Case: estimates versus actuals

A date check starts from history, not hope. Pick the past projects that look like yours: physics has a curriculum, a review, demo lessons and tutor onboarding, like Klasna's two earlier subject launches (exam math and programming for teens). The simulator, mostly a platform build, stays out.

Then pick the rows that are comparable with your plan. The chain to opening is planned at most likely values, so take the past tasks without padding: those estimates were meant the same way.

For those rows, exam math estimated 25 days and took 33; programming estimated 25 and took 34. Together 67 ÷ 50 = 1.34: most likely estimates in launches like ours ran about a third over. That ratio is what you will apply to the physics chain.

When you turn past launches into one overrun factor, the arithmetic matters. Divide total actual by total estimate, not the average of each task's ratio.

The reason is weight. The chain to opening is measured in days, so a long task that runs over moves the date more than a short one. Totals respect that: a task estimated at 10 days that took 12 adds two days, a task estimated at 3 that took 6 adds three. An average of ratios would count the short task's ×2 as heavily as any long task's ×1.2, and a few small tasks that doubled could swing the factor far more than they ever moved a real date.

For Klasna's unpadded subject-launch rows: 67 actual ÷ 50 estimated = 1.34. That number answers the right question: on average, how many days did a planned day become?

Six of the 17 past tasks were padded: their owners said in the retrospective that the estimate already held a private margin. If padding protected dates, at least some of these tasks should have finished early.

None did. Together they were estimated at 78 days and took 85, about 9 % over even with the margin inside. The margin was used, and then some: the work spreads to fill the time, early finishes aren't handed on.

Six rows are a small sample, so don't claim a precise rate from them. But all six point the same way, and they match the reason padding fails. For the physics chain this settles one argument: the owners' padded total of 54 days isn't a safer plan, only a later one.

The inside view of the chain uses the three-point estimates from Estimating in ranges: for each task, E = (O + 4M + P) ÷ 6, and for a chain, the expected values add up.

Klasna's five tasks to opening: the draft (2, 4, 6) gives 4 days; the outside review (4, 6, 14) gives 7; demo lessons (8, 11, 20) give 12; onboarding (5, 7, 9) gives 7; the dry run (3, 4, 11) gives 5. Total: 35 days.

That is three days more than the most likely total of 32. So in expectation the chain uses 3 of its 11 buffer days before anything special goes wrong, and finishes around 19 January. This is normal, not alarming: most likely values leave out the long bad side, and the expected value puts it back. The next question is how far the chain can stretch.

The spread of the chain follows the rule from Estimating in ranges: each task's standard deviation is (P − O) ÷ 6, and you add the variances, then take the root.

For Klasna's chain the standard deviations are 0.67, 1.67, 2, 0.67 and 1.33 days. Their variances add to 0.44 + 2.78 + 4 + 0.44 + 1.78 = 9.44, and √9.44 ≈ 3.07 days. Adding the deviations instead (6.33) would double the spread.

Two standard deviations above the expected 35 gives about 41 days, 27 January, still inside the 43 available. The widest task is the demo lessons (8 to 20 days), which is where Danylo expects trouble and where to look first.

Remember the condition behind the rule: it assumes the tasks are independent. A shared cause that could delay several of them belongs on the risk list, not inside this number.

The outside view applies history to your own plan. Take the plan at the same kind of estimate the history measured, here the most likely values, and multiply by the past overrun factor.

For Klasna: the chain's most likely values add up to 32 working days, and the subject launches ran 1.34 times their most likely estimates. 32 × 1.34 = 42.88 days: nearly all of the 43 available. The same move works per task when the history has it: reviews ran 1.69 times their estimates, so Vira's likely 6 days becomes about 10.

Compare with the inside view: 35 expected, about 41 at the high end. The two views don't have to agree. When the outside view is tighter, it is usually telling you about surprises the plan hasn't imagined yet. Here it says the date is achievable, but there is no room for anything new in the chain, and the review is the task to watch first.

The inside view says the chain is comfortable: 35 days expected, about 41 at two standard deviations, 43 available. The outside view is tighter: the most likely 32 days times the subject launches' 1.34 gives about 42.9 days, and Vira's review, at the 1.69 that reviews ran, becomes about 10 days instead of the expected 7.

Neither view is "right". The honest answer holds both: the date is achievable but tight. Say it with its confidence and its conditions: it holds if nothing new is added to the chain and the review is done by mid-December.

The tempting answers each drop half the evidence. "Guaranteed" ignores the history. "Impossible, 54 days" treats padding as a plan. "Move to 15 February to be safe" gives up two weeks of the exam season for a risk that can still be managed.

A date with moderate confidence is a promise with conditions, so the message to the sponsor names the signals that would change it, each checkable by a date.

For Klasna's 1 February: the outside review, the task history says overruns most, should be done by mid-December; the shared buffer, held by the lead, shouldn't be used faster than the chain is being completed; and nothing new may join a chain that has no room left. If any of these trips, Leo hears it that week, while there are still options: cut scope, add help where it really shortens the chain, or decide on the date with eyes open.

Hours logged and how confident the owners feel aren't such signals: both can look fine while the date slips.

How it fits together: the outside view, the three-point chain and one owned buffer turn "can we promise it?" into a date with confidence and conditions. The next module turns to the people around the plan: Leo's shifting priorities and the stakeholder nobody has talked to yet.