Concept build$LOUPE
Loupe checks the tile each Pearl block opens against the published weights of open models. Every block gets one label: a named model, unknown, or random-range.
Scroll
Is Pearl's work useful? Loupe does not argue it. It reads the tile each block opens and labels the block: a named model, unknown, or random-range.
Popular open-weight models, stored in the exact tiled INT8 form that Pearl's mining software multiplies, using the tiling and quantization of Pearl's official vLLM plugin. The first table is Pearl's own Gemma-4-31B-it-pearl checkpoint. The registry is free and open-source.
On every hit, Pearl's miner opens one Merkle-proved tile of the matrices it multiplied. The scanner matches the weight side of that tile against the registry. Each block is labelled with a named model, unknown, or random-range.
A free, open-source board shows the label for every block and every pool, plus a weekly usefulness report. 'Model weights detected' is the strongest claim a label makes.
Each row is one block or one streamed share. The tile is checked against the block commitment first; only then is its weight side matched against the registry.
| Block / share | Pool / provider | Tile commitment | Proof | Label |
|---|---|---|---|---|
| <height> | <pool> | 0x… | verifies | Gemma-4-31B-it-pearl |
| <height> | <pool> | 0x… | verifies | unknown |
| <height> | <pool> | 0x… | verifies | random-range |
| <share> | <provider> | 0x… | does not verify | share rejected |
Schematic. Placeholder values; no live data is shown on this page.
Named model:
Model weights detected. The weight side of the tile matches a registered model. It does not show that a paying customer was served.
Unknown:
The weight side matches nothing in the registry. A model missing from the registry reads as unknown, not as fake.
Random-range:
Tile values sit in range; the weight side matches no registered model.
Share rejected:
The tile's Merkle proof does not verify against the block commitment.
Providers can stream their mining shares to Loupe. An OpenAI-compatible router, paid per request in USDC over x402, sends traffic only to providers whose receipts match the model ordered.
A provider streams its mining shares to Loupe.
Each share opens a tile the same way a block does.
The weight side is matched against the registry. The stream becomes a running receipt of which model was multiplied.
Paid requests go only to providers whose receipts match the model ordered.
$LOUPE is a utility token on Base. Providers lock it as a bond to receive routed traffic. The board and the registry stay free.
| 01 | Ticker: | $LOUPE | Utility |
|---|---|---|---|
| 02 | Type: | Utility token. Provider bond and routing-fee share. | |
| 03 | Chain: | Base. Agents pay per request in USDC over x402. | Base |
| 04 | Launch: | Fair launch on a bonding curve. No presale. | |
| 05 | When: | Only after the free board has published about 3–4 weeks of cited data. | Open the board |
| 06 | Holding it: | Providers lock it as a bond to receive routed traffic. | |
| 07 | Bond: | If receipts stop matching the ordered model, part of the bond goes to affected customers. | |
| 08 | Fees: | A share of routing fees, charged in USDC, buys back tokens. | 20–30% |
| 09 | Team: | Vested over 2 years. | 10–15% |
| 10 | Board and registry: | Free and open-source, with or without the token. | Free |
The token comes after the measurement, not before it. Open the board
Today, whether Pearl's work is useful is argued with one-off studies and reports nobody can re-run: an arXiv paper by Basu, a hetu-project sample of 913 blocks, the Oregon Dorm DAO report. No standing measurement exists for every block.
Loupe takes no side. It publishes the label so both camps can cite it.
Basu's own proposed fixes include weight commitments signed by model trainers and higher rewards for customer-supplied matrices. Both need a registry like this one.
A match means a model's weights were multiplied, not that a paying customer was served. A model missing from the registry reads as unknown, not as fake. If Pearl changes its number format, as the draft FP8 fork proposes, the fingerprint tables are rebuilt before labels resume.
Unresolved. Loupe depends on the Merkle-opened tile carrying the raw slice of the matrices before noise is added, and on knowing which operand holds the weights in the official vLLM plugin. If the slice is not exposed, or the plugin pre-transforms the weights, public fingerprints never match and there is no label.
Yes. A miner can multiply real public weights by random inputs and be labelled 'model weights detected'. The label says exactly that and no more. It does not claim a customer was served.
A provider can permute its weights so its tiles read unknown. That makes it unverifiable, not falsely verified. The router does not send it traffic, so the routable supply gets smaller.
The PIP-3 FP8 fork is a draft, and may not ship. If it does, every fingerprint table is rebuilt in the new format before labels resume. The first post-fork numbers are published once they can be measured.
Basu recommends a three-tier transparency label. If Pearl adopts one, the board becomes less necessary. The registry stays open for anyone to use.
If Pearl-mined inference stays a market of one or two providers, the router and the bond see little volume. The board stays free either way.
Then it publishes that. Labels are not adjusted for any audience. A low reading may cost the token its natural buyers; the measurement stands regardless.
When a business pays a company to run an AI model, it usually can't check that it got the model it paid for, and some providers quietly swap in a smaller, cheaper one.
Pearl pays computer owners for doing the same arithmetic AI models do, and every payment leaves behind a small, checkable sample of the exact numbers that were multiplied. Loupe compares those samples with the published contents of popular open models, so anyone can see which model a machine was really running, or whether it was crunching random numbers.
The scoreboard is free. Loupe charges only for sending AI requests to providers who keep passing the check.
