Against: controversies, null results, where practice contradicts the evidence

This file collects the evidence against the common rules of architecture design. A source under 10 of 20 is an example only.

Against design patterns as a cure

Zhang 2012: firm support for the claims made for patterns in general could not be identified, in 11 formal experiments.
Zhang 2012: some indication that patterns do not help novices learn design.
Prechelt 2001 (abstract read): in most of nine tasks the pattern design was no worse, so the evidence is mixed rather than against; a few tasks favoured the simpler design.
Khomh 2008 (technical report abstract, secondary): several quality attributes were affected negatively by patterns.
Vokac 2004 (secondary): defect rates of pattern classes ranged from 63% to 154% of the average, so no pattern is uniformly safer.
Prechelt 2002 supports pattern documentation only for pattern-relevant tasks, in small programs of 360 to 560 lines (secondary).

SESE 2020d: low coupling is a trade-off; indirection that lowers coupling can break a real-time budget.
SESE 2012k: the tool metrics give counts, not a verdict on a design.
No study found tests information hiding as a criterion on a real system; its support is argument and one worked example on an indexing problem (Parnas 1972, abstract read in a summary).
The measures of dependency structure are observational: Mo 2016 relates decoupling level to proxies from revision history, not to effort (secondary).

Blinowski 2022: on a single machine the monolith performed better; vertical scaling was more cost-effective in one cloud.
Mendonca 2021 and Noonan 2018: two companies moved from services to a monolith, each reporting on its own decision.
SESE 2017h and SESE 2018d: microservices turn software into a distributed system; benefits come from decoupled components, not from the deployment.
Fowler 2015: monolith first, with the evidence admitted as sparse.
Taibi 2017 (secondary): splitting the monolith and its data stayed hard for those who migrated.
For the other side: 92% of the 1,502 respondents report some success (Loukides 2020); the sample is self-selected from mailing lists.

Mauerer 2022: across 25 open source projects, socio-technical motif congruence was not related to bugs or churn in any temporal scenario.
Cataldo 2013 (secondary) found congruence tied to fewer failures in two industrial projects, so the effect is not robust across settings.
Colfer 2016: mirroring is not supported in 8% of the descriptive industry and firm studies and in 56% of those on open collaborative projects; partial mirroring appears in technologically dynamic industries.
The inverse Conway maneuver reached Trial in 2014 and left the current Radar (ThoughtWorks 2014).
Team Topologies 2026 (example only) gives anecdotes and no study count.

No controlled study shows ATAM or any evaluation improves the system built.
Boehm 2008 (abstract read): the cost gap between minimal and very thorough architecture work is about 18% for small projects and 92% for very large ones. The case for thorough up-front work is therefore weaker for small projects (the paper's own guidance was not read).
Waterman 2015: too little up-front design raises risk and too much delays value; teams use forces and strategies, not a rule.

Vendor guides recommend styles by domain type with no study (Microsoft 2025).
The quality model has nine characteristics, while the cloud framework has six pillars; the two lists are not mapped to each other in the pages read (ISO 2023c; AWS 2026d).
Practice moves on without the evidence: the two Radar entries read are no longer on the current edition.

Eleven formal works were not reached beyond a summary line or the title; no counter-evidence is claimed from them.
The strongest null result is a mapping (Zhang 2012), not a replication.