History

Your interview process has a history. Every version of it is kept.

AI is now editing how you interview — a round added from a sentence, a brief rewritten mid-call and published to the round. So before any change to a preset, a loop, a set of criteria or the notification policy, the app records how things stood. Open any of them and read who changed it, when, and exactly what changed, in plain words.

macOS app, Windows next · free for ten interviews a month, no card

Before, not after.

The version is taken before the write, so what you get back is the state that existed — not the state after somebody tried to fix it. It is the whole configuration, hashed, so ten edits to one round are ten versions and ten saves of an unchanged one are none. Under view-as, the person recorded is the owner who was really acting, with the seat they were in beside them.

The history section on a preset: a list of versions, each with how long ago, who, and what kind of change — edited this preset, changed the materials, changed the criteria.
The history section at the bottom of a preset. Every write that touched it, and only those: when, who, and what kind of change.

Read the difference in sentences.

Pick a version and it is shown against now, field by field, in words rather than in a diff of JSON. A length that changed. A person who came off the panel. A criterion that was added. The lines of a brief that moved.

“Duration was 45 → now 60.”
A field that changed, as was and now.
“Sam came off the panel.”
A person, by name, not an id.
“A criterion was added: Debugging under pressure.”
A row that appeared or went.
“3 lines of the brief changed.”
A paragraph, as the lines that moved — with the rest folded away.
One version of a preset opened against now: a heading giving the date and who made the next change, then lines reading was and now for a duration, a member who was removed from the panel, and three changed lines of the brief shown in place.
One version, opened. As it stood then, against now; the next recorded change is named, because a write that changed nothing leaves no version.

The assistant’s edits are covered without it knowing.

A round the assistant adds to a loop, a brief its counterpart in the room rewrites and publishes — each goes through the same route a person’s edit does, so the version is taken first and the history shows the change like any other. In the list it reads as “added a round from the assistant” or “changed the materials”, with the person who asked as the actor. That is the whole point: when something has changed and nobody remembers asking for it, the history says what, when, and from where.

Where it does the work: the assistant in the app can add a round to a loop and point it at a preset; the interviewer’s assistant in the room can rewrite the brief and the questions and publish them back to the round, or keep them to that one interview. Both leave a version behind. Neither can touch a candidate, and neither is in this history at all when it only reads.

Configuration only. Never candidates.

What is kept is how you interview, not whom. No candidate, interview, scorecard or member row is ever in a version; material attached to one interview belongs to that sitting and is left out; a dataset travels as its fingerprint and size, not its contents. A copy of the pipeline sitting outside every retention rule would be the most sensitive thing we hold, so it does not exist.

Presets
The room a round stages: repositories, layout, terminals, brief, materials, the AI policy and the checks.
Loops
Which rounds a role interviews through, in what order, and who runs each.
Criteria
What the panel rates in a round — the names, the guidance, the kinds.
The notification policy
How long before an interview people hear, how long a scorecard may sit, how fast a candidate is told, quiet hours.
Panels and teams
Who interviews for a role, and which teams a preset belongs to.

What it is for.

Three questions nobody could answer before, and one that will be answerable soon.

“This candidate was assessed under these criteria.”
Answerable a year later, when it matters — the criteria as they stood the week of the interview, not as they stand now.
“Who changed the technical screen, and to what?”
The person, the day, and the difference, in sentences. Nobody has to remember.
“What did the assistant change?”
The same list, with the same actor. An AI edit is a version like any other, which is the only way to be comfortable letting it edit at all.
“Put it back the way it was on Tuesday.”
Not yet. Restore of one object is the next item, and it needs everything above to be safe.

What is not built yet.

Said here rather than implied. The full security posture, including the rows that say no, is on the security page.

Restore a version, one click
One object — a preset, a loop, a set of criteria — never the whole account, and through the same route that wrote it, so the restore leaves a version behind too and is never the one thing you cannot undo. It is the next item; it needs the diff to be safe, and the diff is what shipped first.
A nightly version
Today a version is taken when something changes. A nightly one, deduplicated so an unchanged month is not thirty rows, is what makes “show me our process in March” answerable a year later.
A loop’s own history section
A loop’s versions are recorded and served; the section that shows them on the role is not drawn yet.
The assistant saying what it changed
Its edits are kept like anyone’s. What it does not yet do is announce each one in the conversation with an undo beside it.

Let the assistant edit your process. Keep every version.

Ten interviews a month, free, on the whole product. History is on every plan; a preset’s or a round’s history is readable by anyone who may change presets, and the whole account’s by whoever may configure it.