Against
The evidence against the repository's rules and the controversies. Nothing is dropped to make a rule look stronger. Terms are in frame.
Shape: against
The cost of late defects
- Boehm 2001 says a late fix is often 100 times dearer, and gives 5:1 for small noncritical systems. The word "often" was added in the updated list.
- Menzies 2017 analysed 171 projects and found no evidence for the delayed-issue effect (secondary).
- It suggests the effect is not constant across projects.
- Consequence: the repository must not claim that early work is cheaper by a fixed ratio.
Success figures
- Eveleens 2010 applied the Chaos report's definitions to 5,457 forecasts of 1,211 projects and found the Chaos figures did not reflect the data (secondary).
- Consequence: a failure rate from that report is not used here.
Teaching heuristics
- Hembree 1992, in secondary summaries: heuristics had a clear effect only with trained teachers, and fixed steps scored below flexible approaches.
- Consequence: a fixed list of steps is a convention, not a measured aid.
A taught personal process
- Prechelt 2001b found the improvements from PSP training smaller than proponents assume, possibly from low actual use.
- Consequence: a method written down is not a method followed.
Avoidable rework is not all bad
- Fairley 2005: iteration needs some rework, and too little may mean too little review.
- Consequence: the aim is fewer wasted round trips, not none.
Weak measurement of frameworks
- Serrador 2015 is correlational. The guides of Schwaber 2020 and Chapman 2000 cite no measurement in the text read on the second rating. Chapman's introduction says CRISP-DM was validated on a narrow set of projects: a stated limit, not a measure. Singer 2019 is below the threshold and cites none.
- Ubl 2020 is the owner's claim about its own practice and scores below the threshold.
What a skeptic could say of this review
- Sixteen formal works are located, not opened.
- One reviewer scored the grey sources; no second rater.
- The searches were bounded at 10 results per string.