> ## Documentation Index
> Fetch the complete documentation index at: https://docs.genesiscomputing.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Prover Fuzzer Agent

> Agent for designing synthetic challenge data, separate from Prover behavior fuzzing.

> [!NOTE]
>
> The **Prover Fuzzer** is an early-stage system agent, not a Prover seat or the
> behavior-fuzz workflow. Behavior fuzzing is documented in
> [Genesis Prover → Behaviour Fuzzing](https://docs.genesiscomputing.com/docs/Genesis-Prover/Behaviour-Fuzzing.md).

## Availability

The server rollout setting is `fuzzer.enabled` (boolean, default `false`, read at
call time). When enabled, Genesis seeds the built-in **Prover Fuzzer** agent.
Turning the setting off prevents its tools from running; a previously seeded
agent may remain listed.

## Current behavior

The agent designs synthetic challenge batches from a commissioned request. It
can read source profiles (column types and statistics, not source values), read
the named pipeline text when available, add challenge and filler rows, check a
batch's keys and side limits, and seal finished batches as a read-only dataset.
The seal records the version and hash of the pinned Fuzzer constitution used to
design it, and reports pipeline text the agent could not read and whether source
values were included; the current profile-only workflow keeps real source values
off. Before using Fuzzer tools, the agent reads the constitution and its guidance
for reading pipelines, selecting attacks, and arranging rows.
The Fuzzer does not load a pipeline, compare results, judge an outcome, or affect
a Prover check or certification. The Fuzzer is not connected to Prover.

Start with `fuzz_begin`, passing the request object unchanged. Then read each
commissioned source relation (or the commissioned query) with `fuzz_read_source` and each named pipeline with
`fuzz_read_pipeline`; design one attack at a time with `fuzz_add_challenge`, add
ordinary surrounding rows with `fuzz_add_filler`, and call `fuzz_check_batch`
before `fuzz_seal`. A refusal includes the reason; correct its stated input or
access issue before trying again. A batch is limited by its commission budget,
and a seal can include up to six batches.

After loading a sealed batch into a safe copy, call `fuzz_check_loaded` once for
each side before running that side's pipeline. It verifies the loaded tables
against the seal and returns a receipt; if the inputs differ, a comparison on
that batch is not a pipeline result. This check belongs to the caller, not the
Prover Fuzzer, and never runs either pipeline.

For a Snowflake side, export a sealed batch with `fuzz_export_batch` using
`export_format="snowflake_sql"` and load the returned scripts into a marked
sandbox schema before checking it. The tool reference lists the type and exact-name
overrides for schemas whose columns differ from the seal.

For commission and batch-authoring tools, see
[`fuzzer_tools`](https://docs.genesiscomputing.com/docs/Genbot-Toolkit/Additional-Tool-Groups#fuzzer-tools); for batch
exports and checking the loaded copy, see
[`fuzzer_caller_tools`](https://docs.genesiscomputing.com/docs/Genbot-Toolkit/Additional-Tool-Groups#fuzzer-caller-tools).
