SN99 — score multiplier by miner

Some miners' published score is their replayed q_lcb multiplied by a fixed per-miner constant. This is every miner the replay could verify.

56 / 86
miners replayable
14
sustained multiplier > 1.0
1.030 – 1.150
observed range
±0.000000
replay accuracy where verified

Table

uidmultipliersegment lenconfidenceexact epochsrepo

multiplier = published ÷ replayed, on the longest run of ≥5 consecutive epochs flat to ±0.006. exact = epochs where the replay reproduced the published feed to 1e-4; that is what proves the fit is right, so a row with exact = 0 is unverified and its multiplier may be fitting error.

Multiplier by epoch

The three miners with Docker-graded ground truth, epoch by epoch. 1.00 means the replay matched the published feed exactly.

What this is

published_score = C × [ q_lcb − 0.02 · min(1, cost_per_task / 0.015) ] q_lcb = Σ_b weight_b · wilson_lcb(k_b, n_b) z = 1.645 EWMA λ = 0.5^(1/200) weights: mmlu 0 · math 0 · code 1

Every score on this subnet is reproducible from public data: on-chain commitments give each miner's artifact, its HuggingFace bundle gives the attested proofs, and replaying the evidence accumulator reproduces the published feed exactly — for most miners. For the rest, the published value is the replayed one times a constant.

Those constants land on exact decimals. uid 28 holds 1.03000000 to ±3×10⁻⁷ while its own q_lcb swings 2.5%. A computed quantity does not do that; a stored one does.

What was ruled out

The obvious candidate is headroom_lcb, the documented router scalar that divides by the achievable gap. It is excluded by the owner's own published reference: at uid 117's trigger epoch the gap is 0.0000, and headroom_lcb returns None below 0.05. An independent reimplementation of the frontier reproduces the owner's achievable_gap on 263 of 263 epochs, so this is not a modelling error. Across 270 references, 84.8% of epochs have gap exactly zero — the suite is saturated and the router scalar almost never engages.

Also excluded, each with numbers: pooled and code-only cost (ranges overlap), accuracy level, artifact age, missed epochs, rung diversity, holding a reign seat, eps hysteresis (0.02 maximum against a 0.198 effect), evidence injection, and newer upstream code.

What is not established

The form is verified: an exact per-miner multiplication, confirmed on three miners with a clean control. The cause is not. This observes published output, not the code producing it — a configuration table, an undocumented feature, or a mechanism invisible from outside would all look like this. No claim is made about intent.

Eight miners show sustained segments below 1.0, down to 0.80. None is confirmed against a graded ground truth and each may be replay error from a mis-dated reset. They are shown for completeness, not as a finding.