Verifiable is not replicable

01The inversion — least regulated, most disclosed

The catalogue's implicit assumption was that regulatory pressure is what forces methodology disclosure. The fifth administrator breaks that assumption from below.

Inferra's documentation states, in plain language, five construction facts about its GPU-hour index: the primary input ("real market rates are the main input, drawn from the liquid on-demand rental market"); a secondary input of the platform's own settled trades with an explicit weight cap ("its weight is capped at 20%, so the index can never be driven by Inferra's own flow"); outlier handling ("outliers are filtered out before anything is blended, so a single strange quote cannot move the price"); smoothing ("the result is smoothed over time, so the price tracks the trend rather than every twitch"); and a stale-data rule ("if the market goes thin and there is not enough real data to trust, the price holds at its last good value rather than printing a number that is not there").

Set that against the four administrators the catalogue worked through. Silicon Data publishes a four-step process description and a change log, but no aggregation statistic and no rulebook. Ornn's October 2025 methodology paper was published and then taken down. Kalshi's derived curves rest on a conceptual blog post. Compute Desk publishes nothing. None of the four states a self-trade policy, an outlier rule, or a stale-data rule in public. A permissionless Solana protocol with roughly a month of index history — reportedly live since mid-July 2026 — discloses more of its construction than the administrators CME, ICE, Kalshi, and the American Innovation Exchange intend to settle billions of notional against.

The comparison table is the piece. Same axes as the catalogue, fifth column added.

02The 20% cap — a designed answer to circularity

One sentence in Inferra's documentation does something none of the four incumbents has published an equivalent of.

Our forward-curve piece worked through the general failure mode of derived and self-referencing indices: an index that takes input from prices which settle against it can, in the limit, become a market talking to itself. The section on the derived-curve problem asked what a curve inherits from the index beneath it; the prior question — what an index inherits from its own venue — is the one Inferra's designers evidently asked themselves. Their answer, verbatim from the documentation:

Credit where it is due, and precisely: this is a published, quantified bound on the administrator's own flow as an index input. No equivalent exists in public from Silicon Data (whose weighting is proprietary), Ornn (volume-weighted, contributor concentration undisclosed), Kalshi (whose derived curve is 100% its own order flow by construction — the external anchor sits one level down, at settlement), or Compute Desk (nothing published). A lender running the five-check validation program from our hedge-program design would ask every administrator this exact question — "how much of your index is your own book?" — and Inferra is the only one of the five whose public materials answer it with a number.

Two qualifications keep the credit honest. First, the cap is asserted, not evidenced: no published parameter set, no data, and — see §04 — no recomputable print stream against which a third party could test whether the cap binds in practice. Second, a cap on one input is not a policy on the blend: the aggregation statistic for the 80%+ external leg (mean? median? trimmed? volume-weighted?), the smoothing constant, the outlier thresholds, and the source list are all undisclosed, so the bounded input sits inside an unbounded discretion.

03But the loop is still closed

A 20% cap bounds circularity. It does not break it — and Inferra's own documentation describes the closed loop with something close to pride.

The documentation is explicit that the index is not a detached benchmark: "It is the price a renter pays to provision a machine and the price a trader prices a position against. Real demand for compute anchors it, and trading it feeds back into what the next buyer pays." The marketplace prices provider capacity "at the live GPU-hour index"; perpetuals and dated futures are "marked continuously" against it and cash-settle on it; and the whitepaper completes the circle: "push the price up by trading and you also pay more for the compute you buy, so the loop stays honest."

That last sentence is an incentive argument — a trader who distorts the index pays the distortion back when they rent. It has real force for a participant who both trades and rents. It has none for a participant who only trades, or whose derivatives position is larger than their compute bill, which is what a derivatives venue exists to make possible. The whole reason IOSCO Principle 7 anchors benchmark design in observable market transactions is that "the market" is supposed to mean something outside the administrator's own settlement web. Inferra bounds the direct input channel at 20% — but rentals priced at the index are not independent observations of a market, they are the index echoing back through its own venue, and the documentation does not say whether index-priced Inferra rentals count inside or outside the capped 20%. If marketplace rentals execute at the index and also inform it, the effective self-reference exceeds the nominal cap.

The instructive comparison is the Kalshi case from our forward-curve work. Kalshi publishes less than Inferra — a conceptual blog post against Inferra's five stated construction facts. But Kalshi's ladders settle on Ornn's OCPI, a genuinely third-party index: the one structural property that prevents a market from marking itself. Inferra publishes more and settles on itself. Between an opaque index with an external anchor and a disclosed index with a closed loop, the structural advantage sits with the external anchor — disclosure can be added later; an anchor is architecture.

04Verifiable is not replicable

The piece's central distinction — and it generalises to every "transparent because on-chain" claim in market infrastructure.

The reported architecture — index determinations published on-chain, immutable, timestamped, publicly readable — is a genuinely novel answer to one benchmark-governance problem. IOSCO Principle 9 concerns transparency of benchmark determinations: can users see what was published, when, and know it was not silently revised? On-chain publication answers that mechanically. No administrator of the four incumbents can prove its history was never restated; Silicon Data demonstrated the problem concretely when it restated its full A100 history upward 35–40% in December 2025, as the catalogue records. A print on Solana cannot be quietly rewritten. If Inferra's determinations are on-chain, every one of them is auditable forever — a property worth taking seriously as infrastructure, which our second-settlement-layer work examines in the equities context.

But Principle 9 is the transparency of outputs. Principle 11 concerns the content of the methodology: sufficient detail for a user to understand how the benchmark is derived and to assess its representativeness. The test is replication — could a third party holding the inputs recompute the number? Against that test, Inferra's disclosure fails in exactly the way the incumbents' does: no aggregation statistic for the external leg, no smoothing parameter, no outlier thresholds, no source list, and a reference configuration defined only as "one card, normalized" (the SKU table pins memory and form factor — 80GB SXM, 141GB SXM — but not interconnect, cluster scale, or the multi-card normalization rule). Publishing every output on-chain while withholding these is transparency of the answer, not of the question. A user can verify that a given value printed at a given block time; they cannot verify the value was correct.

And there is a sharper local finding. This review could not confirm the on-chain publication claim itself from Inferra's primary sources. The documentation's reference page lists three addresses — the $INFERRA token, USDC, and the escrow program — and states that "everything on Inferra that touches money is on Solana and public." The whitepaper says "a per-model oracle publishes the GPU-hour price the exchange runs on," without saying where. The roadmap promises a "price API." No published oracle or index account address appears anywhere in the documentation. So as of this reading, a third party cannot even locate the feed on-chain from the project's own materials, let alone recompute it: today, the verifiability is itself an unverified claim. Nothing in this cuts against the design — it may simply be young documentation — but it moves the index's actual disclosure position closer to the incumbents' than the reported architecture suggests: prose methodology, unpublished parameters, and a determination stream you take on trust.

05Four roles, one entity

The conflicts stack, mapped against the governance principles the eventual CFTC-track administrators will be measured on.

IOSCO Principles 1–5 concern governance: the administrator's overall responsibility for the benchmark, oversight of third parties, and — Principle 3 — the identification, avoidance, and management of conflicts of interest. The catalogue found the four incumbents publish nothing on any of these. Inferra publishes nothing on them either — and its conflicts are structurally deeper, because the roles the rest of the complex distributes across firms are collapsed into one protocol:

Each cell compounds the others. The venue profits from volume; the volume settles on the venue's own index; the fees accrue to a token the same entity issued; and the protocol is the counterparty to the leveraged positions marked against the number it computes. A protocol that is long its own token, counterparty to its traders, and administrator of the mark is carrying every conflict IOSCO Principle 3 was written to force into the open, simultaneously, with no published conflicts policy, no oversight function, no revision or cessation policy, and no audit. The reported acknowledgment that the oracle currently runs on Inferra's own servers — with a decentralized feed described as in progress — could not be confirmed from the documentation; if accurate, a young protocol naming its own single point of failure is behaving better than a silent incumbent, but self-disclosure of a conflict is not mitigation of it. The disclosure tells you where the risk is; it does not move the risk anywhere.

One further observation the incumbents deserve in fairness: their unbundling is real. ICE and CME outsource administration entirely; Kalshi lists but settles externally on Ornn; Architect runs two venues on two administrators. The market structure forming inside the perimeter has independent administration as its default. The one forming outside it does not — and the whitepaper-to-docs drift this review caught in passing (the whitepaper has rentals settling in $INFERRA across five models; the current docs settle rentals in USDC across six) shows parameters moving with no published revision policy recording what changed or when. Principle 8 material, in miniature.

06What on-chain publication solves, principle by principle

The optional second table, because the mapping is the argument in compressed form.

07Where this goes

Could an on-chain index pass an Appendix C review — and what is the right end-state?

Nothing about on-chain publication is disqualifying under CFTC Part 40 / Guidance Appendix C — the review asks whether a settlement index is reliable, representative, and resistant to manipulation, not where its outputs live. But the gap list for any on-chain administrator that wanted to carry settlement weight follows directly from the table above: a published aggregation statistic and parameter set; a disclosed source list, or at minimum source-count and concentration statistics per determination; a reference-configuration specification at the level our catalogue demands of the incumbents; an oversight function with named conflicts management; versioned methodology with a change and cessation policy; an external anchor — or at least a demonstration that index-priced own-venue rentals sit outside the capped input; and, for the trust claim to mean anything, a published feed address and a decentralized publication path so that the administrator's own servers are not the sole writer of a benchmark its own derivatives settle on.

The interesting conclusion is that the end-state is a complement, not a substitute. A conventional rulebook without on-chain publication leaves users trusting the administrator not to restate history — a trust Silicon Data's December 2025 restatement shows is not free. On-chain publication without a rulebook is an immutable record of a black box. The two solve different principles — 11 and 9 respectively — and a compute benchmark that published a full parameterized methodology and wrote every determination to a public chain would be strictly more credible than anything the complex currently offers, inside the perimeter or out. The incumbents could adopt the publication layer tomorrow; Inferra could publish the rulebook tomorrow. Whoever does both first sets the disclosure benchmark for the asset class — and the fact that a month-old permissionless protocol is currently closer on one axis than four CFTC-track administrators are on either should discomfort exactly the right people.