Phases Overview

High-level roadmap from Foundation through Advanced Retrieval. Detailed milestones for completed phases live under their respective directories.

Phase	Name	Status	Entry	Exit (summary)
0	Foundation	Complete	Empty repo	CI, docs, compose, proto source, Go skeletons
1	Core Platform	Complete	Phase 0 complete	Migrations, auth, proxy platform, security gate
1.5	Public Web Product	Complete	Phase 1 complete	Unified site at ibexharness.com
2	Single Provider E2E	Complete	Phase 1.5 exit	Authenticated OpenAI-compatible chat through proxy
2.5	Provider Generalization & Foundation	Complete	Phase 2 exit	Multi-provider, tokenizer, response pipeline, embedder, schema pre-work, MCP skeleton, LGTM exit
3	Core Memory Substrate	Planned	Phase 2.5 exit	Memory schema v2 (HNSW), write pipeline, read/ranking, hot cache
3.5	Extraction & Context Assembly	Planned	Phase 3 exit	Extraction worker, context assembly on proxy hot path, MCP memory tools
4	Operator Platform & Multi-Provider	Planned	Phase 3.5 exit	Management API, dashboard, multi-provider routing, hierarchical rate limiting
4.5	Intelligence Layer	Planned	Phase 4 exit	Behavioral fingerprinting, drift detection, directive regression testing
5	Advanced Retrieval & Graph Memory	Planned	Phase 4.5 exit	Hybrid search (dense+sparse+reranking), graph traversal, evaluation harness

Phase 0: Foundation

Status: Complete

Delivered: Documentation-first monorepo, governance CI, Buf proto foundation, local compose, toolchain (Makefile, TOOLCHAIN.md), Go auth/proxy skeletons with health/readiness/metrics.

See phase-0-foundation.

Phase 1: Core Platform

Status: Complete

Goals: goals

See phase-1-core-platform and phase1-exit-audit.

Phase 2: Single Provider E2E

Status: Complete

Exit milestone: 2.6.2 Phase 2 exit gate

Goals: goals

Phase 2.5: Provider Generalization & Foundation

Status: Complete

Exit milestone: 2.5.G7 Phase 2.5 exit gate

Goals: goals

Seven tracks: multi-provider adapters (Anthropic, vLLM), tokenizer registry, response pipeline, pluggable embedding service, schema pre-work (temporal validity, multi-label), MCP server skeleton, and exit/ops pack-up (local LGTM + verify-phase25).

Phase 3: Core Memory Substrate

Status: Planned (redesigned)

Goals: goals

Five tracks: memory schema v2 (HNSW, temporal validity, multi-label), vector store abstraction with composite scoring v2, write pipeline (PII, dedup, temporal conflict), read/ranking with category-conditional decay, and exit gate.

Phase 3.5: Extraction & Context Assembly

Status: Planned

Goals: goals

Six tracks: worker foundation, extraction intelligence, context assembly engine, proxy hot-path integration, MCP memory tools, and exit gate. This is where "the proxy that learns" becomes true end-to-end.

Phase 4: Operator Platform & Multi-Provider

Status: Planned (redesigned)

Goals: goals

Five tracks: management API, hierarchical rate limiting (Redis Lua), multi-provider adapters/routing/resilience, operator dashboard, and exit gate. Merges the original Phase 3 management API/dashboard with the original Phase 4 multi-provider scope.

Phase 4.5: Intelligence Layer

Status: Planned

Goals: goals

Four tracks: behavioral fingerprinting with distribution-aware statistics, drift detection with calibrated thresholds, directive regression testing with LLM-judge evaluation, and exit gate.

Phase 5: Advanced Retrieval & Graph Memory

Status: Planned (redesigned)

Goals: goals

Four tracks: hybrid retrieval (BM25 + RRF + cross-encoder reranking), graph-aware retrieval (multi-hop traversal, contradiction filtering), evaluation harness upgrade (gold set v2, shadow evaluation), and exit gate. Org-wide production hardening remains future roadmap scope beyond Phase 5.

Critical path (dependency order)

Phase 0 → 1 → 1.5 → 2 → 2.5 → 3 → 3.5 → 4 → 4.5 → 5

Pivot triggers and discoveries: findings.