ADR-004 — Organisational knowledge over internet knowledge
Context
Language models already know what the internet knows. They know very little about how a specific company works: who the experts are, which standards matter, what has already been built, why previous decisions were made. Systems that lean on the model's general knowledge produce generic answers, and a competitor can generate those too.
A second failure mode is subtler: treating all internal knowledge as equal. Real organisations are not averages. Some engineers consistently design better systems; some specialists consistently catch what others miss. Averaging across all available information dilutes exactly the judgement that makes the organisation competitive.
Decision
Ground reasoning in organisational knowledge first; use general knowledge only to interpret it. Within organisational knowledge, weight the practices of the organisation's best people (craft excellence) above the mean. The goal is to distribute excellence, not to average it away.
Consequences
- Positive: answers reflect how this organisation actually works and preserve its competitive judgement. Expertise becomes discoverable and reusable instead of siloed in individuals.
- Negative: identifying and encoding craft excellence requires human curation and organisational honesty about whose judgement sets the standard. This is slow and occasionally political. It is also the point.