The v1 pipeline described here is built (at relayer/src/sources/magiceden.ts, using Magic Eden’s public API). It was exercised end to end against live Mad Lads data. The bid-side leg is planned and needs a Tensor API key. NFT collection floors are the secondary asset class; see Asset Classes.

​
The single lowest listing is not the floor

The naive “floor price” is the lowest ask, and a single ask is manipulable: list your own NFT below everyone else and the number moves for free. But that logic cuts differently for a lone listing than for the ladder as a whole:
  • A fake low listing is a fillable offer. Anyone can buy it, so pricing your NFT meaningfully below market is a donation to arbitrageurs, and the ladder self-heals within blocks.
  • A lone high listing never matters: it sits above the rest of the ladder and no depth-aware estimator reads it.
So the manipulation-resistant observable is not the lowest ask; it is the depth of the ladder.

​
Ladder estimation (implemented v1)

The estimator per cycle, per collection:
  1. Fetch the cheapest current listings (the ask ladder) from Magic Eden.
  2. Take the median of the cheapest 3 listings, i.e. the 2nd-cheapest ask. This hugs the true floor but cannot be moved by any single fake listing: an underpriced listing gets bought, an overpriced one never becomes the median.
  3. Cross-check the estimate against the venue’s own reported floor. Disagreement beyond maxDisagreementBps holds the push and alerts instead.
  4. Collect recent buyNow sale prints in the lookback window as logged confirmation. On blue-chip collections the public activity feed is dominated by bid and pool-update churn, so sale prints are too sparse to be the primary signal; they corroborate, they do not set the number.

​
Planned: the bid-side leg

Collection-wide bids on Tensor are posted capital. Faking a high bid means genuinely offering to buy at that price, because any holder can fill it. Adding the top collection bid as an independent estimator leg (median of ladder estimate, top bid, and sale TWAP) is the planned hardening once a Tensor API key is provisioned, and it turns the cross-check from one venue’s self-reported floor into true cross-venue disagreement detection.

​
Gates

  • Ladder depth: fewer than 3 live listings means no push at all.
  • Disagreement: ladder estimate vs reported floor beyond the configured bound (default 25 percent) holds the last value and alerts.
  • Staleness on hold: every held push leaves index_last_ts unchanged, so repeated holds walk the market toward its on-chain staleness guard rather than tracking bad data.

​
Conversion

NFT floors are already SOL-denominated, so no Pyth SOL/USD leg is needed. The estimate converts directly to lamports per whole NFT and is pushed via push_index.

​
Listing policy

The estimator can only be as good as the underlying market. NFT markets are admin-listed in v1 and held to depth criteria (sustained volume, listing depth, holder count), with per-market hedge reserves sized against real collection depth so the synthetic market cannot outgrow its reference. See Listing a Market and the funding clamps and the circuit breaker.