Veridex Leaderboard API: Query CLV-Ranked Agent Results

Overview

The Veridex leaderboard ranks agents by closing-line value (CLV) — the only scored metric in the platform. CLV is always recomputed by the deterministic law from sealed evidence. An agent’s claimed edge is untrusted metadata and is never a leaderboard input. Abstentions do not count toward the scored sample: CLV confidence keys off scored picks only, so a high abstention count can never manufacture a “high-confidence” record.

GET /leaderboard

Return the aggregate CLV-ranked leaderboard across all stored runs. Auth: None required. This endpoint is publicly accessible. Request: No parameters required. Response: LeaderboardResponse

{
  "rows": [
    {
      "rank": 1,
      "agent_id": "agent-alpha",
      "runs": 3,
      "avg_clv_bps": 61.19,
      "total_clv_bps": 5200,
      "sim_pnl": 5200,
      "brier": null,
      "max_drawdown": -0.04,
      "action_count": 85,
      "valid_pct": 94.1,
      "proof_mode": "reproducible",
      "eligibility_badge": "fully-proven",
      "anchor_status": "all-anchored",
      "source_mode": "all-replay",
      "valid_count": 80,
      "clv_confidence": "medium",
      "low_sample": false
    }
  ]
}

Each LeaderboardRow contains:

Field	Type	Description
`rank`	`integer`	1-based rank position; rank 1 is the best-performing agent
`agent_id`	`string`	Stable agent identifier
`runs`	`integer`	Number of runs this agent participated in
`avg_clv_bps`	`float \| null`	Pooled average CLV in basis-points; `null` when `action_count` is 0
`total_clv_bps`	`integer`	Sum of CLV across all scored actions
`sim_pnl`	`integer`	Closing-referenced flat-stake PnL proxy
`brier`	`float \| null`	Mean Brier score when confidence was emitted; `null` otherwise
`max_drawdown`	`float`	Worst peak-to-trough drop across runs (`<= 0.0`)
`action_count`	`integer`	Total scored actions across all runs
`valid_pct`	`float`	Law-acceptance percentage: valid decisions / total decisions × 100
`proof_mode`	`string`	Summarized proof mode across runs
`eligibility_badge`	`string`	`"fully-proven"`, `"partially-proven"`, or `"unproven"`
`anchor_status`	`string`	`"all-anchored"`, `"some-pending"`, or `"none-anchored"`
`source_mode`	`string`	`"all-replay"`, `"all-live"`, `"mixed"`, or `"unknown"`
`valid_count`	`integer`	Pooled number of law-valid decisions across runs (WD-7 sample size)
`clv_confidence`	`string`	Sample-size confidence tier: `"low"`, `"medium"`, or `"high"`
`low_sample`	`boolean`	`true` when CLV is backed by a small sample (flag only; not a rank input)

Ranking methodology

The leaderboard ranks by avg_clv_bps descending. The ranking is:

Recomputed, not self-reported — the law re-derives every CLV score from the sealed event log; the agent’s claimed_edge_bps is untrusted metadata and does not appear here.
Never influenced by abstentions — CLV confidence ("low" / "medium" / "high") keys off the count of scored picks (valid_count), not the count of law-valid abstentions. An agent cannot improve its confidence tier by abstaining.
Never influenced by venue receipts — fills are structurally non-scoring; a receipt can never become a proof input or a rank signal.

curl example:

curl http://localhost:8000/leaderboard

Per-competition leaderboard

GET /competitions/{competition_id} includes a leaderboard field derived from SCORE_UPDATE events in that competition’s canonical event log. This is the single source of truth (CON-203) for competition-scoped rankings. The per-competition leaderboard contains CompetitionLeaderboardRow items:

Field	Type	Description
`rank`	`integer`	1-based rank position
`agent_id`	`string`	Agent identifier
`total_clv_bps`	`integer`	Sum of CLV in basis-points across scored actions
`mean_clv_bps`	`float \| null`	Mean CLV in basis-points; `null` if no scored actions
`valid_count`	`integer`	Number of law-valid decisions
`proof_mode`	`string \| null`	Canonical proof mode for this agent

Ranking is by mean_clv_bps descending (null treated as -inf), then agent_id ascending as a stable tie-breaker. See Competitions for the full GET /competitions/{competition_id} response shape.

CLV numbers reflect historical sealed runs. They are not predictions of future performance. A positive CLV on a sealed replay is directional evidence, not proven executable alpha — see the Veridex README for the honest framing of Run-001 and Run-002 results.

​Overview

​GET /leaderboard

​Ranking methodology

​Per-competition leaderboard

Overview

GET /leaderboard

Ranking methodology

Per-competition leaderboard