Offers and the decision record: writing the reason while you still remember it

Records matter because challenges arrive late. Not the week of the decision, when everyone remembers everything, but months afterwards, when the panel has changed jobs and the only honest thing left is what was written down.

• The criteria used. The scorecard as it stood for this role, on that date. If the criteria changed mid-wave, both versions and the date of the change. A criterion that appears only in the write-up is the classic tell.
• The evidence per rating, per candidate. Not the number, the thing underneath it: the answer given, the sample seen, the incident described. From b5 you already have this, because the kit forced it.
• Who decided and why, in their own words. A name, not a committee. "Dara O. selected candidate B because..." with the reason tied to criteria, not to adjectives about the person.
• The date and the whole slate. Who was in the pool, not only who came out. A record of the winner alone cannot answer any question worth asking.
• The impact check result. The four-fifths ratio for this slate, from b4, run before an offer goes out rather than at the end of the quarter.
• The tool and version in use. Which assistant, which prompt or workspace, what it did and what it did not do. This is what turns "we used AI somewhere in there" into something an auditor can follow.

• Memory does not decay evenly, it decays in your favour. A week later you remember the decision and reconstruct the reasons to fit it. That is not dishonesty, it is how recall works, and it is why the reason has to be written while the evidence is still in front of you.
• Same-day is enforceable, "before the end of the wave" is not. One rule, one day, no exceptions. The moment there is a grace period there is a backlog, and a backlog of rationales is a stack of fiction waiting to be written.
• The layers have different owners on purpose. If the HR partner writes everything, the hiring manager's reason becomes the HR partner's paraphrase of it, and the one sentence that has to be in Dara's own words is the one that is not.
• Retention has two failure modes. Too short and you cannot answer a late challenge. Too long and you are holding candidate personal data with no purpose, which is its own problem under data protection rules. Ask what your retention period actually is, and who set it.
• Record-keeping is an explicit duty in some regimes. The EU AI Act's high-risk regime for employment systems expects record-keeping, logging and traceability; NYC's Local Law 144 expects an independent annual bias audit, a published summary of it, and evidence you gave candidates at least 10 business days' notice. This is education, not legal advice, and the obligations vary by jurisdiction.

• Drafting the offer conversation. What you will say, in what order, how you will explain the band, and what you will say when they ask for more. Rehearsing a hard conversation against an assistant beats rehearsing it live on a candidate.
• Modelling a band position against your own framework. Feed it your levelling framework and band structure - the documents from the b1 pack, which contain no named person's pay - and ask it to check whether the position you are proposing is consistent with your own rules. It is checking your logic, not choosing a figure.
• Writing the honest rejection. The one nobody has time for. Part 3 is entirely about this, because it is the artifact most candidates will actually remember.
• Pressure-testing your explanation. Ask it to play the candidate and push back on your band reasoning. If your justification collapses in four questions from a language model, it will collapse faster with a good engineer on the phone.

These are not guidelines. They are the pay-shaped version of rail 1 and rail 3.

• Useful. Specific enough that the person knows what to do differently. "We went with someone stronger on X" where X is a real capability from the scorecard. Not a coaching essay, one honest sentence.
• True. If they were fourth, do not tell them they were second. The kind lie is the one that gets quoted back when they apply again in six months and get rejected at the screen.
• About a criterion you actually applied. This is the one people fail. Never reach for a plausible-sounding reason that was not in the scorecard and was not in anyone's evidence. A reason invented at the message-writing stage is a criterion invented after the decision, which is the same defect as the reconstructed rationale in Part 1.
• And give them the route. The candidate notice you wrote in b1 says a person decides and says how to ask for human review or an adjustment. The rejection message is where that line has to appear again, in a form someone can actually act on: a named route, not a no-reply address.

• Scale the detail to the stage. Someone rejected at application screen gets a short honest note. Someone who did three interviews and a work sample has earned specifics against criteria, and it is reasonable to offer a short call.
• Say it against criteria, never against the person. "Fewer worked examples of X" is about the job. "Not senior enough in presence" is about the person, means nothing, and is the kind of sentence that reads terribly in a bundle of documents.
• Do not promise a future role you cannot promise. "We will keep you in mind" is either a commitment you will not keep or a sentence with no content. Either say concretely what you will do or say nothing.
• Consistency is part of fairness. Eleven people should get feedback of the same depth against the same criteria. Detailed warmth for the ones you liked and boilerplate for the rest is a pattern, and patterns are what get examined.
• An assistant is genuinely good at this. Give it the criteria and the evidence with the name stripped out, and it will draft something specific and humane faster than you will at the end of a hiring wave. You supply the evidence and the judgment; it supplies the sentences.