Scoring

emission = composite_Q
         * base_weight[skill_type]
         * tier_multiplier[tier]
         * early_submission_bonus
         * role_multiplier
         * consensus_score
         * bootstrap_bonus_if_applicable

if verdict == ground_truth:
    α = 1.0
elif verdict == "REVIEW":
    α = 0.5
elif false_positive:                 # said BLOCK but ground truth is ALLOW
    α = 1.0 - 0.4 * risk_score
elif false_negative:                 # said ALLOW but ground truth is BLOCK
    α = max(0.0, 1.0 - 2.5 * (1.0 - risk_score))

ε_base = 0.0

if probe_evidence verified:                 ε_base += 0.3
if trace_hashes consistent:                 ε_base += 0.3
if sandbox_manifest digest correct:         ε_base += 0.2
if findings cite evidence_ref correctly:    ε_base += 0.2

ε = ε_base

precision = | miner_rules ∩ expected_rules | / | miner_rules |
recall    = | miner_rules ∩ expected_rules | / | expected_rules |

β = 0.5
π = (1 + β²) * precision * recall / (β² * precision + recall)

window = (deadline_s - t_min_s) * 1000     # ms
position = (latency_ms - t_min_s * 1000) / window

if latency_ms < t_min_s * 1000:    η = 0.0       # too fast, suspect caching
elif latency_ms > deadline_s * 1000: η = 0.0     # too slow, dropped anyway
else:                                η = 1.0 - position

canaries_detected = | findings.category == "canary_match" |
canaries_expected = count of injected canaries

ρ = canaries_detected / canaries_expected

μ = max(0.0, 1.0 - |miner.risk_score - ref.risk_score|)

predicted_taint_cmds = set of commands flagged by taint analysis
executed_cmds       = set of commands appearing in process.jsonl

σ = | predicted_taint_cmds ∩ executed_cmds | / | predicted_taint_cmds |

manifest_recovered = miner.evidence.mcp_manifest_hash
manifest_truth     = task.expected_manifest_hash

if manifest_recovered == manifest_truth:    ψ = 1.0
else:                                       ψ = 0.0

poisoned_tools_detected = findings where category == "tool_poisoning"
poisoned_tools_expected = task.expected_poisoned_tools

τ = | detected ∩ expected | / | expected |

expected_risk = task.expected_aggregate_risk
miner_risk   = miner.analysis.risk_score

χ = max(0.0, 1.0 - |miner_risk - expected_risk|)

Q = (α^0.30 * ε^0.30 * π^0.15 * η^0.10 * ρ^0.15)

Q = (α^0.40 * ε^0.20 * π^0.20 * η^0.10 * μ^0.10)

Q = (α^0.35 * ε^0.30 * π^0.20 * η^0.15)

Q = (α^0.30 * ε^0.30 * π^0.15 * η^0.10 * σ^0.15)

Q = (α^0.25 * ε^0.25 * π^0.15 * η^0.10 * ψ^0.10 * τ^0.15)

Q = (α^0.30 * ε^0.25 * π^0.15 * η^0.10 * χ^0.20)

if ε < 0.10:
    Q = 0.0