Engineered against hallucination.

Every Coach verdict is produced by a council of specialist agents running the classic Blackboard architecture, adapted to large language models. The blackboard is a shared, versioned workspace. Every agent reads the whole board; no agent holds private state.

The board is written in six strictly ordered layers. First, the scenario layer pins the input: your hero, position, and both lineups, so every downstream agent argues about the same frame. Second, the facts layer is written by deterministic enumerators, ordinary code that reads the entire knowledge base and renders threat profiles, ally synergy matrices, talent matrices, and item counter matrices. No language model touches this layer, which means the numbers in it cannot be hallucinated. Third, the specialist round: item, skill, and talent specialists write typed, citation-bearing proposals in parallel. Fourth, a coherence checker, again pure code, cross-examines the proposals and writes flags for contradictions. Fifth, specialists respond to the flags with revisions that must cross-reference one another. Sixth, an adversarial Critic audits the revised board and writes formal objections. Only then does the head coach synthesize the final verdict from the full board state.

Every write to the board is parsed and validated through Pydantic schemas. A specialist that returns malformed output gets exactly one retry; a second failure is discarded rather than trusted. The board renderer is pure and deterministic: the same board state always renders to byte-identical text, which is what makes the whole pipeline reproducible and auditable.

Language models answer trivia from training memory, which is where hallucination lives. Quorum does not. Every number a coach cites, cooldowns, mana costs, recipe components, talent values, cast ranges, is retrieved at answer time from a structured knowledge base of the live patch: 127 heroes, 3,084 abilities, 501 items, and 1,068 talent values, assembled from authoritative game data and overlaid with per-patch corrections.

Enumeration questions, such as which abilities apply a given mechanic, never rely on model recall. They execute a deterministic search across the full ability index and return the complete result set, not a sampled or truncated one. If the knowledge base does not contain a fact, the council says so instead of inventing one.

When a Dota patch lands, the knowledge base is rebuilt and audited within 48 hours. The audit is not a vibe check: automated consistency passes verify recipe graphs, talent trees, and ability field integrity before the new patch data is allowed to serve a single answer. Until the audit passes, Quorum keeps serving the previous verified patch and says so in the interface.

Chat's Immortal Mode is anchored to a snapshot rebuilt every day from the game's top public bracket, a rolling weekly window of more than 100,000 matches. For each of the 127 heroes the build compiles pick, win, and ban rates, complete item purchase tables with win rates and average purchase minute, boot choices, talent picks mapped to the 10, 15, 20, and 25 tiers, typical skill-max order, and position distribution with per-position win rates, alongside professional-scene picks and bans. Every section of the snapshot is labeled with its bracket of origin, so the council can be precise about provenance when it cites a number.

The snapshot is versioned and served from our own API. The desktop client caches it and refreshes once a day. Freshness is enforced in code, not by intention: past 3 days without a successful refresh, the tool reports the live meta unavailable rather than let the model cite a dead one. A missing number beats an invented one.

Immortal answers draw twice the metered QI. The multiplier is applied by the same append-only ledger as every other charge, with the same deliver-or-refund rule.

Every answer is metered in QI, and the meter is engineered to be fair in both directions.

Admission control happens before any model is invoked: the proxy checks your pass state, mode access, and QI floor, and refuses cleanly if the turn cannot be afforded. A refusal at the door costs nothing.

Once a turn is admitted, it is guaranteed to finish at full quality. If your balance runs dry mid-turn, Quorum absorbs the overage rather than degrading or truncating the answer. Mid-flight requests carry a per-turn session identifier, so continuation calls of an admitted turn are never refused.

If the system fails to deliver a finished answer for any reason on our side, the QI charged for that turn is refunded automatically. Refunds are written as compensating entries in the same ledger as charges and are idempotent: two concurrent refund attempts for the same turn cannot double-credit, because the second one sees the first one's compensating row and stops.

Every model invocation is recorded as an immutable usage event: mode, agent, token counts, QI charged, and session identifier. Balances are derived from this ledger and reconciled against it; the ledger, not the wallet column, is the source of truth.

Concurrency is handled at the database, not by hope. A Coach run fans out several agents in parallel, so multiple requests routinely finish against the same starting balance. Every deduction and refund first re-reads the license row under a row-level lock, PostgreSQL's SELECT FOR UPDATE, so balance mutations serialize on committed state and the classic lost-update race is structurally impossible. The lock is held for milliseconds at commit time, never across a model call.

We run reconciliation audits that recompute every wallet from the ledger and flag any drift. This is the same discipline banks apply to accounts, applied to game coaching.