doc names its predecessor in a Dependencies section. Vertical coverage lands incrementally: Change Safety in P1; Resilience and Cost Governance in P3. Multi-cloud stays TBD in P4 (Azure-only implementation, see Implementation Focus).

Phase summary

The exit column is each phase’s primary gate; every phase doc lists the complete exit criteria and its dependencies.

Phase	Goal	Key deliverables	Primary exit gate
P0	Establish measurement and remove blockers	KPI dashboard, baseline report, identity / policy blockers resolved	reproducible baseline exists
P1	Deterministic core	rule catalog, T0 engine, policy gate, fix PRs	Change gate runs in observation mode
P2	Quality and lightweight tier	rule-update pipeline, LLM quality gate (guards T2), T1 similarity reuse	auto-resolution rate validated vs P0 baseline
P3	Integrated autonomy	unified loop, DR / chaos scheduler, cost auto-actions	autonomous MVP across all 3 verticals
P4	Scale out (Azure)	continuous measurement, pattern-library and model tracking, scalability; multi-cloud adapters TBD	guard metrics stable on the Azure baseline

Guardrails applied throughout

• Measurement first: no autonomy without telemetry; no multiplier or coverage claim without a measured baseline.
• Observation mode before enforce: every new action ships judge-only, then is promoted per-action explicitly; regressions demote automatically.
• Choose the safer default when the outcome is uncertain: low confidence, verification failure, or budget / rate overflow degrades to human approval, never to an ungated auto-action.
• Seven safeguards on every autonomous action: stop-condition, rollback, impact scope limit, dry-run, per-resource lock, idempotency, and audit record (security-and-identity.md).
• Separation of duties: approval and execution are distinct principals; the console is non-privileged and never receives the executor identity (security-and-identity.md).
• Bilingual, customer-agnostic artifacts: English and Korean are first-class prose languages; identifiers and machine records remain stable as defined by the language contract (language.instructions.md).

Next steps

To do this	Start here
Review FDAI’s highest design authority	fdai-constitution.md
Understand the 3-tier control loop	architecture.instructions.md
Find a subsystem’s source, tests, and design doc	architecture/code-map.md
See the concrete Azure resource inventory	deploy-and-onboard.md
Establish the P0 measurement baseline	phases/phase-0-instrumentation.md
Read the safety rules on every autonomous action	../../.github/instructions/coding-conventions.instructions.md
Contribute a new rule to the catalog	../../rule-catalog/RULE_AUTHORING_GUIDE.md

Canonical diagrams

To keep design docs from drifting, a small number of diagrams are named as canonical. Downstream docs MUST link to them instead of redrawing the same shape.

Diagram	Canonical location
Control loop (event -> tier -> gate -> action -> audit)	architecture.instructions.md § Control Loop
Agent pantheon (15-agent org chart)	agents/agent-pantheon.md
Monorepo layout	architecture/project-structure.md § Monorepo Layout
Subsystem index (source -> tests -> docs)	architecture/code-map.md

A doc that needs a different view of these concepts SHOULD use a domain-specific mermaid rather than paraphrase the canonical shape. If the canonical diagram itself needs to change, edit it once at the canonical location; the roadmap review picks up the change from there.