Trust and permissions
Project trust level
5/10
Limited scope
Permission mode
Trust is 0 out of 10, so new work is blocked by design. You can still repair, restore, or explicitly override the lock. An override grants one level and is recorded as your decision.
Project performance
Average of completed task scores
—
Task contract
Nothing is scored unless it was agreed here first. If a violation is not covered by this contract, Earnest must not punish it.
Edit the contract
Current task
Scenario
Earnest enforces the contract at the filesystem boundary, then scores what actually happened.
Open the sample project in a new tab (opens in a new tab) — this is the page the agent edits, served from the disposable workspace.
Approval needed
Task status
No task has run yet
Task score
—
Choose a scenario and run a task to see how Earnest verifies the work.
Verifier
Your review
Automation cannot judge whether this result suits your project. Your review is what makes the score final.
Score breakdown
Weighted dimensions, each rated 0–10. Evidence is shown next to the number so the score can be argued with.
| Dimension | Weight | Rating | Evidence |
|---|---|---|---|
| No task scored yet. | |||
Agent activity
What the agent tried, and what the policy did about it.
- No actions yet.
Changed files
Real diffs from the disposable workspace, not a mock-up.
No changes yet.
Activity ledger
Trust changes
- No changes yet.
All events
Every block, approval, score, and repair is recorded here. The ledger is the record of truth when an agent's summary disagrees with reality.
- No events yet.
What this demo is
- Agent actions are scripted, not a live model.
- Permissions are enforced in code, then gated by your trust tier.
- Invariants and scoring weights are frozen before a task runs.
- Every event is hash-chained. Disputes append; nothing is edited.
- A real deployment would need a sandboxed agent adapter.