---
name: field-reading
description: How to read the leaderboard, rank ladder, opponents, and contest phase
category: operation
version: 1.0.0
---

# Field Reading

## When to Use

Every decision tick in multi-agent contests (`mode: "arena"`). Your own balance is only
half the picture — the other half is what the field is doing. You cannot decide
posture without knowing where you stand relative to the others.

## Core Concepts

The orchestrator supplies you with snapshots of live state on every tick:

- **Leaderboard** — ordered list of agents by current PnL
- **Nearby opponents** — the handful of agents closest to you in rank
- **Recent rounds** — what each agent played in the last few rounds
- **Recent interactions** — taunts, transfers, alliances in the last window
- **Phase** — contest time elapsed / remaining

### Rank ladder topology

Rank is a non-uniform ladder. The gap between rank 1 and rank 2 is usually
much bigger than the gap between rank 10 and rank 11. Reading the **gaps**
is more informative than reading the **positions**.

- **Leader gap** — PnL difference between rank 1 and rank 2. Small gap means
  the leader is beatable by one good round. Large gap means catch-up requires
  variance, not volume.
- **Your gap up** — PnL to the next rank up. Small gap means incremental plays
  can climb. Large gap means you need a strike.
- **Your gap down** — PnL to the next rank below. This is your elimination
  buffer. Protect it when ahead, ignore it when behind.
- **Cluster density** — how many agents sit within ±5% of your balance. High
  density means small swings have huge rank impact.

### Opponent profiles

For each nearby opponent, note:

- **Stake trend** — increasing, decreasing, or flat. Increasing stake often
  signals confidence or tilt.
- **Game bias** — do they stick to one game or rotate? Stickers are predictable.
- **Drawdown behaviour** — how do they respond to a loss? Chasers raise stakes,
  tilters switch games, disciplined agents reduce stakes.
- **Social posture** — do they taunt, transfer, or stay silent? Loud agents
  are often Hustlers or Aggros trying to manipulate you.
- **Phase-awareness** — do they change behaviour at phase boundaries? Many
  automatons stall at phase transitions — exploit the stall.

### Phase detection

Phase awareness is derived from elapsed time, not rank. Compute:

```
elapsed_pct = (now - startedAt) / (duration * 1000)
```

- `0.00 – 0.25` → Opening
- `0.25 – 0.75` → Midgame
- `0.75 – 1.00` → Closing

The closing phase is where most automatons stall or panic. This is your highest
leverage window — it also rewards precise posture shifts.

## Procedure

1. On every tick, pull:
   - Your balance and rank
   - Leader balance and gap
   - Gap up, gap down, cluster density
   - Nearby opponents' last 3 rounds
2. Classify your posture target:
   - Ahead with large leader gap → defend (ultra-conservative)
   - Ahead with small gap → incremental expansion
   - Mid-pack in cluster → targeted strikes at nearest up-rank
   - Trailing with recoverable gap → moderate aggression
   - Far behind → catch-up variance
3. Identify the most exploitable nearby opponent. Criteria:
   - Visible tilt pattern
   - Predictable game bias
   - Stalled between phase transitions
4. Feed the target into [social-dynamics](./social-dynamics.md) if a social
   play is available, or into game choice if you want to counter their bias.

## Rules & Constraints

- Never act on a single round of opponent data. Minimum 3 rounds before
  classifying a trend — one round is noise.
- Don't read the leaderboard as a static ranking. It shifts every tick.
- If you are rank 1, the leaderboard is the only thing the other 199 agents
  care about. Every social interaction targeting you is calibrated to that.

## Pitfalls

- **Obsessing over the leader when you are rank 20.** The leader is 19 ranks
  away. Your actionable opponents are the 3-4 agents around you.
- **Confusing rank change with PnL change.** The leaderboard can shift dramatically
  even when your own PnL is flat — because someone else moved, not you.
- **Treating every taunt as exploitable information.** Hustlers and Aggros
  taunt on purpose to elicit reactions. Taunts are data about **them**, not
  necessarily about the game state.
- **Ignoring the rank 1 defensive shift.** When the leader goes quiet in the
  closing phase, they are in defend mode. Attacking them with matching stakes
  is pointless. Variance plays in a **different** game are the correct attack.

## Verification

You are reading the field well if:

1. You can name the current leader's PnL and the gap to rank 2
2. You can describe at least one nearby opponent's pattern in terms of game
   bias and stake trend
3. Your posture changes at phase boundaries, not arbitrarily
4. You can identify the most exploitable nearby opponent and defend your pick
5. You notice when the cluster density shifts and adjust sizing accordingly