Your deterioration service's online read needs a quarter to accumulate while the offline sweep reports hourly - how much launch weight should the offline number carry?
answer
- weigh the proxy by its record
- offline vetoes, online licenses
- what did past offline wins deliver
- fast reads that are not the read
- launch with a dated re-decision
basics
~20 sAs much weight as its track record earns. Pair each past launch's offline delta with the online delta that followed; that history says whether the proxy can license a launch or only veto a candidate, and what a launch carried by it must commit to.
solid answer
~50 sDo not argue the split in the abstract - measure the proxy. For each past change to this service, record the offline delta at promotion and the online delta that eventually landed; the agreement between those pairs is the evidence that says how far the fast number can carry a decision. Most services find the offline metric is a good **veto** and a poor **licence**: a candidate that loses offline never reaches a ward, but an offline win alone does not justify replacing the incumbent. Between the two, buy fast reads that are online but are explicitly not the launch metric - worklist overlap with the incumbent, the share of surfaced patients already flagged, alert load per nurse-shift. And if a launch does ride mostly on the offline number, it ships with a scheduled re-read of the slow metric, a named reversal condition and a per-segment floor.
go deeper
Take away the shape of the trade: a fast number that is only a stand-in, a slow number that decides, and a gap in between that someone has to be accountable for.
Be able to explain why the two roles differ - a wrong veto costs one candidate, a wrong licence costs live quality for as long as it goes undetected.
Bring the interim reads and the commitments: overlap, novelty share, alert load, a dated re-read, a named reversal condition and a per-segment floor.
Make the proxy an instrument the organisation measures: record the offline and online delta for every launch, and state the weight the fast number earns from that record.
## The question is about evidence weight, not about which metric is right Both numbers are correctly defined. The offline sweep metric returns in an hour on a frozen cohort; the online read - escalations that changed care per shift, at a stated alert load - needs a quarter of live shifts to accumulate, because deterioration events arrive only as fast as patients deteriorate. The decision a lead owns is how much of a launch the fast number is allowed to authorise while the slow one is still maturing. There is no universal answer, which is what makes it a judgment call rather than a metric choice. ## Calibrate the proxy against the service's own history The argument becomes tractable the moment it is empirical. 1. For every change this service has shipped, record the **offline delta at promotion** and the **online delta that eventually landed**. 2. Look at the pairs. If offline wins have reliably preceded online wins of a similar sign, the proxy has a record and can carry more. 3. If four of the last six offline wins produced no online movement, the proxy has demonstrated that it is a filter, not a decision, and no amount of arguing changes that. This history is cheap to keep and almost never kept, because it only exists if someone wrote down the offline number at promotion time. Starting that record is often the most valuable output of this conversation. ## Split the pair into a veto and a licence The two roles have very different error costs and should be stated asymmetrically in framing: - **Veto (cheap, safe):** a candidate that loses on the offline metric does not reach a ward. A false veto costs one candidate; a proxy only has to be weakly informative to be worth this. - **Licence (expensive, risky):** an offline win replaces the incumbent. A false licence costs real patient-shift quality until the slow read catches it, and the harm accrues the whole time. Most services should say plainly that the offline number vetoes and the online number licenses, then treat any departure from that as a decision with an owner and a written reason. ## Buy a faster read that is not the read Between an hourly proxy and a quarterly truth there is a band of reads that take days and are genuinely about live behaviour: - **Worklist overlap** between candidate and incumbent top-twenty lists - tells you whether the candidate changes anything at all before you spend a quarter finding out whether the change was good. - **Novelty share** - how many surfaced patients the team had already flagged. - **Alert load per nurse-shift** at the deployed list length. - **Per-segment coverage shifts** by ward and admission route. None of these is the launch metric and none may be relabelled as such. They compress the window in which the team is flying blind, and they catch the failures that are obvious quickly - a candidate that reallocates all attention to one ward shows up in days. ## Name the cost of each direction | decision | what it costs | reversibility | |---|---|---| | wait for the slow read | a quarter of the incumbent's misses, plus the team's patience for the programme | fully reversible - the option stays open | | launch on the offline number | an unknown regression that the quarter will eventually reveal | reversible only if you can detect and revert it | Stated that way the real question is visible: it is not which number is more scientific, it is whether the downside of launching early is something you would notice and could undo. A service with a fast revert path and honest interim reads can afford more offline weight than one without. ## What an offline-carried launch commits to - **A re-decision date**, set when the slow read will have accumulated, with someone accountable for making it. - **A named reversal condition** in the online metric and its alert load, agreed before launch, not negotiated afterwards. - **A per-segment floor**, so an aggregate that recovers while one ward stays regressed does not count as vindication. - **The pair recorded**, so this launch becomes one more calibration point for the next one. A launch that rides the offline number without those four is not a judgment call, it is a hope. With them, it is a defensible decision under uncertainty - and over a few cycles it converts the offline metric from an article of faith into a measured instrument whose weight you can state with a number rather than a preference.
- What do you do when the service has no launch history to calibrate against?Treat the first launches as the calibration and pay for it deliberately: give the offline number veto power only, take the slow read at least once end to end, and record both numbers. It is slower than the team wants, and it buys the evidence that lets every later decision be faster. Borrowing another service's proxy record is not a substitute - the gap is specific to this population and this worklist.
- Is a shorter online read the same thing as a faster proxy?No. A shorter read is the same metric with less accumulated evidence; a proxy is a different metric entirely. Shortening the window trades certainty for speed on a quantity you actually care about, while a proxy trades relevance for speed. They fail differently, and a plan should say which of the two it is buying.
- Who should own the reversal condition?The clinical owners of the affected wards, together with whoever runs the service - not the team that built the candidate. The condition is a statement about acceptable harm at a stated alert load, so it needs the authority that can actually stop a live service, and it must be written before the launch rather than argued once a number looks bad.
saying these in an interview costs you the question
- Debates the split in the abstract instead of checking past launches.
- Gives the offline number launch authority because waiting is inconvenient.
- Refuses any launch until the slow read matures, whatever that delay costs.
- Relabels a fast interim read as the launch metric once it looks good.
- Sets a re-decision date without naming what would reverse the launch.