The decision doc
Record one decision in 600 words — options, numbers, the call, the figure that reverses it, a pre-mortem — so a price change reopens a number, not an argument.
On this page
A decision doc records one decision so that the people who were not in the room can act on it and the people who were can be held to it. Under a one-sentence question, five parts carry it — the options, the numbers, the decision, what would reverse it, and a pre-mortem paragraph — inside 600 words. A doc that needs more is either two decisions or one that has not been made yet.
Five parts, in the order a reader needs them
The parts follow the decision loop every exercise in this course is scored against, with one change and one addition. Restate becomes the question. Options move ahead of the numbers, because a reader has to know what is being compared before the arithmetic means anything. Decide and what-would-change-your-mind keep their order. The pre-mortem is new: a document outlives the meeting, so it has to say how the decision is most likely to fail.
The budget is a ceiling, not a target: the routing doc below runs to about 350 words before its header.
The doc that argues
The first draft of the routing decision ran to 1,900 words. It opened with two paragraphs on how fast model prices are falling, compared three vendors' current lines model by model, explained routing from first principles, and reached its recommendation in paragraph eleven: "a hybrid approach offers the best balance". No option was rejected with a number. Nothing said what would change the answer. When the next price change arrived, the team re-read all 1,900 words and held the meeting again.
Every one of those faults has one cause: the draft was written to persuade a meeting rather than to record a decision. Persuasion grows; a record has a fixed shape. "A hybrid approach" is the tell: it names no option that was on the table.
Worked: routing short replies to the small tier
The rewrite, part by part, with the numbers from model selection as a product decision and tokens as a cost model.
Question. Do we route short-reply tickets to the small model tier, and on what condition?
Options. A: stay on the mid tier with the 150-token output cap on short replies, $44,550 a month. B: the cap, plus short replies routed to the small tier at a tenth of the price, gated on the eval, about $27,500. C: everything on the small tier, about $5,400, with no eval on the policy and escalation slices. C is rejected: it swaps a known saving for an unmeasured loss, and each point of pass rate on 200,000 drafts a day is 2,000 wrong drafts, about 11 agent-hours a day at 20 seconds each, roughly $10,000 a month at an assumed $30 an hour.
Numbers. The budget is $30,000. A is 1.5× it; B is inside it with $2,500 to spare. B's risk is paid in agent time: each point lost on the 140,000 routed drafts a day is 1,400 wrong drafts, 7.8 agent-hours, about $7,000 a month. Routing saves about $17,000 a month on the input line, so two points lost costs $14,000 and still nets $3,000; three points costs $21,000 and the lever loses money.
Decision. Mid tier for policy and escalation tickets; small tier for short replies while the routed path scores within 2 points of the mid tier on that slice, with no factual error.
Reverse if. The routed slice falls more than 2 points below the mid tier on any weekly run; or either tier's price moves by more than 20%, which re-runs the numbers; or any model update, which re-runs the eval before the router sees traffic.
Pre-mortem. It is six months from now and routing has been switched off. The likeliest story: a model update cost the small tier 2.5 points on short replies; the weekly eval ran on the blended 200-case set, where 2.5 points on 70% of traffic shows as 1.75, under the 2-point alarm; agents absorbed 3,500 extra wrong drafts a day for five weeks before a tenant complained. The decision already carries the mitigation: the eval runs on the routed slice alone.
What each part must contain
Options — two or three, at least one rejected with its number. Staying where you are is usually one, here A. An option nobody proposed is padding.
Numbers — only the arithmetic the decision turns on. The vendor price table goes in a dated appendix, because it will date before the decision does.
Decision — one sentence that can be pasted into a message. Its condition is written in the unit that triggers it: eval points, not dollars. A dollar condition goes stale with every price change; a 2-point condition does not.
Reverse if — a figure and a direction that someone other than the author can observe. "If circumstances change" is not one; "the routed slice more than 2 points below the mid tier on a weekly run" is, and the eval in evals as the feature's test suite produces it every week.
Pre-mortem — Gary Klein's technique (Harvard Business Review, 2007), narrowed to a paragraph: where Klein's team lists every reason the project failed, the doc assumes the decision has failed and writes the likeliest story in the past tense with a number in it, then the line that prevents it. One story, the most likely rather than the most dramatic. A list of twelve risks is a register, and nobody acts on a register.
Why 600 words
At an assumed 250 words a minute, 600 words take under two and a half minutes: inside the time a decider gives a document before a meeting, and short enough for a phone. The ceiling also works as a test. A draft at 1,100 words is carrying something the decision does not need — history, an option nobody proposed, the vendor table — or it is two decisions.
The routing doc and the caching doc are separate for that reason. Caching depends on a vendor rate nobody has read yet, which is why the token bill counts it last, and a doc that waited on it would hold routing hostage to a page nobody had opened.
Reopened, not rewritten
When a reverse-if figure arrives, the doc is not edited to fit. A new doc supersedes it, links back, and says which figure moved; the old one stays, dated, so the next PM can see what was known when the call was made. The weekly cadence in working with an FDE and with research ends in one decision a week, and this is the form it is filed in — beside the one-pager and the PRD, which record the calls made before the work.
In an interview
The behavioural question "tell me about a hard decision" and the execution question "how do you decide with incomplete data" both ask for this document said aloud. Give the options with one number each, reject one out loud, state the decision in a sentence, and name the figure that would have reversed it — then say whether it did. The communicating trade-offs lesson makes the same point for design rounds: the rejected alternative, with its reason, is the highest-signal sentence you say.
The mistake that loses points is a decision with no reversal condition. The interviewer hears certainty about a model's score or price six months out, which tells them the candidate has not yet watched either move. The second is "we went with a hybrid approach": an option that was never one of the options, chosen because nothing was rejected.
Check yourself
1. Your draft decision doc is 1,100 words. Two of its five options were never proposed by anyone, and it opens with a page of vendor history. What goes first?
The two unproposed options, probably 300 words between them, because they make the real rejected option look weaker than it was. Then the vendor history, to a dated appendix. If it is still over 600, look for a second decision inside it — caching, or one tenant's exception — and split it out; a doc that will not fit usually holds two.
2. The vendor halves the small tier's price. Does the routing decision change?
The reverse-if fires, since the price moved more than 20%, so the numbers are re-run, not the decision re-argued. The routed input line is 140,000 × 1,500 tokens × $0.30 per million = $63 a day; it falls to $31.50, so routing saves about $950 a month more. The condition is written in eval points, which did not move, so the decision holds, and a five-line superseding doc records the new figure.
3. The weekly eval alerts on a 2-point drop in the blended pass rate. How large a drop on the routed slice can hide under that alert, and what does it cost?
The routed slice is 70% of traffic, so 2 ÷ 0.7 ≈ 2.9 points on the slice shows as 2 on the blend. At about $7,000 a month per point, 2.9 points is roughly $20,000 a month in agent time against the $17,000 routing saves — underwater before the alarm rings. The alert has to run on the slice, which is what the pre-mortem found.