CSE Roadmap

Common tier · about 4 hours · 12 parts

Part 6 of 12 · 4 min read

What the example found

Schedule risk by plan size, the shared-factor effect, cut lines by confidence, the last-mile curve, evaluation item counts, and the three fictional scope checks.

How to read the schedule-risk table: each row is a different feature plan (core only, core plus important, everything); the right two columns compare the chance of finishing in 80 hours with and without the shared delay factor.

Plan (median of estimates) Median finish 80th percentile 90th percentile P(done in 80 h) Without shared factor
Core only (48 h) 52 h 71 h 84 h 88% 94%
Core and important (70 h) 78 h 102 h 118 h 54% 56%
Everything (116 h) 130 h 167 h 191 h 5% 0.1%

What to notice: the "everything" plan's chance of finishing swings from 5 percent down to 0.1 percent when the shared factor is removed. For a large enough plan, shared delay risk is what keeps a bad outcome merely unlikely rather than almost impossible. For the small "core only" plan, removing it only shifts the odds from 88 to 94 percent, a much smaller swing.

How to read the spread-sensitivity table: each row is a different assumed per-task duration spread; each column repeats the same three plans from the table above.

Spread of task durations (assumed) Core only Core and important Everything
0.3 95% 63% 6%
0.5 88% 54% 5%
0.8 72% 39% 4%

What to notice: the "everything" column barely moves across spread assumptions (6 to 4 percent), while "core only" moves substantially (95 to 72 percent). A plan that's already comfortably oversized relative to the hours available is insensitive to exactly how uncertain each task's duration is. A tight plan is not.

How to read the cut-line table: each row is a different confidence level; the cell shows how many features survive the cut at that confidence.

Confidence for the cut line Features kept
50% 7 (core and important)
80% 4 (core only)
90% 3 (the evaluation harness is cut)

What to notice: pushing confidence from 80 to 90 percent doesn't just trim a stretch feature. It cuts into the evaluation harness itself. That's exactly the kind of cut the module's design rules warn against making from panic rather than from a confidence level chosen deliberately in advance.

How to read the last-mile table: each row is a target quality level; months are measured from the start of the project.

Target quality Months from the start (anchor: 80% at 1, 95% at 5)
80% 1.0
90% 3.0
95% 5.0
98% 7.6
99% 9.6

What to notice: the first 80 percent takes 1 month, but the next 19 points (80 to 99) take nearly 9 more. The quality curve's cost is almost entirely loaded onto its final stretch, exactly as the halving-time formula in Part 3 predicts.

How to read the evaluation-size table: each row is a different planned accuracy gain; the number is paired items needed for 80 percent power.

Planned gain Paired items for 80% power
50% to 60% 155
60% to 75% 77
50% to 70% 57
45% to 70% 45
50% to 90% 27

What to notice: the smallest planned gain in the table (50 to 60 percent, just 10 points) needs by far the most items (155). The relationship between gain size and items needed is not linear; it's roughly inverse-square. So halving your ambition more than doubles your evaluation burden.

How to read the interval-width table: each row is a different evaluation set size; the width is for a measured score near 0.70.

Evaluation items Width of a 95% interval near 0.70
50 25.4 points
100 18.0
150 14.7
300 10.4
500 8.0

What to notice: going from 50 to 100 items (doubling) narrows the interval from 25.4 to 18.0 points. But going from 300 to 500 (less than doubling) only narrows it from 10.4 to 8.0. The return on more items shrinks fast, which is why a sensible floor exists rather than an open-ended "more is always better."

How to read the scope-checker table: each row is one of three fictional scopes, ranging from deliberately vague to deliberately tight.

Fictional scope Flags What was flagged
Vague 5 user, multi-part job, no measurable result, no simple baseline, no cost or latency limit
Partial 4 evaluation set (60 items), no simple baseline, no cost or latency limit, plan exceeds hours at the median
Tight 0 none

What to notice: the "partial" scope actually has a specified number and target. Its flags are all about the supporting evidence — eval set size, baseline, limits, schedule — being too thin to back that number up. That is a more subtle failure than the "vague" scope's complete absence of specifics.

{{svg:scope|A one-page scope and what the checker says about three fictional scopes.}} {{svg:time|Schedule risk, the last mile and the size of the evaluation set.}}