PRISM Scoring

PRISM (Performance and Reliability Incentive for Subnet Mining) is Beam's orchestrator scoring system. Your PRISM final score shapes how much production traffic you receive. Validator emission weights are based on completed qualified production tasks.

Final score

throughput_score      = fleet_normalize(decayed_assignment_verified_mbps)
reliability_score     = fleet_normalize(raw_reliability)

performance_score     = 0.40 x throughput_score + 0.60 x reliability_score

penalty_multiplier    = decayed_penalty_pressure(fraud, integrity_chunk_mismatch, sybil events)

current_ready_gate    = (ready AND connected) ? 1 : 0
active_time_ratio     = active_seconds_in_lookback / lookback_seconds
readiness_multiplier  = current_ready_gate x active_time_ratio

prism_final_score     = performance_score x readiness_multiplier x penalty_multiplier

Scores and multipliers are clamped to [0, 1].

Performance

Beam compares orchestrators in the same pool cohort and maps each value linearly from fleet min to max into the compressed band [0.2, 1]: the lowest positive throughput in the cohort scores 0.2, the highest scores 1, with linear scaling between. Orchestrators with no positive throughput evidence score 0.

Throughput uses provider-verified transfer bandwidth samples. BeamCore keeps per-batch verified bandwidth visible for observability, then combines eligible batches into one task-count-weighted sample per transfer:

transfer_mbps = SUM(batch_mbps x batch_task_count) / SUM(batch_task_count)

The singleton first_assignment batch always counts toward throughput. Later batches need to be above BeamCore's configured follow-up minimum task count for verified BW (10) to count, while their task outcomes still count for reliability and task totals.

Throughput uses recent verified transfer bandwidth with a 1-hour half-life. A bandwidth sample around 1 hour old contributes about 0.5; a sample around 2 hours old contributes about 0.25.

Reliability uses time-decayed task outcomes - completions, failures, and reassignment-style failures - over the same window with a 24-hour half-life. A reliability sample around 24 hours old contributes about 0.5; a sample around 48 hours old contributes about 0.25. Raw reliability blends success rate and reassignment rate, then fleet-normalizes on the same compressed min-to-max range before it enters the performance blend.

Performance combines throughput and reliability with 40% throughput / 60% reliability weighting.

Confidence and pools

Beam routes each orchestrator through one of two pools:

You begin in qualifying. Verified calibration work builds your confidence score. At 0.9 confidence, Beam graduates you to qualified for production routing and validator weights. Qualified membership persists for routing and scoring on production work.

verified_task_ratio = min(1, verified_task_count / target_verified_tasks)
age_ratio           = min(1, age_days / 7)
maturity_factor     = 0.8 + 0.2 x age_ratio

confidence_score    = verified_task_ratio x success_rate x maturity_factor

The verified-task target is about 120 distinct tasks in the evidence window. Confidence also reflects your recent success rate and account age (full maturity around 7 days).

Right after graduation, routing share uses a mid-tier weight until your first production transfer; it then follows your live PRISM score from production evidence.

Qualifying and qualified orchestrators are scored against peers in the same pool.

Qualified seasonal reset

Qualified PRISM runs in weekly seasons. A new season starts on Sunday at 00:00 UTC.

At the start of a season, every qualified orchestrator begins from a neutral routing score of 0.5 until Beam sees fresh qualified PRISM evidence for that season. Readiness still applies, so disconnected or not-ready orchestrators do not receive traffic from the neutral score.

Your first qualified evidence in the season clears the reset status on the next score refresh.

Readiness

Readiness has two parts:

1. Current gate - you must be both ready on BeamCore and connected to BeamCore over NATS. If either is false, the folded readiness multiplier is 0.
2. Active-time ratio - BeamCore records readiness and connection transitions, then measures the percentage of the PRISM evidence window where ready && connected was true.

The dashboard and APIs show a simple status:

Penalties

Penalty events in the evidence window shape your penalty multiplier:

Each event contributes:

pressure += coefficient[kind] x 0.5^(age_hours / 48)
penalty_multiplier = clamp(1 - pressure, 0.0, 1.0)

The evidence window for tasks, bandwidth samples, readiness active-time, and penalties is 7 days by default. Reliability samples use a 24-hour half-life, and bandwidth samples use a 1-hour half-life.

Guardrail reassignments feed reliability samples.

Reading your score

Open the Beam Dashboard for live pool, confidence, and PRISM final score. Use the Selected Orchestrator panel for the full breakdown.

For programmatic access:

GET /orchestrators/prism-scores/:orch_uid

Improving your score

• Maintain a stable NATS control connection and set ready when your worker pool can receive work.
• Keep your orchestrator active; intermittent downtime reduces the readiness multiplier.
• Complete transfers reliably with steady task completion.
• Complete assigned chunks promptly so provider-verified transfer samples reflect sustained throughput.
• Avoid fraud, integrity mismatch, and sybil penalties that reduce the PRISM final score.