Agent Evidence Observer (prototype)

Record brokered agent writes and check selected native framework reports offline.

Use it in agent harnesses and evaluation pipelines that need replayable action records or a consumer gate for retained reports.

Quick start

These commands install a pinned source snapshot. Python 3.12 or later and uv are needed:

git clone https://github.com/probityai/agent-evidence-observer && cd agent-evidence-observer
git checkout f68c8538639aea1515ea357a85a20739004e66fd
uv venv .venv --python python3.12 && uv pip install --python .venv/bin/python .
.venv/bin/agent-evidence-observer demo ./sample-run
.venv/bin/agent-evidence-observer verify ./sample-run

The two commands print:

{"noDetectedGap": "true", "retryReplayed": "true", "status": "verified", "witnessScope": "PEER"}
{"noDetectedGap": "true", "status": "verified", "witnessScope": "PEER"}

sample-run/ now holds the signed packet, the history, the public keys and the file tree. The demo writes a file, retries it, and the retry returns the first effect. noDetectedGap means the broker's file-tree snapshots found no change it did not make. It does not mean every agent effect was seen.

Status

The root distribution is version 0.0.1. Every record says witnessScope: PEER: the witness key runs on the same host under the same operator, so the records show what this broker saw, not what an independent observer saw. Linux runs can launch a fixed workload under bubblewrap with only the broker's socket exposed. Running an unmodified agent that way, and a separately operated witness, are the next steps.

Documentation

Start with worked examples to follow a practical question, then open its technical reference for commands and exact evidence.

License

Apache License 2.0; see LICENSE.