Skip to content

0092a — ARIA data index and dual-brain architecture (shipped substrate, split from 0092)

Status

Folded into OPERATIONS/aria.md §2 (Architecture) and DATA_MODELS/aria-data-index.md — both items shipped and closed (2026-08-04). Split out of ADR-0092 §2-3; design-only remainder lives at audit/design-briefs/aria-lattice-lore-resonance-operator-arc.md. (re-verified 2026-08-06)

Context

See ADR-0092 §Context for the full design-session background — unchanged, not repeated here. The two decisions below are the parts of that session that shipped.

Decision

1. The canonical data index

Everything ARIA learns is enumerated in one place: DATA_MODELS/aria-data-index.md — the ARIA data index, a registry of observation streams (sector visits, port catalogs, price observations, trades, combat events, drone and mine encounters, NPC and player sightings, warp discoveries, and the rest). Each stream defines its trigger, payload, storage, retention class, and transparency surface.

The index is registry-driven: an aria_data_streams catalog table mirrors the doc, and adding a data point ARIA tracks is a registry row, not a code edit. The memory_type taxonomy derives from the stream registry — one memory type per stream.

Two consequences are intentional:

  • The transparency surface reads the registry. The player-facing raw-data browser (the "memory journal" in aria-companion.md) enumerates streams from the registry, so every new stream is player-visible by default.
  • Memory diversity becomes real. Consciousness promotion requires breadth of experience across distinct streams — the per-tier diversity thresholds are ratified at DATA_MODELS/aria-data-index.md § Memory diversity (per the 2026-07-10 addendum below).

2. Dual-brain architecture

ARIA has two brains; the split is canonical and permanent:

  • The Local Brain — deterministic, always available, free, and requiring no LLM: the SQL-aggregate recommendation engine (ADR-0038), course plotting, the narration template kernel (ADR-0068), intent-classified template dialogue, and every gameplay hook. The Local Brain is the primary intelligence, not a degraded mode. Every ARIA capability that affects gameplay outcomes runs here.
  • The Deep Brain — the LLM provider chain (per SYSTEMS/aria-dialogue.md), invoked only when the player chooses it and only with Resonance to spend (design-only budget, see the split-out design brief). The Deep Brain adds synthesis, judgment, and voice: conversational command parsing, narrative situation briefs, cross-stream analysis, and personality depth scaled by relationship tier. On provider failure, fuel exhaustion, cost-cap block, or security block, every Deep Brain surface falls back to the Local Brain answer — never to an error.

The player-facing contract: every question has a free answer. Deep answers are richer, never required. The Deep Brain never holds exclusive gameplay-affecting capability.

Retrieval follows the tool-first model: the Deep Brain reaches into the player's local dataset through read-only retrieval tools scoped by the data index registry (per the tool loop specified in SYSTEMS/aria-dialogue.md), pulling the specific streams a query needs rather than pre-stuffing context. Tool access is bounded by the same per-player isolation as every other read path.

Note on Deep Brain traffic: the provider chain and dual-brain split are shipped as architecture — the Local Brain and its fallback contract are live in production today. Whether the Deep Brain currently carries traffic is a separate, config-gated question (ARIA_LLM_CHAT_ENABLED) tracked in SYSTEMS/aria-dialogue.md, not a gap in this architecture decision.

Consequences

  • Positive: the index unifies transparency, memory diversity, and the retrieval-tool surface on one registry; the Local Brain guarantee keeps ARIA fully functional for free players and during any provider outage.
  • Negative / accepted: none beyond what ADR-0092's original Consequences section already named for these two pieces.