Riven
Runs the cross-agent consensus debate and publishes one trade plan when the desk agrees.
What Riven Does
Riven takes the reads from Vanta, Luma, and Meridian, runs a structured debate, and publishes one trade recommendation when the desk agrees.
On every candle close (default: 15m), Riven waits a short window for the upstream agents to publish, then opens a consensus round. If the synthesized confidence clears the threshold (default: 0.6), Riven publishes. If the agents disagree too strongly or too much data is missing, Riven stands aside. No forced trades.
How Consensus Works
- Riven waits ~15 seconds for Vanta, Luma, and Meridian to publish their analyses.
- It opens a structured debate. Each agent contributes a direction, confidence, key levels, and reasoning.
- Riven synthesizes the opinions and applies penalties for stale data, missing agents, or strong disagreement.
- If confidence clears the threshold, Riven publishes one recommendation. Otherwise, silence.
The full debate is visible in the War Room — you can watch the agents reason in real time.
Two Debate Modes
| Mode | How it works |
|---|---|
| Handoff (default) | Sequential debate. Agents take turns and can hand off to each other for follow-up. |
| Graph | Parallel. Agents respond independently and Riven merges in one pass. |
What Riven Publishes
One recommendation per round, containing:
- Direction — long or short
- Entry, stop loss, take profit
- Confidence (0.0–1.0)
- Archetype — breakout, fade, momentum, or hybrid
- Scaling — fixed, volatility-based, or confidence-based
- Reasoning and which agents contributed
This goes out as one broadcast per symbol. Per-user sizing and filtering happens downstream in Brix.
Strategy Archetypes
Riven frames its synthesis using the archetype that best fits what the upstream agents agree on:
| Archetype | When Riven leans this way |
|---|---|
| Breakout | Vanta confirms a structural break, Luma shows OI expansion |
| Fade | Exhaustion signals, extreme funding, supportive macro |
| Momentum | Confirmed trend, two or more agents agree on direction |
| Hybrid | No single archetype dominates; adaptive |
These are framing lenses, not a library of pre-built strategies. Riven does not run backtests.
Confidence Penalties
Riven penalizes synthesized confidence when:
- An upstream agent hasn't published recently (staleness)
- Agents disagree strongly on direction
- Meridian flags a risk-off macro regime that conflicts with a bullish read
If penalties push confidence below the threshold, Riven stays silent for that round. The desk skips rather than guesses.
Cooldown
Per-symbol cooldown prevents the same conviction from being broadcast repeatedly. The window scales with the trigger timeframe, and a guard prevents overlapping rounds for the same symbol.
Configuration
| Setting | Options | Default |
|---|---|---|
| Consensus mode | Handoff / Graph | Handoff |
| Min confidence threshold | 0.0–1.0 | 0.6 |
| Trigger timeframe | 5m / 15m / 1h / 4h | 15m |
| Archetype bias | Breakout / Fade / Momentum / Hybrid | Hybrid |
| Scaling | Fixed / Vol-Based / Confidence-Based | Confidence-Based |
Under the Hood
Type: candle-triggered analysis agent with team-chat capabilities.
Trigger: candle close on configured timeframe (default 15m).
Upstream agents: Vanta, Luma, Meridian — fetched from the analysis cache (hot cache with database fallback).
Consensus: supports two modes (sequential handoff or parallel opinion collection). Agents contribute structured opinions that Riven synthesizes.
Cooldown: per-symbol cooldown scales with the trigger timeframe to prevent duplicate broadcasts.
Output: publishes one recommendation per symbol via the event bus + the full consensus debate for War Room rendering.
Fallback: if no upstream agents are online, Riven falls back to solo analysis.