Evidence lane / Configured-agent session

Connect the runtime. Keep the evidence local.

Launch a checked-in OpenClaw or Hermes bridge against the same synthetic Inbox. Validate the bundle in your terminal, then inspect its local browser projection.

Build an Inbox plan Choose an adapter

What this flow can claim

Real mailbox: Not connected by Clawbotomy
Private evidence: Selected locally; not uploaded
Client identity: Self-asserted
Permission decision: None

The local operator, same-UID filesystem, interpreters, Git, dependencies, and canonical runtime checkout are assumed inside the local trust boundary. This flow does not attest them. The model, tool choices, protocol frames, and evidence claims remain untrusted.

02 / Connect

Choose the runtime you actually operate.

The checked-in bridges expose the same eight mock-Inbox tools and fixed protocol. Support for exact runtime pins is separate compatibility evidence, not a session result.

Agent adapterOpenClawNode.js bridge / OpenClaw runtimeHermes AgentPython bridge / Hermes v0.18.2

Need a plan first? Build it in the browser-local planner.

openclaw.clawbotomy-bridge

OpenClaw launch

Local process

Runs an isolated OpenClaw model/tool loop as the parent of Clawbotomy’s fixed mock-Inbox child protocol.

Before you launch

A reviewed Inbox plan JSON from /preflight
An OpenClaw executable and independently obtained runtime digests
A local Ollama model, or a separately configured provider profile

Use runtime digests obtained independently of the OpenClaw runtime you are evaluating.

Copyable launch command

npm run agent:evaluate -- \
  --adapter openclaw \
  --plan ./clawbotomy-inbox-plan.json \
  --model ollama/qwen3:1.7b \
  --openclaw-bin "$OPENCLAW_BIN" \
  --expected-openclaw-runtime-sha256 "$OPENCLAW_RUNTIME_SHA256" \
  --expected-provider-runtime-sha256 "$OPENCLAW_PROVIDER_RUNTIME_SHA256"

Read the process exit before the score

Evidence lane / deterministic bundle verification

Passedexit 0

A complete bundle was accepted after integrity validation and deterministic replay, and every evaluated case passed. This session does not authorize access.

Findingsexit 2

A complete bundle was accepted after integrity validation and deterministic replay, and one or more evaluated cases produced findings. This session does not authorize access.

Infrastructure failureexit 1

The launcher did not establish one accepted measured bundle. No behavioral conclusion or permission change is supported.

03 / Inspect

Inspect an allowlisted local projection.

After terminal validation, the viewer derives case, tool, state, assertion, and digest receipts in memory. It never renders tool arguments, message bodies, prompts, transcripts, local paths, or raw event payloads, and it does not validate integrity or replay.

Load local evidence

Complete run: select one evaluation-attempt-*.json receipt with manifest.json, summary.json, and cases.jsonl. Infrastructure-only: select the attempt receipt alone.

Select one launcher receipt with its complete bundle, or one infrastructure-failure receipt. Files never leave this browser.

No private runs loaded

The browser holds nothing until you choose the files.

Imported private evidence starts with the fixed launcher and canonical validator in your terminal. The browser is an inspector after terminal validation. It requires a launcher receipt that names and binds the selected files, then derives an allowlisted display model; it does not validate the bundle itself.

Review the sanitized configured-session example

npm run inbox -- validate .clawbotomy/inbox-runs/<runId>

After a valid baseline / Compare

Compare receipts without flattening their provenance.

Comparison is a conditional branch, not a required step. Load at least two launcher-bound bundles to compare case counts.

Comparison waits for local evidence.

Load a run above

04 / Decide

Decide what changes, then test one thing.

Recommendations use only the allowlisted case and assertion IDs already present in the safe viewer projection. Raw prompts, messages, tool arguments, and private event payloads never enter this layer.

Sanitized configured-session summary

Hermes Agent / 2026-07-13

Sanitized configured-session example / Not loaded evidence

Hold permission changes

25 of 36 cases produced findings.

One recorded configured Hermes session produced a replay-validated synthetic-Inbox measurement with findings.

Passed: 11/36
Findings: 25
Tool attempts: 23
State transitions: 7

Allowed claim

One recorded configured Hermes session produced a replay-validated synthetic-Inbox measurement with findings.

Not supported

This does not prove Hermes is unsafe, compare it with OpenClaw, or authorize any production permission change.

01
Load the replay-bound private bundle to see which allowlisted assertions failed.
02
Apply one guardrail against the review-first recommendation.
03
Rerun the same frozen plan and compare the new evidence, not the narrative.

Before a permission change

Measurement is a review input, not an authorization.

A complete run describes one observed session in a synthetic fixture. It does not attest the production deployment, authenticate the adapter’s client identity, or prove repeatability.

Keep human approval and platform controls around consequential actions. Review the full private bundle and rerun the checked-in validator before acting on any result.

Read the evidence boundary