Review: turning captures into memory

Review is every gate in the product that is waiting on a person. Nothing on it is published until somebody decides. It replaced seven separate lists, and it is the only surface a steward needs open.

The layout is a lane switcher over a split pane: the queue on the left, the item on the right. Deciding does not cost a page navigation, and the lane and filter you are on live in the URL, so a link you send a colleague opens on the same row you are looking at.

All open

Everything waiting. The default.

Overdue

Past the decision window the workspace policy sets.

Urgent

Raised as urgent by whatever created the item.

Assigned

Assigned to you specifically.

A filter a lane cannot honour is not applied. The console will not show you a filtered view that quietly ignores the filter and lets you conclude a full queue is empty.

Select a row in the Claims lane. The right pane shows the task: every claim in it, the evidence each is quoted from, any open conflicts on the asset, and, importantly, the audience it would reach if published, with its sensitivity and the number of principals.

1. 1

   Read the claims, not the summary

   A task carries several claims. You are deciding all of them together unless you mark individual ones for rejection. The approve button restates the tally, so it reads Approve 4 claims and not a bare Approve. That wording is deliberate: a slip of a key cannot approve more than you believe you marked.

2. 2

   Check the conflicts block

   If the asset has open conflicts they are listed with their kind and detail. A conflict is a claim that contradicts another already in memory. Resolve the contradiction before you certify either side of it.

3. 3

   Check the audience

   The header states audience and conflict counts. Publishing mints a certificate against a per-item audience, so read who that is before you decide, not after.

4. 4

   Choose the verb that matches your intent

   See the next section. The four are not interchangeable.

Approve

A · Marks the claims verified. They become answerable by Ask immediately. No certificate is minted and no new audience is reached beyond the one that could already see the evidence.

Approve & publish

P · Approves and certifies. It mints a provenance certificate against a named audience. There is no un-publish. The key opens a typed confirmation asking for a reason and for the word publish, deliberately never a single keystroke. Needs knowledge.publish, which reviewers do not hold.

Request changes

C · Sends it back to draft with a note for the author. Use it when the claim is plausible but the evidence or the wording is not yet right.

Reject

R · Refuses the claims. Costs a written rationale; with one already written the keystroke commits and the queue advances, without one it opens the dialog to collect the reason.

S skips: move on without deciding, and the item stays in the queue. It is not a decision and nothing is recorded.

A decision is held for twelve visible seconds before it is sent. The row leaves the queue immediately, a pending-commit strip appears, and you can undo it in that window. It exists because a queue you work at speed is a queue where you will mis-key eventually.

• Approve and reject with a written rationale defer. The queue advances while the commit is held.
• Publish does not defer. It goes through the typed confirmation and commits.
• Enrolments never defer. Admitting an endpoint rotates its credential and returns the token exactly once, so a decision committed twelve seconds later, with the pane long since moved on, would have nowhere to put that token and the endpoint would be locked out permanently. That lane decides through a dialog that shows the token to a human, now.

Relationships

Each item is one proposed edge in the meaning graph: that a person owns a system, that a project depends on a service. Approving mints a durable, queryable edge that the Memory Map and expertise ranking both read. The detail view shows the evidence behind the edge and its blast radius: what else changes if you accept it.

Meaning & judgement

Judgements and automation boundaries a model proposed. They stay proposed until a human confirms them. This is the lane where a model’s interpretation is treated as a suggestion rather than a fact, and it is the one most worth reading slowly.

Automation

A workflow run has stopped at a human-review step. Nothing after that gate has executed. Deciding here resumes or abandons the run, and needs workflow.manage rather than workflow.read.

Source requests

Somebody is asking about a folder on their own machine, either to have it indexed, or to object to it being indexed. Approving one widens a data boundary, which is why it has its own permission. Answer these promptly: it is the only channel a person’s objection has.

Enrolments

Endpoints awaiting admission. Admit endpoint mints the credential and shows it once, under a heading that says so. Copy it before you close the dialog.

Model promotions

Model and role pairs that cleared every hardware floor on real fleet hardware and have had no answer. Eligibility is evidence, not a decision. Promoting changes what every device in the ring loads on its next reconcile. Chapter 14 has the detail.

• Empty the source-requests lane first. A person is waiting on it, and it is the only lane where the delay is felt by somebody outside the console.
• Use Overdue, not All open, as your daily view. The sidebar badge counts the three knowledge lanes together (claims, relationships and meaning), so “clear” there means all three are clear.
• Write the rationale even when you are not forced to. Rejections and publications are the two decisions somebody will ask you about in six months.
• Do not publish to clear the queue. Approve clears it just as well and is reversible.