How this is written
No model writes a sentence here. Here is what does.
Last updated September 11, 2026.
Argusic publishes hundreds of pages about software it has tested, and every measurement on them was produced by automated agents. That is a fair reason to ask how the words themselves were made, so this page answers it plainly and names the files you can check.
No sentence on this site is written by a model
Not the verdicts, not the summaries, not the weekly reports, not the video narration. The system does call AI models, but only in one direction: the agents being measured are AI, and dispatching them is the whole of it. Nothing in the publishing path asks a model for prose.
What produces the words instead is a template filled from a database row. If the row is empty the sentence does not appear, which is why pages here have gaps rather than guesses. The weekly report generator states the same rule in its own first lines: every sentence is a template filled from a row, and a model writes nothing.
Where the words actually come from
- Numbers
- Read from the run that produced them: install time, errors hit and fixed, tokens, cost, the pinned commit, the container image. Each one links to that run.
- The agent's account of what happened
- Quoted exactly as the agent recorded it during the run, including the errors it hit and the minutes each cost. Not summarized, not tidied, not turned into advice. When it is quoted, it is in quotation marks and attributed to the run.
- Strengths, limits, and comparisons
- Assembled from the same rows: a benefit is a measurement, never an adjective. "Installed in eight minutes" is a row. "Easy to set up" is an opinion, and does not appear.
- Explanatory text and page copy
- Written by a person, in the dashboard or in the repository, and reviewed like any other change.
A person approves before anything publishes
Results are created as drafts. A human approves them before they appear, and that approval is recorded. Measurement is automated on purpose, so it is repeatable and nobody can lean on it; publication is not, so there is always somebody answerable for what is on the page. The operator is named on the about page and corrections go to the same address.
Evidence is never edited
Recordings and logs are hashed on arrival and stored unchanged. Three transforms are applied before serving and none of them touches the substance: live credential values are redacted, key material a project generates as its own function is masked, and the agents under test are named under Argusic's own neutral identity rather than their vendors'. If a measurement turns out to be wrong we run it again and the page changes with the new run. A number is never quietly rewritten over an old log.
Why say all this
Because the alternative is asking you to take it on trust, and this site exists to argue that nobody should have to. Everything above is visible in the code: orchestrator/src/lib/report-weekly.ts for the reports, web/src/lib/subject-analysis.ts for what a subject page may say about a project, and orchestrator/src/lib/video-brief.ts for what a video says out loud. Each of those files carries tests whose job is to fail if a sentence ever appears that the data cannot support.
The full measurement procedure, which is a different question, is on the methodology page.