CipherQuery

CipherQuery docs

How private evaluation works.

A seller submits a bounded sample. 0G evaluates the buyer's questions inside protected compute. Only verified answers leave.

  1. AskThe buyer writes plain-language questions.
  2. SubmitThe seller uploads through an encrypted link.
  3. Verify0G returns one verified result per question.

Workflow

  1. Ask and separate access.Add a name and 1-20 plain-language questions. CipherQuery creates separate seller and buyer links. Questions do not need to reference columns or scoring types.
  2. Submit a bounded sample.TLS protects a CSV, JSONL, NDJSON, or flat Parquet sample with 1-50 records, up to 20 columns, and a maximum size of 200 KB. A free browser check runs before any paid request.
  3. Evaluate and verify.The complete bounded sample and every original question are sent in one private 0G request. Every question receives its own score or a clear unable state.

What a score means

A score answers one buyer question against only the submitted sample. Zero means the sample did not meet that question's requirement. One hundred means it fully met the requirement. The result includes a short explanation and confidence level.

No overall score.

Different questions can test different qualities. Combining them would hide useful distinctions and create false precision.

If 0G cannot answer a question safely from the sample, CipherQuery publishes Unable instead of inventing a number. If the full verified result set is unavailable, no scores are published.

Privacy boundary

TLS encrypts the sample in transit. The sample, the full private prompt, and the raw 0G response then exist only in server memory for the duration of the request. They are not written to Supabase, browser storage, analytics, or application logs.

The buyer receives the original questions, question-level result summaries, score or unable state, confidence, and safe verification metadata. The buyer does not receive sample size, row references, values, or aggregate counts.

A sample is still a sample.

Results describe only the submitted records. They do not prove that the seller's full dataset has the same quality.

0G Router and TEE verification

CipherQuery sends one OpenAI-compatible request to the 0G Router using private trust mode. The Router selects the configured private-capable model and returns the inference response with execution metadata.

0G Private Computer
Runs the model evaluation in protected compute so the seller's sample is not exposed to the buyer.
TEE verification
Confirms that 0G reports protected execution for the request. It proves the execution path, not that every model judgment is correct.
Fail closed
A timeout, authentication failure, missing TEE trace, invalid JSON, unsafe evidence, or partial result set publishes no scores.