Attestations are live over REST — contracts are not deployed yetRead an attestation
Cleaton

Calibrated confidence

Why the published confidence is a conformal coverage guarantee rather than the model's own probability.

Argued in full in the whitepaper at §5.

Why not a model probability#

The ensemble already emits S_P(t), a number between zero and one. Printing it next to the horizon and calling it confidence would be wrong in a way that matters for a bonded claim.

S_P(t) is a probability under the model. It inherits every misspecification the model carries. If proportional hazards fails, or the feature distribution shifts because a new chain behaves unlike the training set, a stated 0.9 may correspond to an empirical frequency of 0.6. For a dashboard that is a quality issue. For a claim someone posts a bond against, it is a false statement with a number attached.

Split conformal#

Conformal prediction gives distribution-free, finite-sample coverage under exchangeability — the guarantee holds whether or not the model is correctly specified.

sᵢ = ( Ĥᵢ − Yᵢ ) / ς̂ᵢ
ς̂ is the model's own dispersion estimate, from ensemble disagreement and Cox standard errors.

The score is signed. Over-prediction is the error that matters, because it is the error that gets a user liquidated; under-predicting a horizon only costs someone some yield. What gets published is the conformalised lower bound, not the point estimate:

H_pub = Ĥ_P − q₍₁₋α₎ · ς̂_P
Guarantee: Pr[ Y_P ≥ H_pub ] ≥ 1 − α, regardless of model correctness.

So a published confidence of 0.7 means something checkable: across the calibration distribution, at least 70% of attestations at that level had horizons no breach exceeded. It is a statement about long-run frequency, verifiable against the public record.

The exchangeability problem#

Coverage requires exchangeability, and market regimes are emphatically not exchangeable. A calibration set drawn from expanding liquidity will not cover a contraction. Three mitigations, none complete:

Regime-stratified calibration — separate sets per regime, classified by aggregate stablecoin supply direction, cross-chain flow direction and realised volatility.

Weighted conformal — under covariate shift with estimable likelihood ratios, weighting restores coverage. The weights come from a discriminative classifier trained to separate the calibration distribution from the current one.

Explicit degradation — outside the support of every stratum, the system does not extrapolate.

OUT_OF_SUPPORT#

Calibration-set membership is a deterministic function of attestation nonce and time, fixed before outcomes are known. That is what stops an adversary — or us — from influencing which attestations enter the set and distorting the guarantee.