Why enterprise AI has no reference works

Ask a structural engineer what they consult and they will name a code and a handbook, both revised on a cycle, both carrying the marks of every argument the profession has had. Ask an accountant, a pharmacologist, a pilot. Mature practices produce maintained reference works, and the maintenance is most of the value.

Ask somebody running an enterprise AI programme what they consult and you will get a list of newsletters, a conference they went to, two analyst subscriptions their employer pays for, and a group chat. All useful. None of them is a reference work, and the difference is not prestige. It is that a reference work is accountable to its own past claims and a feed is not.

The feed cannot help being wrong

A post is finished when it is published. That is the whole economic model, and it produces three behaviours that make the format unsuitable for reference.

It rewards novelty over accuracy, because attention accrues to the new claim and not to the correction. It has no mechanism for retraction that reaches the people who read the original. And it dates everything to the moment of writing, which means a reader six months later has no way to tell whether the author still believes it.

None of this is a failure of character. Writers correct things constantly. The format simply gives them nowhere to put the correction where it will be found by the person who needs it.

What a discipline needs instead

The properties that make a handbook trustworthy are unglamorous and largely mechanical.

• Numbered, stable addresses. You can cite Entry 004 in a meeting and everyone can find the same thing.
• Revision that is real rather than cosmetic. Not a last-updated date that ticks when somebody fixes a typo. A record of substantive change, kept because the editor needs it, and surfaced to the reader as the one fact they can act on: when this last actually changed.
• Corrections marked as corrections. Folding an admitted error into a general update is a way of not admitting it.
• Maintenance that happens whether or not anyone can see it. This is the hard one, and it is hard for a reason that has nothing to do with tooling.

The part everyone gets wrong

The obvious move, once you accept all that, is to publish the machinery. Show the revision count. Put a badge on anything that has gone sixty days without a read-back. Publish the queue of unmaintained pages. It feels like the honest thing, and it is the first thing I built.

It is wrong, and it took seeing it rendered to understand why.

A staleness badge does not tell the reader something useful. It hands them a decision that was mine to make. “This has gone sixty days without a read-back, you decide whether to trust it” is not transparency. It is an editorial obligation with the label turned outward, and the reader has no way to discharge it: they cannot know what would have to be true for the entry to still hold.

There are only two honest states for a page in a reference work. Either it currently holds, in which case it should be there and the badge is noise. Or it does not, in which case it should be corrected or taken down, and a badge is a way of leaving it up while disclaiming it.

The same goes for a public changelog. A reference work is not a repository, and a reader who wants to know whether their understanding from last year is safe is not asking for a diff. They are asking one question, and the answer is a date: when did this last substantively change.

What honest maintenance actually looks like

Do the work, and let the reader see the one output that concerns them.

Behind this site an automated job re-reads the citations every week, refreshes the measured data, and tells me which entries have gone too long without somebody checking them against current practice. That queue exists. It is detailed, and it is genuinely useful, and it reaches me by email. It is deliberately not published, because it is a status report about my own housekeeping and the reader did not come here for that.

What the reader gets is the entry, the date of its last substantive change, and a correction marked as a correction when I get something wrong.

The discipline is not less real for being invisible. It is more expensive, because nothing about doing it earns credit with anybody. That is the actual reason the field has no reference works, and it was the reason before anybody thought about badges.

Why nobody has built one

Partly because the field is young, and a canon requires enough settled practice to be worth writing down. That objection is weaker than it sounds. Plenty here is settled: the funding pattern behind stalled pilots has not changed in three years, and the measurement failures repeat so exactly across organisations that you can predict the conversation.

The larger reason is that maintenance is unrewarded, and it stays unrewarded even if you advertise it. Publishing a new piece produces a spike. Re-reading an old one and confirming it still holds produces nothing, and dressing that up as a badge does not change the economics, it just moves the cost onto the reader's judgement.

So it has to be worth doing for its own sake, or it does not get done. That is an uncomfortable place to end and I think it is the true one.