In short: Almost everything about an enterprise brain changes with size: the tools, the owners, the governance, the automation, whether it can learn. Five things do not. Get those right at two people and they carry you to ten thousand. Get them wrong at ten thousand and no platform will save you.
Key points
- Record every decision with what was decided, why, and what would prove it wrong by when.
- Keep the evidence behind a decision, with timestamp and origin; summaries are views.
- When the organisation changes its mind, link the old view to the new one with a reason and date.
- Without someone checking the falsifier and recording what happened, nothing learns at any size.
- The record must be exportable with history and survive replacement of any tool, vendor or person.
These last ten pieces have described a ladder: a founders' document, a ledger strangers can read, a glossary, decision classes, a governed trace, a federated learning layer. Each rung is different. Underneath them is a short list of properties that are the same at every size. They are what make the record a record of judgment rather than a pile of content.
The five invariants
1. Decisions are recorded with a reason and a falsifier. What was decided, why, and what would prove it wrong by when. At two people it is a line in a document. At ten thousand it is a required field in every workflow. Same idea.
2. Sources are kept, and summaries are views. The evidence behind a decision survives the summary of it, with a timestamp and an origin.
3. Views are superseded, never overwritten. When the organisation changes its mind, the old view remains, linked to the new one with a reason and a date.
4. Outcomes are written back. Someone checks the falsifier on its date and records what actually happened. Without this, nothing learns, at any size.
5. The organisation owns the record. It can be exported, with history, and survives the replacement of any tool, vendor or person, including the people who built it.
What scales on top
Volume, which makes learning possible. Automation, which makes capture cheap. Governance, which makes the record defensible. Federation, which lets many records coexist. Each matters more as the organisation grows. None of them substitutes for the invariants, and all of them amplify whatever the invariants are, good or bad.
One question per rung
If you remember nothing else from these pieces, ask the question for your size:
- Two: would we both describe last month's big decision the same way?
- Ten: can a new hire find out why we do this without asking a founder?
- Fifty: do our teams mean the same thing by our ten most important words?
- Two hundred: does a new manager see the three closest prior decisions before making one?
- One thousand: could we reconstruct an AI-assisted decision from three months ago, end to end?
- Ten thousand: do we know, by decision class, whether our judgment is getting better?
Back to the thesis
This series began with a claim: the enterprise brain that compounds is not the one that remembers most, but the one that can show what it decided, why, what would have proved it wrong, and when it changed its mind. The five invariants are that claim, made operational. They cost almost nothing at two people. They are the hardest thing to retrofit at ten thousand. That asymmetry is the best argument there is for starting small, and starting now.
In practice
- Score your organisation against the five invariants today: present, partial or absent.
- Fix the first absent one before buying or building anything else.
- Ask your rung's question at the next leadership meeting, and write down the answer.
What would prove this wrong. Organisations that lack one or more invariants at small scale and add them later at large scale, at low cost and without losing the history they skipped.
Common questions
What are the core principles of a company decision record?
Decisions are recorded with a reason and a falsifier, sources are kept, views are superseded rather than overwritten, outcomes are written back, and the organisation owns the record.
When should a company start keeping a decision log?
Now, and small. The five invariants cost almost nothing at two people and are the hardest thing to retrofit at ten thousand.
Can automation or governance make up for a weak decision record?
No. Volume, automation, governance and federation matter more as an organisation grows, but none substitutes for the invariants and all amplify whatever the invariants are, good or bad.
How can a ten-person company test its decision record?
Ask whether a new hire can find out why the company does something without asking a founder.