Steward review capacity and the arithmetic of how many records fit
Steward review capacity is the number of candidate records one person can decide in the time available, calculated from measured seconds per decision rather than assumed. A steward averaging 99 seconds a pair clears 36 an hour and 109 in three focused hours, so a queue larger than that needs a threshold change or a second reviewer.
Three weeks before doors, somebody asks the data lead how the duplicate cleanup is going. The queue holds 4,000 candidate pairs. One person works it between two other jobs, and nobody has ever measured how fast that person goes. Steward review capacity is a number you can work out in an afternoon, and until somebody works it out, the plan for that queue is a wish with a deadline attached.
The arithmetic is short. Seconds per decision, decisions per hour, hours per day that survive contact with a show cycle. Multiply, compare against the queue, and the conversation about thresholds becomes a conversation with numbers in it.
What does one steward decision actually involve?
A candidate pair arrives as two registration rows that a scoring stage thinks might be the same person. The steward opens both, reads name, company, email domain, phone and event history, and decides to merge, to leave them separate, or to escalate because the evidence points two ways.
The easy pairs are quick. Same person, one character different in the surname, same work email, same company: twenty seconds. The hard ones eat the clock. Two buyers at the same twelve-person distributor, one shared inbox, similar first names, both registered by an office manager who filled the company field differently each time. That pair needs somebody to look at the badge history, and sometimes to send an email.
So the mean matters more than the impression. Time twenty real pairs with a stopwatch. A plausible result: twelve pairs at 45 seconds, six at two minutes, two at six minutes. That is 540 plus 720 plus 720, which is 1,980 seconds for twenty decisions, a mean of 99 seconds. The median is 45. Anyone planning capacity off the feeling of a typical pair will overstate throughput by more than double, because the tail is where the time goes.
DAMA International's Data Management Body of Knowledge, second edition, 2017, describes stewardship as accountability and responsibility for data and for the processes that ensure effective control and use of data assets. Accountability has a name on it and responsibility has hours behind it, which is the part that gets skipped when a stewardship programme is announced without anybody's calendar changing.
The capacity number, worked from your own file
Take the 99 second mean. An hour of uninterrupted review clears 3,600 divided by 99, which is 36 decisions. Three focused hours a day gives 109. Five days gives 545 a week.
Three hours rather than eight is the assumption worth arguing about. A steward who is also the person answering exhibitor questions does not get eight clean hours, and decision quality drops long before the eighth hour anyway. Pick the number your own team can defend, then hold it constant so the capacity figure means something across editions.
Now the comparison. Four weeks before the badge file freezes, the queue holds 4,000 pairs and is still growing. Capacity over those four weeks is 2,180. The gap is 1,820 pairs that will not be looked at by anybody, and that gap exists whether or not it appears on a slide.
Two more figures make the position concrete. At 545 a week, clearing 4,000 pairs takes 7.3 weeks, so the queue as it stands is a seven week job with four weeks left. And each extra hour a day of review adds 182 pairs a week, which means a full extra hour every day for four weeks buys you 728 decisions, well under half the shortfall.
One correction to apply before trusting any of this. A share of the queue gets escalated instead of resolved, usually a pair that needs the registration platform checked or a call to the exhibitor whose staff list is a mess. Those pairs consume the review minute and then come back later, so they get counted twice in throughput and once in progress. If one pair in eight is escalated, effective clearance is 545 times 0.875, which is 477 a week, and the four week figure drops from 2,180 to 1,908. Track the escalation share for a fortnight and put it in the capacity formula, because it moves the answer by more than most threshold tweaks do.
What happens to the pairs nobody reviews
They get decided anyway. Whatever your pipeline does with an unresolved candidate is the decision, and in most builds that default is to leave both records separate, so the duplicates stay and the audience count carries them.
That default deserves to be written down and reported, because an unreviewed queue is a quiet policy. If 1,820 pairs go unlooked at, the honest line in the post-show report is that 1,820 candidate duplicates were not adjudicated, alongside however many were merged. Reporting the merges without the backlog makes the cleanup look complete.
Fellegi and Sunter set up the three-way decision in the Journal of the American Statistical Association in 1969: link, non-link, and a middle region held for clerical review. Their optimality result is worth reading with a staffing eye, because the rule they prove optimal is the one that minimises the size of that middle region for fixed bounds on the two error rates. The theory treats human review as the scarce resource, which is exactly how a show team experiences it.
Should you move the threshold or add a reviewer?
Both are legitimate. The mistake is choosing without pricing the option you reject.
Say the review band currently runs from 75 up to the auto-merge line, and 2,300 of the 4,000 pairs sit in the lowest five points of that band. Pull 200 of those 2,300 at random, have the steward label them, and suppose nine turn out to be true matches. That is 4.5 per cent. Raising the review floor by five points removes 2,300 pairs from the queue and abandons roughly 2,300 times 0.045, which is 104 true matches. The remaining queue is 1,700 against a capacity of 2,180, so it fits with room for the pairs that arrive next week.
The comparison is then 104 known unmerged duplicates against 1,820 unreviewed pairs whose composition nobody knows. I would take the 104 every time. It is bounded, it is measured, it can be quoted in the report, and next year you can lower the floor again once you have the hours. The unreviewed queue has none of those properties, and it grows.
The sampling step above needs an interval around it before anyone acts on 4.5 per cent, and pulling a defensible sample from a file you cannot review entirely is K32's subject rather than this one.
Sizing capacity to the show calendar
Registration arrives in bursts, so candidate pairs do too. If 70 per cent of the year's candidates are generated in the eight weeks before doors, capacity has to be planned against those eight weeks and not against an annual average.
Work it through. Nine thousand candidates a year, 70 per cent of them in the peak, gives 6,300 pairs in eight weeks. One steward at 545 a week covers 4,360. A second reviewer at half time adds 272 a week, or 2,180 across the eight weeks, which brings capacity to 6,540 and closes the gap.
That second reviewer does not have to be a hire. The two people who know the exhibitor base best are usually in sales operations, and a four hour rota on Tuesday and Thursday mornings during the peak is easier to get signed off than a headcount request. Whoever it is needs the same decision rules written down, otherwise you have doubled throughput and halved consistency.
The other lever is moving the work earlier. Running the matcher against last edition's file in February generates candidates in a month when somebody can actually clear them, and a person resolved in February stays resolved when they register again in September.
Where this arithmetic stops
Capacity maths assumes decision quality holds steady, and it does not. Sustained review work degrades after the first hour or so, and the errors it produces are invisible because a wrong merge looks like a completed task. The cheap control is to slip the same thirty pairs back into the queue a fortnight later and measure how many decisions come out the same way. If agreement with the earlier pass is under 90 per cent, the throughput number is optimistic and the training material needs work before the rota does.
The second limit is that capacity is the wrong lever when the queue is full of junk. A generator producing thousands of low-evidence pairs will exhaust any staffing plan, and the fix sits upstream in blocking and scoring. Compute the share of reviewed pairs that end in a merge. If it is under one in ten, you are paying people to say no, and the threshold or the blocking key is what needs attention.
The third limit is the show week itself. A steward who is on the floor from Tuesday has no capacity that week, and any plan that quietly assumes otherwise fails in the week the data matters most. Build the rota with show week at zero and see whether the arithmetic still works. This is also the point where the numbers stop being a data question and become a resourcing decision that belongs to whoever holds the domain, which is the split between deciding and doing that K34 covers, and it feeds everything downstream in the unified data layer.
Start this week by timing twenty pairs. Note the seconds for each, take the mean rather than the median, multiply by the hours you actually have, and put the resulting weekly number next to the current queue depth in one line of your Monday report. If the two numbers are far apart, you now have the argument you need about the threshold, in a form nobody has to take on faith.
Questions people ask about steward review capacity
- How many records can one data steward review in a day?
- Measure it before you assume it. Time twenty real decisions with a stopwatch, take the mean rather than the median, and multiply by the hours the person genuinely has. A mean of 99 seconds gives 36 decisions an hour, so three protected hours a day clears about 109 records.
- What should happen to the candidate pairs nobody has time to review?
- They fall to whatever your pipeline does by default, which is usually leaving both records separate. Write that default down and report the count every week. An unreviewed queue of 1,800 pairs is a decision to keep an unknown number of duplicates, and it should appear in the same report as the merges.
- Is it better to move the match threshold or add another reviewer?
- Sample the band you would drop before deciding. Label 200 pairs from the range you plan to exclude, count the true matches, and multiply that rate by the band size to get the merges you are giving up. Compare that known cost against the unbounded cost of a queue nobody finishes.