CSE Roadmap

Common tier · about 4 hours · 12 parts

Part 3 of 12 · 2 min read

Plans slip because estimates are medians, and totals and tails are longer than the sum

Adding up a list of "most likely" time estimates systematically underestimates how long the whole project will actually take.

Think of planning a road trip by adding up the "usual" drive time for each leg. Some legs will run long due to traffic. If you hit bad traffic on several legs in the same week — a storm front moving through — the delays don't cancel out. They stack.

In plain words: each feature's time estimate is right-skewed. It can run much longer than expected but not much shorter. On top of that, shared causes of delay — an unfamiliar library, a bad week, shifting requirements — hit every task in the plan at once. That widens the tail of the total far more than adding up independent per-task errors would suggest.

A worked example. A plan with 48 hours of "core only" feature estimates has a median finish time of 52 hours. It needs 71 hours to be 80 percent confident, and 84 hours for 90 percent confidence. The uncertainty band is wider than the median estimate itself. A fuller 116-hour "everything" plan has just a 5 percent chance of finishing within 80 available hours. Removing the shared project-wide delay factor from the simulation — pretending every task's risk is independent — makes the 48-hour plan's chance of finishing in 80 hours jump from 88 to 94 percent. The shared factor alone accounts for a meaningful share of the real-world risk.

The key lines. From code/mini-scope/schedule.py:

# from code/mini-scope/schedule.py
def draw_durations(features, trials=20000, sigma_task=0.5, sigma_shared=0.25, seed=0):
    """Matrix (trials x features) of hours, with the shared factor already applied."""
    rng = np.random.default_rng(seed)
    med = np.array([f.likely_hours for f in features], float)
    task = rng.lognormal(0.0, sigma_task, (trials, len(features)))
    shared = rng.lognormal(0.0, sigma_shared, (trials, 1)) if sigma_shared > 0 else 1.0
    return med * task * shared

Read it as: shared is drawn once per trial and multiplied into every feature's duration in that trial. A single bad draw of the shared factor stretches the entire plan at once. That's structurally different from each task having its own independent bad luck, and it's what makes the tail of the total so much longer than you'd expect from the median estimates alone.

Cut line

The cut line is the point in a priority-ordered feature list below which work gets dropped if time runs short. It is decided from the schedule simulation's confidence level, not from whatever's left undone in the final week.

Example: keeping features in priority order until the 80th-percentile total fits in 80 available hours leaves only 4 features standing. At 90 percent confidence, that drops to 3, cutting the evaluation harness itself. The cut line moves depending on how much schedule risk you're willing to accept. It's a decision made against the simulation, not against panic in week four.