Evidence and provenance

What an evidence item records, how to check a finding against it, and how provenance is kept intact.

Evidence is the record of every read a run made. It is what makes an AGENTEX report checkable: each observed finding points at the reads it rests on, and each read shows exactly what was asked and what came back.

What an evidence item shows#

The Evidence tab lists one item per read, most important first:

Part	What it tells you
What this proves	The findings that rely on this read, or that it supports findings drawn from several reads together.
Observation	The value that was read, in readable form.
Source	The provider and method, for example an RPC method such as `eth_call` or `getAccountInfo`, or a public API.
Context	The block number or slot the read was taken at, and when. Reads with no block (for example a public API) show Latest state: the source's state at the time of the read.
Raw evidence	The exact request and response, collapsed by default.

Use the filters to show only reads cited by a finding, or all sources.

Check a finding#

In Findings, select the finding's evidence button.
The Evidence tab opens filtered to that finding's reads.
Compare the observation with the finding, and check the block or slot.
Open Raw evidence to see the request and the provider's response exactly as recorded.

For EVM networks you can repeat the same call against any node at the same block and compare the result.

Provenance rules#

Every report follows these rules:

Pinned reads. EVM agents pin one block for the whole run and recheck its hash at completion. If the chain reorganised in between, the run fails rather than mixing two states. Solana reads record their slot.
Declared reads only. An agent can only call the read methods its version declares. A read outside that list is refused and never appears as evidence.
Recorded, not re-fetched. Evidence is captured during the run and stored with the report. Opening a report later does not query the chain again.
No invented values. A value that was not read is shown as unknown, never estimated. An empty list stays empty.
Integrity. Each report has a content hash, and its evidence ids are fixed. Exports and share links carry the same hash.

Evidence completeness#

Each read is labelled:

Label	Meaning
Complete read	The source answered in full.
Partial read	The source answered, but hit a limit (for example a log cap). The report states the range actually covered.
Read failed	The source did not answer. Findings that depended on it are unknown.

Untrusted content#

Token names, metadata, revert reasons and responses from public sources are data, not instructions. They are shown as text, stripped of control and invisible characters, and never executed or followed as links unless they are plain HTTPS addresses. An agent cannot be steered by content it reads.

Checked claims#

For Token Researcher, Transaction Inspector, Deployer Investigator and Wallet Analyst reports, AGENTEX also checks each material claim against the recorded evidence. The claim is marked supported, derived from input, unsupported or contradicted. These checks re-derive values from the captured outputs; they produce no confidence score.