Operations

The debrief that names the disagreement

Your panel's scorecards become a debrief that quotes both sides of every disagreement, and refuses to average a 2 and a 4 into a 3.

Join the waitlist See what it produces

Hiring loop

Run the hiring loop

Who it is for: Hiring managers and heads of talent whose debrief is run from memory, and decided by whoever speaks first.
What starts it: A candidate's ticket enters your Interviewing or Debrief column; On a board move

Coming soonRun the hiring loop is built, but not ready yet, so it cannot be turned on today. In Aestus you can ask to be told when it is ready.

A hiring loop does not break in the interviews. It breaks in the last ten minutes, in a room where four people are trying to remember what they thought a week ago.

Two of them rated the same answer 2 and 4. Somebody says "so, a 3", and everybody nods, because a 3 is a number and nobody wants to be the one holding the meeting open. What has actually happened is that the most informative thing the loop produced — two experienced people watched the same answer and disagreed about whether it was good — has been deleted, and replaced with a figure none of them would defend.

So this use case has one job, and the rest follows from it: show the panel where it disagrees, with both quotes, and never average them. An attribute everybody agreed on does not appear in that section. An attribute two people read differently appears with what each of them wrote, the time they wrote it, and one sentence saying whether they are describing the same moment or two different ones. That last part is usually the answer: one person scored the first attempt and the other scored the attempt after a hint, and the panel has never agreed whether "corrects with a hint" counts at that level. That is a question a room can settle in ninety seconds. A 3 is not.

It never scores anyone. Every number in the debrief is a number an interviewer wrote down, quoted with their name against it. The document computes nothing about a person — no average, no composite, no total — and the review step in front of it refuses one, quoting the sentence it found. It also runs on one candidate at a time, so there is no version of this that ranks a pile of CVs; that is the shape in which an automated judgement about people does the most damage and is the hardest to see.

The recommendation comes from five words and nothing else, and one of them is "no decision — evidence gap". It is not a hedge. It is mandatory when two scorecards conflict on a must-have that no third interview tested, or when a must-have was never tested at all — which happens constantly and is almost never said out loud, because the loop is over and everybody would rather decide than book another hour. Saying it is the useful output: it names the one interview that would close the gap.

The person who never wrote a scorecard is named. Not softened, not omitted, not averaged into the three who did. Their interview time is there and so is how long ago it was, because a debrief that reads as complete while a quarter of the evidence is missing is worse than one that admits the gap.

What happens before the interviews matters as much. Each interviewer gets a pack naming the attributes they own and the attributes the other three are covering, so two people do not spend an hour on the same system-design question. Three things from the CV worth probing, each quoted with the file it came from. A gap in dates may be named as a gap, with the instruction to ask what the person worked on and not to ask why. And a standing line of what nobody may ask — age, family, health, citizenship beyond work authorization, salary history, current compensation — because that question usually gets asked by somebody trying to be friendly.

Nothing about a candidate is written into workspace memory. That is not a toggle somebody could turn on: the steps carry no memory tool. A durable statement of fact about a job applicant, recalled into a prompt in some unrelated run months later and read by somebody who never saw the evidence, is the harm worth designing against, and the only reliable control is not to write it.

There is no applicant tracking system behind any of this. Greenhouse, Ashby and Lever are not connected in Aestus, and a page that implied otherwise would be selling something that fails the moment somebody trusts it. What exists is the candidate's ticket: the panel block, the two attached files, and the comments your interviewers already write. Every document it produces prints the list of what it could not see, in full, on every run — a section that only appeared when something went wrong would be read as an error rather than as the standing truth it is.

What you get

The artifact this use case produces, in the product view you would read it in.

Example — fictional team and data. Not a live run or customer result.

Worked example

document

Debrief — Staff Backend Engineer · LIORHEBE-CAND-412

Produced by: Read-only · document artifact on LIORHEBE-CAND-412 · no applicant tracking system read · demo values

Recommendation: No decision — evidence gap. Two of 4 panelists tested the same must-have and reached opposite conclusions about the same answer, and nobody tested it a third time — and 1 attribute in the role brief was tested by no interview at all. The value is one of five the document may use, and this one is mandatory in exactly those two situations, so the paragraph under it names the interview that would close both gaps rather than picking a side.
Where the panel disagrees: The section the whole use case exists for. One panelist rated the answer 2 — "chose range sharding for a write-heavy key and did not notice the hot shard" — and another rated the same question 4, "then caught the hot shard himself when I asked about p99". Both quotes are printed with their author and the time they were sent, and the last column says whether the two are describing the same moment. Averaging them into a 3 is refused by the review step, which quotes the offending sentence back.
Missing feedback: 1 of the 4 interviewers never wrote a scorecard, and is named here with the interview time and how long ago it was — not softened, not omitted, not averaged into the 3 who did. A debrief that reads as complete while a quarter of the evidence is missing is worse than one that admits the gap.
Role brief attributes and coverage: All 6 attributes the role brief lists, each with the interviews that tested it and one of four verdicts — Met, Partly met, Contested or Not tested. "Greenfield design" reads Not tested with no interview against it, which is the row that usually never gets said out loud because the loop is over and everybody would rather decide than book another hour.
Sources: Every quote is a comment on LIORHEBE-CAND-412, read on 2026-09-11, and the role brief is `staff-backend-role-brief.pdf`, revision 1, attached to the same ticket. The paragraph closes with the standing sentence: "No rating was computed; every number is a value an interviewer wrote."
What this did not read: Printed in full on every document, whether or not anything was missing: 4 lines naming what this could not see — No applicant tracking system is connected. Nothing here was read from Greenhouse, Ashby, Lever or any other ATS, and nothing is written back to one. No mailbox, no calendar, no interview recording and no HR system were read either.

How it runs

Each step is one node in the workflow graph the interpreter runs. Nothing here is a prompt chain hidden in a route handler.

Workflow steps

01A candidate's card enters Interviewing
You point that column at this use case once. Every card that arrives afterwards starts one run anchored to that ticket, so everything it writes lands where the candidate's record already lives.
02It reads the ticket, and only the ticket
The panel block your recruiter wrote, the CV and role brief attached to the card, and the scorecards your interviewers left as comments. No applicant tracking system is connected, no mailbox and no calendar are read, and every document it writes says so in full.
03Each interviewer gets a prep pack
The attributes they own, who is covering the others, and three things in the CV worth probing — each quoted, with the file it came from. A must-have nobody owns is named as unowned rather than quietly given to two people.
04A late scorecard gets one chase
Twelve hours after the interview, two sentences naming the role, the ticket and the hours elapsed — and nothing about the candidate. Nothing is sent until a person approves it; decline and the chase lands as a comment on the ticket instead.
05The debrief is written before the meeting
Agreements with their quotes, disagreements with both quotes side by side, everyone who never wrote a scorecard, and every attribute nobody tested. A review step refuses a paraphrase presented as a quote, an invented rating, or a panelist left out.

By hand, and on Aestus

The same job, on either side.

By hand
Four people open the CV at 09:58 for a 10:00 interview.
On Aestus
A prep pack per interviewer when the card enters Interviewing, with their attributes and three quoted things to probe.
By hand
Two people ask the same question and a must-have goes untested.
On Aestus
Attributes are split so no two interviewers own the same one, and anything nobody owns is named in the pack.
By hand
A missing scorecard is chased in Slack, or forgotten.
On Aestus
One two-sentence reminder, carrying nothing about the candidate, sent only after someone approves it.
By hand
The debrief is written from memory, in the meeting.
On Aestus
A document before the meeting, every claim quoting a scorecard with its author and the time it was sent.
By hand
A 2 and a 4 are averaged into a 3 and the room decides on a number nobody believes.
On Aestus
The disagreement gets its own section with both quotes, and the recommendation can read "no decision — evidence gap".
By hand
The person who never wrote anything is forgotten by the time the room votes.
On Aestus
Every panelist with no scorecard is named, with their interview time and how long ago it was.

What Aestus records

Every run is auditable after the fact, which is what makes handing the job over survivable.

The documents
Prep packs and the debrief are artifacts on the candidate's own ticket, with revisions. A second run is a new revision, never a second document somebody has to choose between.
The run
Cost, tokens, duration, every step and the review step's verdict, on the same ledger as every other run in the workspace. Open it and check a quote against the comment it came from.
Nothing else
It creates no ticket, moves no card, changes no field, and writes nothing about a candidate into workspace memory. There is no memory tool on any step of it.

What it connects to

Only the tools this workflow actually uses. Each one has a setup guide.

Questions teams ask

Does it score or rank the candidate?

No, and it cannot. No step of it produces a number about a person: there is no average, no composite and no percentage anywhere in what it writes. Every rating in the debrief is a value an interviewer typed into their own scorecard, quoted with their name and the time they sent it. It also runs on one candidate at a time and has no way to read a pile of applications, so it cannot screen or shortlist either.

Does this replace our applicant tracking system?

No. Nothing is read from Greenhouse, Ashby, Lever or any other tracking system, because none of them is connected in Aestus, and nothing is written back to one. What it reads is the panel block on the candidate's ticket, the CV and role brief attached to it, and the scorecards your interviewers write as comments. Every document it produces prints that list of what it did not read, in full, whether or not anything was missing.

What does it do with the candidate's data?

It reads the files on that one ticket and the comments on it, and writes back to that same ticket. Nothing about a candidate is recorded in workspace memory, so nothing it read can surface in an unrelated run months later — that is not a setting, it is that the steps carry no memory tool. The weekly pipeline digest, which spans candidates, carries no CV text, no scorecard text, no quote and no rating at all.

What if the panel disagrees?

That is the point of the document. A disagreement gets its own section, with both quotes and the times they were sent, and a sentence saying whether the two people are describing the same moment or different ones. When two scorecards conflict on a must-have that no third interview tested, or a must-have was never tested at all, the recommendation reads "no decision — evidence gap" rather than picking a side.

Can it email or message the panel?

Only the late-scorecard reminder, and only after a named person approves it. That message names the interviewer, the role, the ticket and the hours elapsed, and carries nothing else — the step that drafts it has no way to read a CV at all. Decline the approval and nothing is sent: the chase lands as a comment on the ticket and the run still finishes. The prep packs and the debrief never leave Aestus.

All use cases How access is governed

Sources

lib/domain/use-cases/run-the-hiring-loop/seed.ts
lib/domain/use-cases/run-the-hiring-loop/prompts.ts
lib/domain/use-cases/run-the-hiring-loop/panel.ts
lib/domain/use-cases/run-the-hiring-loop/debrief.ts
lib/domain/use-cases/deliver.ts
lib/domain/use-cases/seed-contract.ts

Hand over Run the hiring loop.

Tell us how your team does this today and we will set it up with you.

Talk to us

Run the hiring loop

Who it is for

Hiring managers and heads of talent whose debrief is run from memory, and decided by whoever speaks first.

What starts it

A candidate's ticket enters your Interviewing or Debrief column

On a board move

Coming soonRun the hiring loop is built, but not ready yet, so it cannot be turned on today. In Aestus you can ask to be told when it is ready.

Questions teams ask

Does it score or rank the candidate?

Does this replace our applicant tracking system?

What does it do with the candidate's data?

What if the panel disagrees?

Can it email or message the panel?

Sources

lib/domain/use-cases/run-the-hiring-loop/seed.ts

lib/domain/use-cases/run-the-hiring-loop/prompts.ts

lib/domain/use-cases/run-the-hiring-loop/panel.ts

lib/domain/use-cases/run-the-hiring-loop/debrief.ts

lib/domain/use-cases/deliver.ts

lib/domain/use-cases/seed-contract.ts

The debrief that names the disagreement

Run the hiring loop

What you get

Debrief — Staff Backend Engineer · LIORHEBE-CAND-412

How it runs

By hand, and on Aestus

What Aestus records

The documents

The run

Nothing else

What it connects to

Questions teams ask

Other use cases

Hand over Run the hiring loop.

The debrief that names the disagreement

Run the hiring loop

What you get

Debrief — Staff Backend Engineer · LIORHEBE-CAND-412

How it runs

By hand, and on Aestus

What Aestus records

The documents

The run

Nothing else

What it connects to

Questions teams ask

Other use cases

Hand over Run the hiring loop.