Aston University · MSc dissertation JBKS1
Does the agent build
what you asked for?
Five ways of producing the same software, measured against one fixed
specification: a human baseline and four agents. Every figure below
comes from an immutable CSV in data/reports/, and every
card opens the full metric breakdown for that run.
Experiment reports
newest firstHow to read these
What the numbers are, and are not
Each run reports five metrics. Three of them are properties of the code alone and any linter could compute them. The one that matters here is scope drift, which needs a specification and therefore cannot be computed by a tool that does not know what was asked for.
A reading is taken at a commit and dated. Two corrections to published figures are recorded as errata in the repository, and the figures they correct are kept beside the corrected ones rather than replaced.