> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentlinkops.com/llms.txt
> Use this file to discover all available pages before exploring further.

# How records, checks and usage fit together

> Understand candidates, watches, observations, local intent, asynchronous jobs, frozen pagination and shared check budgets before building a workflow.

The [AgentLinkOps overview](/introduction) starts with your task. The records beneath that task separate what you want, what a source reported and what a check established. Keeping those records separate makes interrupted work and uncertain evidence easier to review.

## Records and ownership

| Record         | What it means                                                                    |
| -------------- | -------------------------------------------------------------------------------- |
| Candidate      | An imported or discovered source/target row with provenance; it may be unchecked |
| Watch          | An explicitly monitored source placement or destination                          |
| Observation    | A dated result from a specific check, including uncertainty                      |
| Event          | A recorded change for a feed consumer                                            |
| Local intent   | Your declaration that a link is wanted, expected or retired                      |
| Action receipt | Your declared work; later observations do not establish attribution              |

The publisher URL is the source. The cited URL is the target. Source presence and target health are independent. Read [evidence states](/evidence) before combining their results.

Your [ledger](/ledger) stores local intent and notes. Hosted records supply observations and history. [Repository context](/guides/local-context) distinguishes manual judgments from dated search and site facts.

## Jobs and retries

A queued response confirms admission. It does not confirm the page was fetched or a link found. Save the job identifier and read its terminal result. Even a successful job can produce an unknown observation.

For an idempotent write, preserve both the request's idempotency value and its arguments. Replay that request after a lost response. A different intended check needs a new value. Read operation-specific errors before retrying; permission refusal and provider unavailability need a change outside the request.

## Pagination and snapshots

List operations return bounded pages. Treat cursors as opaque values and keep filters consistent while paging. Discovery candidate cursors freeze a stored result snapshot; restart without a cursor when you need newer results.

Event consumers apply a full page before storing the next cursor. Source, destination and workspace event feeds have independent cursor histories. Expiry requires the supplied recovery procedure, not an assumption that nothing happened. See [sync and export](/guides/sync-and-export).

## Usage and coverage

Source and destination checks share allowance. Reserved, consumed and released units describe different stages of work. Discovery and overview usage remain separate from checks. A configured schedule does not guarantee an exact completion time.

Reports describe the selected tracked or imported dataset. Missing supplier or corpus rows cannot establish whole-web absence. [Competitor research](/guides/competitor-research) keeps coverage attached to comparisons; [availability](/capability-status) describes which lanes can run.

The [common schema reference](/reference/schemas) explains recurring field shapes. [Error recovery](/reference/errors) maps response codes to next actions, and the [documentation changelog](/changelog) separates new reference coverage from product releases.