Live in this page · no account · no upload
Don’t trust the model.
Check it.
Every RAREXT answer names the weights, the engine, and the exact token trace that produced it. This page runs a committed model on your device right now, then shows you the arithmetic that proves it — the same check anyone can repeat.
Live proof
running…A committed model class runs here. Change the seed, or tamper with one token, and watch the verdict follow the bytes. Token boundaries are colored; the output is their exact concatenation.
- model
—- trace
—- replay
- —
Runs on load, in your tab. Nothing is sent anywhere.
- 0 accounts
- Value before signup
- 0 bytes uploaded
- The model runs locally
- 0 third parties
- No fonts, trackers, or CDNs
How it works
Four steps. None of them asks you to believe a server.
The model and the engine are separate content-addressed artifacts. A provider fetches both, verifies every shard, and runs exactly what it fetched. You get bytes you can check.
-
01
Commit the weights and the engine
Each is sharded into RAREXT blobs under one digest. A provider that cannot reproduce the digest refuses to serve the model — not “logs a warning”, refuses.
-
02
Decode deterministically
Greedy or sampled, with temperature, top-k, top-p, min-p, typical-p, penalties, logit bias, and a stop set. The seed is part of the claim, so a run is repeatable.
-
03
Record the trace, not just the text
Every answer carries the token ids it decoded, per-token logprobs where requested, and the bound it ran inside. The text is a rendering of the trace.
-
04
Verify by recomputation
A reader re-derives the result from the trace, the model, and the seed. Lossy work is a measured distance against a committed reference — never an assertion.
Measured, including the gaps
What it does. What it does not.
The capability claim is checked live against a real 504 MiB Qwen2.5-0.5B container served from 64 verified shards. The audit reports the misses as results, not footnotes.
| Capability | Status | Evidence |
|---|---|---|
| Chat, completions, streaming | yes | 24 SSE chunks = 24 trace tokens; concatenation equals the non-stream text |
| Sampling, stop, logprobs | yes | Seeded replay is byte-identical; 6 scored tokens = 6 trace tokens |
| JSON-schema output | shape | Grammar constrains the decode; it guarantees shape, not truth |
| Embeddings | yes | 2 × 896-dim vectors from the same committed weights |
| Tool calls | model-weak | Protocol implemented and tested; the 0.5B missed the live prompt |
| Vision, audio, image, speech, video | no | Refused by name: HTTP 400 unsupported_content_part |
| Model classes | 1G only | 6/7/8 GB classes are declared shipped: false |
| Throughput on this host | 0.50 tok/s | llama.cpp does 15.2 tok/s on the same 0.5B model (Q4_K_M vs Q8_0) |
| Weights + engine by digest | yes | Both are artifacts; the engine fingerprint is folded into the receipt |
| Tamper refusal, replay, receipts | yes | 34/34 verification phases, including a corrupted shard being refused |
Evidence
The numbers have commands behind them.
Every claim on this page is produced by a script in the repository. These are the four that carry the most weight.
14 / 15 live probes
0 critical failures. Chat, streaming, replay, stop, logprobs, schema output, embeddings, refusal, and receipts — measured over HTTP.
demo:web4-model34 / 34 phases
Includes tamper refusal, engine-as-artifact, a determinism witness, quorum, and enclave isolation.
public registryA bundle you can rebuild
The bundle, the seed list, and the mirror — the record the ecosystem publishes into, with its Merkle root.
galleryArtifacts that check themselves
Exact bytes, a bounded reconstruction, and a generative replay, each verified its own way.
Against the top AI brands
One harness, one weighting — performance 30%, accessibility 35%, UX 35% — on each brand's best-measured surface. The top six are shown; the full table, the method, and the six surfaces that refuse an anonymous visitor are in the benchmark.
| # | Measured surface | Perf | A11y | UX | Overall |
|---|---|---|---|---|---|
| 1 | RAREXT front door | 100 | 100 | 100 | 100 |
| 2 | Google Gemini (DeepMind) page | 98 | 95 | 100 | 98 |
| 3 | Microsoft Microsoft Copilot | 88 | 96 | 93 | 93 |
| 4 | Anthropic Claude product page | 83 | 93 | 100 | 92 |
| 5 | Mistral Mistral site | 83 | 99 | 80 | 88 |
| 6 | You.com You.com | 79 | 83 | 93 | 85 |
The full comparison: method, per-app observations, and screenshots →
Reproduce it in one shell
git clone https://github.com/999svs999/RAREXT-protocol
cd RAREXT-protocol
npm ci
npm run test:runtime # 181 / 181
npm run audit:ai # 14 / 15, 0 critical
npm run test:showcase # 44 / 44
Limits
The claim says what it does not cover.
RAREXT is a runtime and a protocol, not a model. Its quality ceiling is the model an operator commits. Today that is a 0.5B model class on this host, and the ladder says so in public.
- Small models, for now. The shipped class is 1G; 6/7/8 GB need a host that can hold them.
- Slower than native. The JavaScript engine is ~30× slower than llama.cpp here, and it is not claimed to compete.
- Text only. No vision, audio, image, speech, or video path exists.
- Software enclave. The trust anchor is a shipped key, not fused silicon; a hardware TEE is host-side work.
A verifiable answer is not a true answer. It is an answer whose bytes, run, and claim you can reproduce — and it only means something if the claim says what it does not cover.
Open by default
Build with us
Run a provider, write a verifier, port a model, build an app, or argue with the design. The protocol is open, and the reference implementation is public.
Contribute
Setup, tests, commit rules, and DCO sign-off. Start with one focused change.
Port a model
Commit a container and an engine; the same verifier checks the runs they produce.
Write an RFC
Design changes start as an argument in writing. That is where the argument happens.
Verify this registry
Rebuild the Merkle root from the live bundle and compare it with the root the chain holds.