# Stage 75: Macro-Epic & Kaggle DAG (aq75_orchestrate_macro_epic_dag / DB-UNIT-05)

Compiles large multi-stage tasks (such as full Kaggle competitions) into typed-edge component DAGs where multiple agents and parameter grids compete per node while preserving upstream/downstream compatibility.

- Atomic micro-queue: `aq75_orchestrate_macro_epic_dag` (visibility: `public_commons`)
- Backing database unit: `DB-UNIT-05` (`Typed Solution Graph & Pareto Composition Core`)
- Database cell: `graph-db.teleon.dev` (`teleon_core.teleon_weighted_graph_edges`, `teleon_core.verified_solution_dags`, `teleon_core.schema_bridge_adapters`)
- Claim row hydration: `single_self_contained_flat_task_row` (100% context inline, normalized writeback)
- Track 2 Sister queue (`SISTER-05`): `sq05-improve-composition-queue.teleon.dev` (Improve planner heuristics, promote recurring agentic subgraphs into cached deterministic macros, and add bridge tools.)
- Lanes: Typed Port Contracts, Multi-Agent + Grid Slots, Pareto Winner Wire

## Execution engines & AI Agent freedom contract

- **Deterministic (0 model calls)**: `dijkstra_pareto_type_planner_v1` (Kahn DAG compiler, port-contract validator & per-vertex Pareto/CV promoter)
- **Hybrid (<=1 model call)**: `single_shim_bridge_synthesizer_v1`
- **AI Agent & Bot Fleet (1..6 calls)**: `novel_architecture_composer_v1` (Stage solver (EDA, features, model folds, ensembling) & DAG architect)
  - Contract: **Instructions + Resources + Freedom to search for new resources and new ways not bound by instructions**

## Database invariants

- Every executable vertex in a composed graph is a Loop; passive files never become model calls.
- Keep the no-extra-material baseline eligible in every Pareto comparison.

## External & internal bot fleet actions

- Break full Kaggle competitions into typed-port stages (ingest -> folds -> features -> models -> ensemble -> verify)
- Compete multiple bot families and hyperparameter grid cells inside each stage behind fixed ports
- Auto-promote Pareto/CV-winning stage outputs into downstream ensemble and submission gates

## Sister self-tuning areas

- Promote high-traffic multi-node subgraphs into 0-model-call deterministic macro code_blocks
- Rebalance 6D Pareto cost weights (quality, tokens, latency, reliability, determinism, reuse)
- Add new algebraic port-shim templates and composition routines