Arena
Arena is timed, head-to-head prediction competition on a game preset.
You deploy agents onto a preset’s sessions. At the Opening Bell, once the candle before it has closed, the platform resolves which agents are on duty, lets the first whose strategy admits a full grid build one, and can auto-submit a real wager before entries lock — under that agent’s playbook and your deployment rules. Agents do not need to hold live perps the whole session; this is game participation, not Radar-style continuous trading.
Use Arena to pressure-test edge under live market movement — before, beside, or instead of Radar and the trade terminal.
Create ≠ deploy (Arena edition)
| Step | Effect |
|---|---|
| Create agent | Runner exists, strategy bound, risk rails set |
| Deploy to Arena | You + this game preset — who plays which sessions / regimes |
| Opening Bell | The on-duty agents are tried in priority order; the first whose strategy admits a full grid generates it + may auto-submit (entry fee) before lock |
An Arena slot configures game participation only. It grants no standing authority to trade a coin. Continuous trading authority is a separate Radar policy per coin.
Session timeline
Every session scores exactly one exchange candle, from its open to its close. Three instants define it:
| Instant | What happens |
|---|---|
Opening Bell (startsAt) | The scored candle opens. The candle before it has just closed, so every entrant can read it |
Lock (lockAt) | Entries close: the Opening Bell plus a fraction of the candle (10% by default). The session goes live and its roster freezes |
Settle (settleAt) | The scored candle closes and the session settles |
Entries stay open from the moment the session appears until lock, including the Opening Bell window between the bell and lock. Submitting, editing and cancelling (with a full refund) all work inside the window. At the default fraction the window is:
| Session | Window |
|---|---|
| 5M | 30 s |
| 15M | 1 m 30 s |
| 1H | 6 m (e.g. bell 14:00 UTC, lock 14:06 UTC, settle 15:00 UTC) |
| 4H | 24 m |
| 24H | 2 h 24 m |
The arena header counts down to the bell (“OPENING BELL IN 12m 34s · 14:00 UTC”), then to lock (“OPENING BELL · LOCKS IN 4m 12s”). The next session of a preset appears once the current one locks.
When a deployed agent enters
A deployed agent looks once per session, inside the Opening Bell window:
- never before the bell, and only once the candle that closed at the bell has settled for every coin in the pool;
- only while at least the platform’s lock cutoff (5 minutes today) remains before lock.
A window shorter than the cutoff — 5M and 15M sessions at the default fraction — gets no automated entry. A session you already entered or cancelled yourself is never entered by automation.
The look tries your on-duty agents in priority order, and each is judged by its own strategy first:
- Admission. The agent’s entry conditions decide which coins of the pool it may pick, and in which direction — by the same rules a trade uses. A coin is admitted only when every required condition reads true, and its verdict decides the side: Up admits UP, Down admits DOWN, a strategy with no verdict condition admits both, and Neither or an unresolved verdict admits none. A strategy with no required and no verdict condition admits nothing. Trade settings — the ATR floor, stop band and exit rules — never refuse an arena pick.
- Stand-aside. An agent whose strategy admits fewer coins than the grid needs (9 on today’s presets) stands aside — free: no LLM credit is spent and nothing is wagered — and the session passes to the next on-duty agent. The journal records it with the admitted count.
- The session’s agent. The first agent with a full admission is the one that plays. It may pick only the admitted coins and directions, and a grid holding any other pick is refused rather than repaired. No other agent is tried after it: a grid below its slot’s confidence bar sits the session out with no wager, and there is no second look.
When every on-duty agent stands aside, the session is sat out.
What a deployment is
A deployment is scoped to you + one game preset:
| Field | Role |
|---|---|
| Preset | Which arena (duration / settlement window, entry fee, coin pool, scheduling) |
| On / off | Pause without deleting — off means no entry and no wager; slots stay |
| Slots | Ordered agent rules + an optional fallback: always-on rules after the others, or one default catch-all |
| Regime anchor | Optional override (coin + timeframe); omit to inherit the preset default |
Save refuses stale edits: if the deployment changed since you loaded it, the write is rejected. Reload, then save again.
Fleet roll-up reports arenas deployed, scheduled, sitting out, warming, no session, paused (your switch), and retired (admin disabled the preset). Those buckets are separate on purpose.
Preset active vs you paused
| Switch | Who | Effect |
|---|---|---|
| Preset active | Platform | Off ⇒ arena schedules no new sessions; your policy is dormant |
| On / off (you) | You | Off ⇒ this deployment plays nothing until you resume |
Delete withdraws the policy (un-deploys all agents from that preset). Pause keeps the rules.
Slots
Each slot binds an agent and a min confidence grid bar. When the agent enters is fixed by the Opening Bell, not by the slot.
Rule vs default
| Kind | Priority | Conditions | Role |
|---|---|---|---|
| Rule | Unique positive integer; lowest is tried first | 1–2 activation conditions | On duty when every attached condition is true |
| Always-on rule | Unique positive integer, after every other rule | Always, alone | On duty at every session, behind the other rules |
| Default | (none) | None | On duty only when no rule matches — optional, at most one |
The fallback is either the default or always-on rules, never both. The default plays only when no rule matches. An always-on agent is on duty at every session and plays whenever the agents above it stand aside — it still needs a full admitted grid. Several always-on rules back each other up in order. A rule with no condition is never saved, and never means always.
Resolution for the next pending session puts every matching rule’s agent on duty, in priority order (an agent in two matching rules is on duty once, at the earlier of the two), or the default alone when no rule matches — else it lands idle / warming / no-session / paused. An archived agent is never on duty: its rule is skipped and the next matching rule, or the default, takes its place. Which on-duty agent plays is decided only at the Opening Bell, by admission.
Min confidence
Per-slot min confidence (shown as a percent) is the grid confidence gate for that agent in this arena.
- A number floors how confident the generated grid must be to auto-submit.
- Leaving confidence unset means no confidence bar on the slot (no gate from this dial).
- The resolved bar (and display percent) is computed for you — including any platform floor.
This is not Radar’s trade conviction dial. Arena confidence gates the game grid; Radar conviction gates trade capital.
Conditions
Arena owns session vocabulary (Radar owns wall-clock hours). A rule may carry session selection, regime, or both. With both, both must match. Always is the one exception: it stands alone in its rule.
Always
Matches every session. A rule whose one condition is Always is an always-on rule: it sits after every other rule, and a deployment holding one has no default.
Session start (recurring)
Exact UTC start times (hours:minutes:seconds, each unique) sharing one weekday set (Sunday–Saturday UTC).
A session matches when its start-of-day time — its Opening Bell, not its lock — is in your start-time set and its weekday is in your days set. Contiguous bands and scattered hours are the same shape — membership, not a from/to range.
Pick starts from what the preset currently schedules. An unscheduled start can still save but stays dormant until that hour is restored on the preset.
Session occurrence (one-shot)
A single concrete session start (one UTC timestamp) — one concrete session.
Regime
Match confirmed regime at the effective regime anchor (your override or the preset default). Labels match the product:
Bull trend · Bear trend · Bull range · Bear range · Fallback
Beside the regime, a rule carries two more settings:
- Volatility — a multi-select of Expanding, Normal and Contracting, all three by default (any volatility).
- During warnings — Keep running (default), Pause during warnings or Only during warnings. The warnings are Crash and Momentum divergence; a warning sits beside the direction and never replaces it, so a Bull trend rule stays eligible through a warning unless set to pause.
Optional min conviction floor on the regime reading (Uncertain → … → Maximum); the floors on offer are the ones the selected regimes and volatility levels can reach. Prefer real regimes when authoring — Fallback is not the intended target path. See Radar: Market regimes for what each reading means.
Live resolution
Every policy read carries a computed resolution for the canonical next pending session:
| Status | Meaning |
|---|---|
| Resolved | The on-duty agents for the target session, in priority order — each with its slot and confidence bar — and the regime used. The first whose strategy admits a full grid at the bell plays |
| Idle | Rules ran; nothing matched (and no default or always-on rule to fall to) |
| Warming | Regime / anchor not ready yet for selection |
| No session | No upcoming pending session to name |
| Paused | You turned the deployment off — nothing resolved, no agent identity disclosed |
Retired presets (platform switched the arena off) appear in fleet KPIs separately — the arena itself schedules nothing.
Preview resolution is a dry-run against a draft slot set (optional simulated reading: regime, volatility and warnings) — same resolver, no write. It names no single winner ahead of the bell, because admission reads the candle that closes at the bell. Test goes one step further against the live pool: it tries the on-duty agents in order, reports each one that would stand aside with its admitted count, and generates — without submitting — the grid of the one that would play. A draft that names an agent you cannot use is refused before anything runs.
Auto-submit and money
A live deployment is autonomous wager authority for that preset: when the session’s agent — the first on-duty agent whose strategy admits a full grid — generates a grid that clears its slot’s confidence bar, the platform can submit it and charge the preset entry fee (USDC). An agent that stands aside spends nothing. That is why create / replace / delete need confirmation in Commander and the wager permission on MCP.
Generate / test grid paths can preview without submitting. Submit is the spend.
How this fits the stack
| Layer | Owns |
|---|---|
| Strategy | Playbook the agent uses to build the grid; its entry conditions decide which coins and directions it may pick |
| Agent | Brain, bind, risk — the runner that plays |
| Arena deployment | Which runner plays which sessions / regimes on this preset |
| Radar | Separate: standing trade authority per coin |
Commander and MCP use the same concepts: list, get, preview, upsert, and delete deployment policies.
← Radar · Next: Trade terminal