Human review limits
Plan evidence capture, proposal size, event history, and generation inputs.
Review limits apply to the saved proposal, its captured evidence, and any generation or projection work. They are separate from general storage quotas and App-specific retention policies.
| Review resource | Maximum |
|---|---|
| Sources per review | 20 |
| Captured bytes per review | 20 MiB |
| Payload per review | 1 MiB |
| Reviews per App | 1,000 |
| Events per review | 1,000 |
Generation and projections
| Resource | Maximum |
|---|---|
| App reviews considered by a generation or projection | 100 |
| Source bytes | 40 MiB |
| Parsed characters | 5 million |
| Result size | 6 MiB |
| Quote in an evidence view | 20,000 characters |
A coordinated generation plan can contain one to four requests. Its bounded correction allowance can correct a completely received answer that fails compilation; it does not authorize automatic replay of failed or uncertain provider calls. The complete generation deadline is 600 seconds.
Workspace model budgets also apply. Generation reserves estimated usage before work, settles known usage, and can retain reservations when usage is unknown. Inspect an uncertain receipt before creating another paid intent.
Plan useful reviews
Choose only the relevant sources and keep payloads focused on the decision. Splitting a proposal should preserve meaning and evidence rather than hiding a dependency between separate approvals.
A source accepted by general storage may still exceed the review capture limit. Capturing evidence and generating proposals are different stages, so a valid capture does not establish that the combined generation input fits every bound.
The ledger is not a separate legal archive or retention guarantee. Whole-App purge removes associated reviews. See history, source capture, and uncertain requests.