Plant leadership wants one monthly quality number for the 30-day failure model while most of the month's predictions are unresolved - what do you publish, and what do you commit to?

## Three numbers, all defensible, none interchangeable | figure | latency | what it actually claims | how it misleads | |---|---|---|---| | matured-cohort figure | horizon + reconciliation lag | every prediction in this cohort has an adjudicated outcome | describes a model version that may already have been replaced | | censoring-aware provisional estimate | days | given partial observation windows, quality is probably in this band | a wide interval gets quoted as its midpoint | | proxy nowcast | hours | a correlated operational signal is moving this way | a measured bias gets forgotten and the number is read as quality | The error to avoid is not choosing wrongly among them. It is publishing one number whose provenance is unstated, so a reader cannot tell which of the three they are holding. ## The calendar the horizon forces Work it through rather than asserting it. Horizon 30 days, reconciliation lag 7 days, monthly report published on the seventh of the following month: - a prediction made on 1 January closes its window on 31 January and is adjudicable on 7 February; - a prediction made on 31 January is not adjudicable until 9 March; - so on 7 February exactly one day of January is final, and the last month that is complete end to end is December. That is uncomfortable and it is arithmetic. The choice is which discomfort to take: a final number about a period that already feels like history, or a fast number about the present that will move. ## The reporting contract 1. **Stamp every figure** with the cohort dates it covers, the as-of time it was computed, and the share of that cohort already matured. A number without those three travels badly and is misread within a week. 2. **Mark provisional against final** visually and in the data, so a chart shows where the final section ends and the estimated section begins. 3. **Publish the restatement policy before the first restatement.** It states what may change, the maturity threshold at which a figure becomes final, who is notified when a published number moves and by how much, and that prior versions are kept rather than overwritten in place. 4. **Forbid mixed-maturity comparisons.** Comparing this month's provisional figure with last quarter's matured figure produces a gap that is an artefact of maturity; if a comparison is wanted, compare like-aged cohorts. 5. **Say who may act on a provisional number.** Reversible, cheap decisions can run on an estimate with an interval; irreversible ones - retiring a line, standing down a maintenance contract - wait for a matured cohort. ## What the review budget buys Recall does not arrive on its own: without a sampled audit of the assets the model stayed quiet about, the quality number is one-sided no matter how mature it is. Cutting the audit does not make the report cheaper, it makes it partly fictional while leaving it looking identical. If the budget genuinely will not stretch, the honest move is to publish precision alone and say so in the same sentence, rather than to let a one-sided number be read as overall quality. ## The judgment a lead actually owns - **How much lag the business will absorb against how much bias it will tolerate.** This is a business decision with a technical floor, and it differs per decision rather than per report: an operations review can work from a band, a capital decision cannot. - **Which single figure is the headline.** Publishing three numbers with equal weight guarantees that whichever is most flattering gets quoted. Name one as official and demote the others explicitly to supporting evidence. - **What the organisation is told when a number moves.** Restatement is normal here and it is still a credibility event the first time it happens unannounced. The policy exists to make the first restatement boring. - **When to stop paying for an estimate nobody uses.** If the provisional series has never changed a decision in six months, its interval is too wide to be actionable and the budget belongs on shortening the reconciliation lag instead. - **What honesty costs.** Saying "the number for this month does not exist yet" is more defensible than producing one, and it is also the harder sentence to say in a plant review; the contract above exists so it only has to be said once, as policy, rather than argued every month.