Scoring + the incentive shift

Canonical source: reliquary-forge/docs/scoring.md. Synced into this site as apps/web/content/synced/scoring.md.

The shift: from "do as many rollouts as you can" → "find me the rollouts that I need to train on." Miners now compete on the best inferences, not the most. Two pricing knobs together collapse the quantity-farming exploit:

1. Per-rollout quality — each accepted, unique verdict adds its dense_reward (fraction of clauses satisfied) to the miner's total. Mass-producing low-reward completions earns less than a few high-reward ones.
2. Superlinear weighting — totals are squared before normalization. weight ∝ total². A miner with 2× the raw reward earns 4× the weight; 3× → 9×.

Pipeline

Rollouts → Verdicts → Uniqueness Dedup → Score Aggregation
        → Superlinear Weighting → Normalized Weights

Uniqueness dedup

Each rollout is keyed by SHA-256 of its comma-joined token ids. The first occurrence of (challenge_id, uniqueness_key) is marked is_unique=true; duplicates are excluded from scoring entirely. Submitting the same completion twice across a window earns nothing extra.

Acceptance gate

A verdict is accepted when both:

• proof_valid = true — the sketch proof passed verification
• evaluation.accepted = true — the SAT evaluation completed successfully

Anything else (failed proof, malformed environment, copycat, log-prob drift) is rejected before scoring.

Score aggregation

For each accepted, unique verdict the miner's total accumulates:

totals[miner] = Σ dense_reward across all accepted, unique verdicts

Dense reward is the fraction of clauses satisfied — partial credit for partial solutions. Validators recompute it independently from the clause evaluation; miners cannot self-report.

Superlinear weighting

powered[miner] = totals[miner] ^ SUPERLINEAR_EXPONENT  // = 2.0
weight[miner]  = powered[miner] / Σ powered[*]

Weights normalize to 1.0. With exponent = 2.0, an Alice-vs-Bob ratio of 2.18× in raw reward becomes 4.75× in onchain weight.

DAPO zone filter (the trainer-side dual of the incentive)

Per §9.1 of the paper, Forge's GRPO trainer only consumes rollout groups where σ ≥ σ_min — the within-group reward standard deviation crosses a threshold. Below that, the normalized advantage (rᵢ − μ)/σ collapses to zero and the group carries no gradient signal.

The dashboard surfaces this directly in the Training Signal · Zone Filter panel (zone-hit ratio per window, group σ histogram, mean reward inside zone). Miners producing rollouts that LAND IN the zone are the ones whose work actually trains the model — and the ones the incentive curve concentrates weight on.

Worked example

Alice has 2.18× Bob's raw reward but receives 4.75× the weight.

Sybil resistance

Splitting identity is punished by the same superlinear curve. One miner with total 5.0 earns 5²=25 units of weight. Two miners with total 2.5 each earn 2 × 2.5² = 12.5 — half. The math actively discourages identity fragmentation.

The nine stages (verifier pipeline)

Verdicts must clear:

1. schema — verdict envelope well-formed
2. tokens — every token in vocab
3. prompt — prompt matches window policy
4. proof — sketch receipt matches within tolerance
5. termination — completion ended cleanly
6. environment — the env's own evaluator accepts
7. reward — reward value matches declared formula
8. logprob — per-position log-prob drift < 0.15 on ≥ 51 % positions
9. distribution — median importance ratio in [0.85, 1.15]

Stages 1–7 are hard rejects. Stages 8–9 can soft-flag (accepted but observed).