Independent, multi-signal analysis of whether registered Flare providers' submitted values resemble the stock example implementation. This is an implementation-likeness model, not an accuracy ranking.
Running an independent production provider is required for reward eligibility and is enforceable by the FTSO Management Group.
Choose a section below to navigate directly to it, or select Show everything to display the complete page.
| Provider | Likelihood | Classification | Weighted VP share | Evidence |
|---|---|---|---|---|
| Loading validated report… | ||||
Retrieve the complete provider census, scores, evidence components, network summary, and methodology through the versioned read-only JSON API.
API index · Provider data · Cohort history · OpenAPI document
The collusion-evidence workspace is in private beta testing. Detector development, event evidence, graphs, provider comparisons, and raw analysis interfaces are not publicly available.
The detector tests whether a provider's submitted-value process resembles the stock implementation. Each candidate is evaluated independently against two anchors—the captured stock provider and the passive on-chain Flare Network Test provider. This is a probabilistic implementation-likeness model, not an accuracy ranking or an inference of intent.
For candidate value v and anchor value a at the same feed and voting round, the model measures exact equality, proximity, and relative mismatch magnitude.
exact = count(v = a) / nIt also measures update-direction agreement, direction during volatility, and whether both values cross a robust population band on the same signed side.
Volatility spikes and signed band excursions are binary boundary events. Coincidence matters only when shared events exceed the rate expected from both series' marginal activity.
expected shared = nP × nR / NThis discounts ordinary market-wide movement and sporadic coincidence.
Each anchor contributes exact rate, near-5-bps rate, transformed error, movement direction, volatility direction, and signed out-of-band agreement. Two regime families add precision, recall, and transformed lift: 18 model inputs in total.
error feature = −ln(1 + mean error bps)Cadence and the raw harmonic anchor score remain diagnostics; neither replaces the calibrated final posterior.
A class-balanced logistic model is fitted using asserted example users as positive controls and known independents plus canonical primary-band performers as negative controls.
class weight = N / (2 × Nclass)Weighted cross-entropy is constrained by L2 shrinkage. The strongest tested penalty that separates every asserted control at 0.5 is selected; if none succeeds, calibration fails closed.
Update/freeze timing is reported separately using Cohen's κ over feeds and lags −1, 0, and +1. Chance agreement from each series' marginal update rate is removed.
po = (n11 + n00) / NThe displayed cadence is a reference-information-weighted diagnostic, not a final logistic input.
Detection requires calibrated implementation likelihood strictly above 65%. Coverage is derived from usable evidence at both anchors and is descriptive—not a model weight.
detected cohort: P > 0.65Vote power is calculated after classification. It describes the detected cohort's network weight and cannot influence a provider's likelihood.