## The problem
Two onchain instruments both advertise 8%. One is a 30-day T-bill wrapper you can
exit at par; the other is a leveraged loop that pays 8% until it doesn't. The APY
is the same number. The risk is not, and neither is what you actually receive
after credit and execution costs at the size you intend to trade.
Nobody publishes the comparison. Aggregators list yields; none of them tell you
what fills at $250,000, or whether your wallet is even allowed to hold the asset.
## What Fenorium does
**Prices at your size, not at the screen.** You set position size, holding period,
execution cost and per-tier credit haircuts. Every rate on the page is recomputed
against *your* assumptions. They are labelled as the reader's assumptions, not
our opinion — we publish no view on which instrument is preferable.
**Refuses to score what it cannot measure.** Evidence is either *measured* (read
from a chain or an API) or *attested* (a document, a verbatim quote, and its
sha256). Price, size, liquidity and redemption accept measured evidence only. An
instrument with nothing admissible is not scored, not estimated, not interpolated
— it is excluded with the reason attached. Of 295 instruments, 74 publish no rate
at all and 113 have no measured depth. We say so instead of inventing a number.
**Never merges rate kinds.** Executable, indicative, realized and issuer-quoted
stay separate, and the basis travels with the figure. A trailing vault APY and a
live borrower bid are not the same claim and are never averaged.
**Probes eligibility against the real contracts.** Compliance registries are read
directly on mainnet, not assumed: BlackRock's BUIDL gates through Securitize's DS
Protocol, Ondo's OUSG through its own KYC registry, USDY transfers freely despite
carrying a sanctions oracle. Four distinct identity systems, none of them ERC-3643.
This part had to be built by reading contracts rather than simulating transfers.
A transfer simulation misclassifies BUIDL, because DS Protocol checks balance
*before* compliance and returns a balance-shaped revert to a wallet that would
also have failed the gate. Registry detection is authoritative; the simulation is
the fallback.
## The QVAC track: what a local model cost, and what it broke
The dashboard ships with a local assistant. The model never states a number. Its
only job is to choose one tool and fill typed arguments; the arithmetic is done by
ordinary code against data that never leaves the machine. Ask "find instruments I
could buy at $250,000 with TIER_A credit maturing within 200 days" and it returns
every match, unsorted, alongside the count excluded by each individual reason —
because a ranked list would be a recommendation, and the counts are what make an
empty result legible.
We measured three models over five questions covering five tools:
| model | tool chosen correctly | mean |
|---|---|---|
| **Qwen3-1.7B** (shipped) | **5/5** | **17.5 s** |
| Qwen3-4B | 5/5 | 80.5 s |
| Llama-tool-calling-1B | **0/5** | 2.1 s |
The 1B is the result worth reporting. It is eight times faster than what we ship,
and despite being named for tool calling it never emitted a tool call. It did not
fail loudly — asked which instruments matched, it answered from parametric memory
about Treasury bills that are not in this universe, and contradicted itself inside
four sentences ("the 200-day maturity would be a 30-day T-Bill").
Every guard in this codebase assumes the model uses the tool interface. A model
that quietly declines to use it walks past all of them at once. Buying latency by
shrinking the model costs the one property this product cannot trade away, so the
middle model ships. The benchmark is in the repository and reproduces.
## Buying, and the eligibility gate
Base Sepolia mocks make the permissioned path demonstrable without real money —
necessarily, since the real funds are closed to anyone who is not a qualified
purchaser. Connect a wallet and the table relabels itself for *your* address.
mPT is holdable. mGT refuses you by name. Press **Get verified** and a real EAS
attestation is issued to your address; you register it with the token, and the
contract independently re-checks schema, attester, recipient, revocation and
expiry before opening the gate.
The gate fires *before* balance arithmetic — deliberately mirroring what we
measured on Ondo's OUSG rather than what we assumed.
## Verification
363 Python tests and 44 Forge tests pass. The ingest tests replay recorded API
responses rather than calling any venue, so the suite runs offline and
deterministically. An anti-lookahead guard rejects evidence postdating the
as-of date; it has fired on our own refresh scripts twice.
Deployed on Base Sepolia — EAS schema 22 Aug 18:36 UTC, both mocks 22 Aug 18:57.
All 19 deployer transactions fall between 22 Aug 18:36 and 23 Aug 12:44 UTC.
## What it does not claim
Fenorium is not registered with or regulated by the SEC. It takes no custody,
exercises no discretion, publishes no view on which instrument is preferable, and
receives no compensation from any issuer or venue. Sorting is on objective factors
only. You sign from your own wallet; we never send the transaction.
Hacki