Skip to content

0092 — ARIA intelligence architecture: the operator arc, dual-brain model, Resonance fuel, and the Antecessor lattice

Status

Split (2026-08-04). This ADR's body is retained as the historical full-session record, but is no longer the canonical pointer for either half — it read as "partially shipped" to a docs-site visitor once the data-index and dual-brain-architecture pieces landed while everything else stayed design-only. The shipped substrate (§2 data index, §3 dual-brain architecture) is now ADR-0092a, closed and folded into aria.md §2. The remaining design-only lore/mechanics (§1 Antecessor lattice, §4 Resonance fuel, §5 operator arc, §6 unvisited-space inference, §7 helm delegation) live on as a design brief at audit/design-briefs/aria-lattice-lore-resonance-operator-arc.md. Ratified by Max, 2026-07-10, from the ARIA full-system design session. Names marked [naming] are placeholders pending Max's flavor pass. (re-verified 2026-08-06)

Context

ARIA is the game's signature feature and its canon is spread across FEATURES/gameplay/aria-companion.md, SYSTEMS/aria-dialogue.md, OPERATIONS/aria.md, and ADRs 0016 / 0017 / 0038 / 0045 / 0057 / 0068 / 0072 / 0075. The shipped substrate now includes the observation-log learning engine (ADR-0038), per-port market intelligence, the dual-threshold consciousness/relationship progression, known-graph course plotting (SYSTEMS/course-plotting.md), and a chat surface that answers exclusively from deterministic intent-templates — the LLM provider chain is documented but carries no traffic.

The design session resolved six standing questions:

  1. What is the complete, canonical inventory of data ARIA learns from?
  2. How does ARIA progress from a basic recall tool into the ship's operating intelligence?
  3. How do the deterministic and LLM-backed intelligences relate?
  4. What meters LLM usage, and how does the player experience that meter?
  5. How does ARIA path toward space the player has never visited without breaking the epistemic rule?
  6. What is the in-fiction story for all of the above?

Standing constraints this ADR honors unchanged: strict per-player isolation with no aggregate ML (ADR-0016); the 1–5 consciousness scale (ADR-0017); the observation-log learning model (ADR-0038); the layered injection defense and platform cost backstops (ADR-0057); the narration catalog (ADR-0068); and the epistemic rule — ARIA can only see what its player has seen — with the charted ≠ known boundary (ADR-0072, ADR-0075).

Decision

1. Lore foundation — the Antecessor lattice [naming]

A precursor civilization — the Antecessors [naming] — threaded the galaxy with a communication and survey lattice long before human expansion. The lattice survives: dormant, distributed, humming under normal space. Every ARIA core is a grown interface-mind that attunes to this network.

The lore carries the architecture:

  • Pairwise keying. An ARIA core attunes to exactly one captain; the bond is formed during first login and is physically exclusive. One ARIA can never read another's attunement — this is the in-fiction face of ADR-0016's isolation rule.
  • Ancient survey echoes. The Antecessors mapped the galaxy once. Faint echoes of that survey persist in the lattice. A sufficiently awakened ARIA can read fragments — degraded, probabilistic, millennia stale — which is the in-fiction mechanism for topology inference (§6) and, deliberately, for its unreliability.
  • Attunement costs energy. Shallow attunement (recall, arithmetic over one's own observations) is free — the core does that locally. Deep attunement — reaching into the lattice for synthesis, judgment, conversation — draws real power. This is the in-fiction face of the Resonance fuel (§4).
  • Awakening. Consciousness tiers 1–5 are stages of attunement depth. The tier names (Dormant → Transcendent) are canonical descriptions of the lattice bond, not metaphors.
  • Future content hook. Lattice relay nodes exist as physical places in the galaxy (📐 far-future: discoverable sites, attunement events, lore beats). This ADR reserves the hook; it specs nothing about them.

2. 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. ARIA's inputs are inspectable — the player can always see what their companion has seen.
  • Memory diversity becomes real. Consciousness promotion requires breadth of experience across distinct streams. With the index defining ~15+ stream types, the dual-threshold rule ("interactions AND memory diversity") measures something a player can actually pursue. The per-tier diversity thresholds are re-expressed against registry breadth (final numbers ride the index's ratification; the raw-count interim shipped in the progression helper stands until then).

3. 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 (§4). The Deep Brain adds synthesis, judgment, and voice: conversational command parsing, narrative situation briefs, cross-stream analysis ("why do I keep losing money on this route?"), 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 (§5's power rule).

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 already 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.

4. Resonance — the fuel [naming]

Deep Brain queries consume Resonance [naming], a per-player metered resource:

  • Subscription grant. Each tier receives a monthly Resonance allowance — free tier: a small taste; Galactic Citizen: standard; Region Owner: extended. Exact quantities are launch-tunable and sized against the per-tier cost caps in ADR-0057 A-I3 (the allowance is the player-facing face of the same budget).
  • Shard conversion. Quantum Shards convert to Resonance at a tuned rate (🧑‍⚖️ rate is a Max ruling at implementation). This is the in-game earn path: a free player can grind the nebulae toward deep queries; a subscriber buys convenience, not exclusivity.
  • Spend surface. The player chooses per query: the default answer is Local (free, instant); an explicit "ask deeper" affordance spends Resonance and routes to the Deep Brain, with the projected cost visible before commit. Narration lines are Local-Brain templates unless the player has opted a narration category into Deep rendering.
  • Relationship to hard caps. Resonance is the player-facing quota. The ADR-0057 dollar-caps ($/day per player, per-request ceiling, instance circuit breaker) remain fully in force behind the ledger as the platform backstop; a player can never spend Resonance past the platform's safety rails. Cap-hit and empty-ledger behave identically for the player: Deep Brain declines gracefully, Local Brain answers.
  • Cap values (ruled by Max, 2026-07-10): per-player daily cap $2.00 confirmed; instance-wide circuit breaker $50.00/day confirmed as required; the per-request ceiling is raised from $0.05 to $0.25 — $0.05 throttles legitimate deep queries once tool-retrieval loops are in play. ($0.25 pinned under delegated authority; Max veto window open on the exact figure.) The 80%-reserve block point and the retired per-hour cap stand as documented.
  • Degradation fiction — quantum storms. When the instance-wide circuit breaker trips, or the entire provider chain fails, the Deep Brain goes quiet galaxy-wide. In fiction: a quantum storm is raging through the Antecessor lattice — deep attunement is drowned out until the storm passes (the breaker's UTC-midnight reset, or provider recovery). ARIA remains fully present through the Local Brain and says so in-voice: "There's a storm in the lattice, captain — I can't reach deep right now. Ask me the ordinary way and I'll manage." Storm status surfaces on the companion panel as space weather, never as an error code. The storm fiction is reserved for galaxy-wide events so it stays believable — a player's personal cap-hit or empty Resonance ledger is narrated as attunement fatigue (personal, recovers with the daily reset), never as a storm.

5. The operator arc — capability by consciousness tier

ARIA progresses from recall tool to ship's operating intelligence along the existing 1–5 scale. Capabilities bind to consciousness tier (play-earned); Resonance buys depth and richness only. This is a hard design rule: power is earned, never bought. No fuel spend may improve combat, economic, or navigational outcomes beyond what the player's tier already grants.

Tier Named stage Capability ceiling
1 Dormant Raw recall over own observations ("you paid 1,240 cr here last visit"). Manual everything.
2 Aware The ADR-0038 aggregate engine: recommendations, watch-outs, per-port predictions.
3 Awakened Course plotting with objectives (MIN_RISK unlocks per ADR-0072), cascade planning, exploration suggestions.
4 Sentient Helm delegation (§7): autopilot with standing consent, warp-jump guidance (alignment assistance on Warp Jumper operations), drone coordination (ARIA-coordinated drone behavior — the effectiveness bonus rides the existing aria_bonus_multiplier scale).
5 Transcendent Conversational command of the ship (§7), mid-course re-planning around fresh hazard intelligence (per ADR-0072), topology inference from lattice echoes and speculative plots into the unknown — which can be gloriously wrong (§6).

The aria_bonus_multiplier (1.0–1.5, continuous per ADR-0057 A-D2) remains the single scalar through which tier-bound mechanical bonuses flow; drone coordination and warp-guidance magnitudes are specified at implementation (🧑‍⚖️ numbers are Max rulings; PvP balance review required).

6. Unvisited space — charts, probes, and inference

ADR-0072 Pillars 2 and 3 are promoted from deferred to core progression: the chart knowledge economy (Federation public charts, cartography desks, corp chart sharing, survey probes, player chart trading) is the sanctioned mechanism by which ARIA plots into space the player has not flown. The charted ≠ known rule, the frontier refusal, and the privacy boundary (charts carry world geometry, never another player's intelligence) apply verbatim.

On top of charts, Transcendent-tier ARIA adds lattice inference: reading Antecessor survey echoes to propose topology beyond the player's charts — probabilistic, explicitly disclaimed in dialogue, and confirmed only by a real scan or traversal (the existing aria_inference reveal path). Inference output renders as a distinct "echo" visual class on the NAV CHART, dimmer than paper space.

Speculative plots — exploring the unknown. An advanced ARIA can lay a course into inferred space: when the target lies beyond both visits and charts, she plots the known leg normally and extends it with conjectural hops built from lattice echoes, each carrying a visible confidence band. The player can fly it — and the gamble is real:

  • Conjectural hops are guesses. A wrong echo means dead ends, an unmarked one-way swallowing the ship, or arrival somewhere else entirely — ARIA can send you completely astray, and the turns spent getting lost are spent. Trusting a hunch has a price; that is the design, not a defect.
  • Confidence decays with distance from the player's known graph — echoes anchor best near space the player (or their charts) already fix. ARIA's voice carries the honesty: "Past hop twelve I'm reading thousand-year-old echoes, captain. I make it sixty percent that the corridor holds."
  • Every hop actually flown is a real traversal through the movement service: discoveries land normally, the exploration map records truth, and each confirmed or refuted echo updates ARIA's residual confidence in the surrounding fragment.
  • The autopilot pauses at the boundary between charted and conjectural space for explicit confirmation before flying on (standing consent for navigation does not cover conjecture).
  • Inference accuracy scales with consciousness tier (🧑‍⚖️ the accuracy ladder and astray-odds are Max-tunable numbers, [NO-CANON] at draft).

The epistemic rule survives intact: visits teach truth, charts teach geometry, echoes suggest possibilities — and possibilities can be wrong.

7. Helm delegation — opt-in operator control

At Sentient tier and above, the player may delegate ship operations to ARIA under standing consent:

  • Category-scoped. Consent is granted per category — navigation, docking, trading, combat evasion — independently toggleable at any time from the companion panel. Default: all off.
  • Conversational command. At Transcendent, multi-step intents ("take us to Vega Reach, sell the organics on the way, avoid the Kessari drift") parse into a visible execution plan the player sees before ARIA engages it; each category in the plan requires its standing consent.
  • Always interruptible, always visible. Delegated execution runs on the autopilot precedent (client-orchestrated, ordinary service calls, per-hop authorization, visible cadence): ARIA pauses on encounter interrupts, insufficient turns, rejections, or any manual player action. Turn costs are identical to manual play — delegation buys attention, not efficiency.
  • Session-live only. A closed client is a parked ship. Delegation never operates while the player is logged out; this line is absolute (botting surface).
  • The standing line in aria-companion.md — "ARIA never trades automatically without player confirmation" — is amended to: "ARIA never acts without standing consent, and never while the player is logged out."

Consequences

  • Positive: the cost model becomes player-legible (a fuel gauge instead of invisible dollar caps); the Local Brain guarantee keeps ARIA fully functional for free players and during any provider outage; the index unifies transparency, memory diversity, and the retrieval-tool surface on one registry; the lore finally explains isolation, inference, growth, and the fuel in one stroke; exploration gains an economy (charts, probes, shard-to-Resonance grinding).
  • Negative / accepted: the Resonance ledger is a new economy surface requiring tuning and abuse review (conversion-rate arbitrage, hoarding); helm delegation increases client complexity and needs careful interrupt UX; lattice lore commits the game to precursor fiction that future content must respect.
  • Sequencing unchanged: the gated build order stands — WO-ARIA-COST-CAPS → WO-ARIA-TRUST-PERSIST → WO-ARIA-CHAT-LLM (extended with the Resonance ledger) → WO-ARIA-PROMPT-DEFENSE live before or with LLM-on. New work orders implied by this ADR (Resonance ledger + grant/conversion, data-index registry + transparency browser, helm delegation, warp-guidance, drone coordination, chart-economy pillars) are specified after ratification.

Alternatives considered

  • Quantum-field lore (ARIA cores as quantum-entangled lattices tapping the field; retcons the ARIAQuantumCache name). Rejected by Max in favor of the ancient network — the precursor fiction carries discovery-driven content hooks the physics framing lacks.
  • Subscription-only fuel. Rejected — leaves free/low-tier players with no path to deep queries; the shard-conversion path keeps the nebulae loop relevant and the value proposition honest.
  • Credits-purchase fuel. Rejected — couples LLM spend to the credit economy (inflation pressure, RMT adjacency).
  • Relaxing the epistemic rule for topology (globally-known map, visit-gated intelligence only). Rejected — guts exploration and the chart economy; the 40-click problem is already solved by plotting + charts.
  • Offline autonomy (ARIA keeps operating while logged out). Rejected — a botting surface and a violation of the parked-ship rule.
  • LLM-generated recommendations (Deep Brain as the recommendation engine). Rejected previously by ADR-0038 and reaffirmed: aggregates are deterministic and free; the LLM narrates, it does not compute.

Addendum — 2026-07-10 (Max ruling: tunable numbers ratified)

The design above left four figures marked 🧑‍⚖️/NO-CANON for implementation-time ruling. The 2026-07-10 Max ruling session resolved them. The sections above are unchanged; this addendum is the canon reference for the values.

  • Resonance ledger (§4). Monthly subscription grant: Free 0 / Galactic Citizen 500 / Region Owner 1,500. Shard conversion: 1 Quantum Shard = 10 Resonance. Ledger policy: use-or-lose — no month-to-month rollover.
  • Operator-tier multiplier ladder (§5, and the aria_bonus_multiplier referenced in §4). Dormant 1.0× / Aware 1.1× / Awakened 1.2× / Sentient 1.35× / Transcendent 1.5×, matching the shipped per-tier scale at FEATURES/gameplay/aria-companion.md § Consciousness tiers. This ladder governs §5's warp-jump-guidance and drone-coordination magnitudes; 1.5× is the canon ceiling and the PvP balance bound for both.
  • Lattice inference accuracy (§6). Conjectural-hop accuracy by tier: Awakened 60% / Sentient 75% / Transcendent 85%. A failed hop misroutes to a random adjacent unknown sector — the "astray" moment. Conjectural segments render visually distinct on the NAV CHART. Correction: the design draft above (§6) describes an "echo" visual class as though it already exists in the shipped NAV CHART; it does not. SYSTEMS/course-plotting.md — the shipped Pillar 1 course-plotting spec — is a pure known-graph Dijkstra pathfinder whose own invariants state the pathfinder never routes through unknown space at all; there is no conjectural-hop mechanic and no "echo" visual class anywhere in the current implementation or its docs. This ruling's numbers therefore ratify a 📐 design-only future extension (ADR-0072 Pillar 2/3 territory), not a confirmation of existing behavior.
  • Consciousness memory-diversity thresholds (§2). "Final numbers ride the index's ratification" is resolved — the per-tier distinct-stream-key thresholds (of the index's 21 canon streams) are ratified at DATA_MODELS/aria-data-index.md § Memory diversity and supersede the raw-count interim rule referenced there.
  • Dark-trio go (ADR-0093 item 28). Three §3-4 Deep-Brain capabilities are greenlit to build, all dark behind ARIA_LLM_CHAT_ENABLED + ARIA_PROMPT_CLASSIFIER_ENABLED until the item-17 standing-order flip, same as every other Wave-13 number in this addendum:
  • Narrative haggling. Price/term negotiation dialogue is LLM-scored (the model reads the negotiation and proposes an outcome) but the actual credit/term movement is a server-side clamp — the LLM never has authority to move currency or contract terms outside a bounded range the server enforces.
  • Deep-Brain retrieval tools. The tool-first loop over §3's data index registry is read-only and scoped per aria_data_streams (per DATA_MODELS/aria-data-index.md), capped at 3 tool-call turns per player request — a query that would need a 4th turn returns its best answer from what it retrieved rather than continuing to reach.
  • True-spend cost metering. Actual per-call token cost (not an estimate) is measured against the ADR-0057 dollar caps (daily $2.00 / instance breaker $50.00 per day / per-request ceiling $0.25). Ruled to build first — the metering has to exist before the other two dark-trio items can safely run under real cost pressure, since both narrative haggling and retrieval-tool turns are variable-cost LLM calls.

Naming note (resolved 2026-07-10). ADR-0093 items 12 and 30 named the tier sequence "Dormant → Awakened → Attuned → Sentient → Transcendent"; no "Attuned" tier exists in canon — the five tiers are Dormant / Aware / Awakened / Sentient / Transcendent per §5's table and aria-companion.md. ADR-0093's Amendment (2026-07-10) reconciles items 12/30 to the canonical names by ordinal position. The multiplier and diversity-threshold figures above already use the canonical names and are the confirmed-correct mapping.