Dark factories: generation is real; Level 5 trust is not

Peter Roelants’ Dark Software Factories (v3.66, 23 Sep 2026) is a long, dated evidence dossier — not a product pitch — on what public systems have actually shown. Working definition: people specify intent, constraints, and acceptable behavior; agents produce and validate implementation. In the strict form, people neither write nor routinely review the generated code.

The report’s load-bearing conclusion is uncomfortable for pure overnight optimism: strict darkness remains unproven. Named cases (OpenAI, Intercom, Honeycomb, Tessl, Ramp, Spotify, LaunchDarkly, and others) support bounded agent pipelines — maintenance, research-intern tasks, high merge share under human direction, risk-routed specialist review — not a general-purpose factory owning live software unattended across years of incidents and migrations. Generation is the strongest claim; repeatable operation has young support; the trust claim that automated evidence can replace routine source reading is the least proven. LaunchDarkly’s stalled legacy rewrite until humans restored steering is kept as a counter-example, not a footnote.

Instead of treating the prompt as the new source, Roelants argues for a versioned source bundle (intent, contracts, acceptance scenarios, architecture and safety policy, factory program, provenance) with explicit authority classes so the implementer cannot silently weaken protected acceptance. Pull requests stay useful operational envelopes; the semantic compilation unit is the change package. Public stacks show recurring layers — executable checks, capability and release controls, operational feedback, sometimes review agents — whose coverage and lifetime effectiveness remain open. Maintainability and lifetime economics are called out as missing from throughput-first first-party reports.

Why it matters for code review

Harnessed overnight momentum still beats babysat Claude bursts inside the envelope humans actually trust — and this dossier marks that envelope. Continuous verification under volume is necessary and insufficient: without governed specs, independent acceptance, and an honest halt when evidence is weak, “always-on” is just faster amplification of unverified claims. Use the report as the complication: ship the continuous harnessed loop, keep humans on consequential semantic choices, and do not confuse a green night shift with Level 5 ownership.