Rough Consensus Was Borrowed From a Culture With Nothing at Stake

The phrase comes from a slide. David Clark, at the twenty fourth meeting of the Internet Engineering Task Force in July 1992, in a presentation the proceedings record as “A Cloudy Crystal Ball” with the subtitle “Visions of the Future”:

We reject: kings, presidents and voting.

We believe in: rough consensus and running code.

Crypto took the second line and used it as a governance model for money. The first line is the one that turned out to matter, and the reason is that “we reject voting” is only half a sentence. If you are not voting, something else has to decide, and in the culture that produced this slogan the something else is a named human being with the authority to say the argument is over.

That person does not exist in Bitcoin. The mechanism was borrowed without them.

What the norm actually specifies

The IETF wrote its own norm down properly in June 2014, twenty two years after the slide, in RFC 7282, “On Consensus and Humming in the IETF” by Pete Resnick. It is worth noticing that they had to. The originating culture found its own practice had drifted far enough that participants were asking, in the document’s words, “Why don’t we just vote?”

The RFC’s section headings are the specification, and they are unusually blunt:

• Lack of disagreement is more important than agreement.
• Rough consensus is achieved when all issues are addressed, but not necessarily accommodated.
• Humming should be the start of a conversation, not the end.
• One hundred people for and five people against might not be rough consensus.
• Five people for and one hundred people against might still be rough consensus.

The last two are the whole design. Consensus here is not a count and cannot be turned into one. What determines the outcome is whether an objection has been genuinely engaged with, and the RFC is explicit that this is the point:

It’s the existence of the unaddressed open issue, not the number of people, which is determinative in judging consensus.

Head counting is rejected on purpose, and the document names what it is rejecting it for. The word it uses is “vote stuffing”. The worked example in section 7 is a participant who loses an argument on the merits, then recruits a hundred colleagues from their own company to post the same one line message, and the chair correctly disregards all hundred because none of them can answer a question about the technical point.

The load-bearing part

Read that example again and notice who does the work. Not the process, not the document, not a rule. The chair.

If the chair of a working group determines that a technical issue brought forward by an objector has been truly considered by the working group, and the working group has made an informed decision that the objection has been answered or is not enough of a technical problem to prevent moving forward, the chair can declare that there is rough consensus to go forward, the objection notwithstanding.

Determines. Declares. This is a person exercising technical judgement and then making a call that binds the group, and the RFC is honest that the call is hard: it asks, rhetorically, how many of us are willing to say that the vast majority of people in the room are stonewalling.

Rough consensus is therefore not an alternative to authority. It is a specific allocation of authority, to somebody with enough standing to overrule a majority, held in check by the fact that an appeal exists and that everyone involved has a professional reputation to lose. Remove the chair and you have not got a purer version of the norm. You have got a slogan.

What crypto kept, and what it did not

The improvement proposal process borrowed the structure of the standards culture almost verbatim, including the numbered documents, the status fields and the mailing list discussion. What it did not borrow was the office.

BIP 2, the process document, is direct about this. It assigns consensus building to the person proposing the change:

The BIP author is responsible for building consensus within the community and documenting dissenting opinions.

And it defines the only role that resembles a chair, in a sentence that reads like it was written specifically to close the door:

The BIP editors are intended to fulfill administrative and editorial responsibilities.

Administrative and editorial. Not judicial. An editor assigns a number, checks the formatting, verifies the thing was posted to the list, and merges the pull request. Nobody in the process is empowered to look at an objection and rule that it has been addressed.

That was a choice, and it was defensible on its own terms. A network holding value has a much worse worst case for capture than a protocol working group, and a chair with the authority to declare an argument over is a chair somebody would have a very strong incentive to become. Declining to create the office is a real defence against a real threat. It also removes the component the norm was resting on.

Why the norm breaks specifically here

The failure is not that people behave worse when money is involved. It is narrower and more interesting than that, and it survives the assumption that everybody is entirely sincere.

The RFC’s test is whether an objection has been addressed. Applying that test requires an observer to evaluate an argument on its technical merits. In a standards working group that evaluation is difficult but tractable, because a participant’s stake in the outcome is mostly professional and mostly legible: they work for a vendor, everybody knows which vendor, and the argument is about a protocol nobody’s savings are denominated in.

In a network where the participants hold the asset, the same evaluation asks an observer to separate two things that produce identical text. A technical objection and a position in the outcome generate the same posts. They generate the same posts even when the objector holds no position at all, because the audience cannot see that either. And the RFC’s own worked example depends on being able to ask an objector a follow up question and judge whether the answer is informed, which is precisely the judgement that becomes unavailable when every participant can be suspected, plausibly and unfalsifiably, of arguing their book.

I want to be careful here, because this is the point at which most writing about protocol disputes goes wrong. The claim is not that anybody in any particular argument was motivated by their position. That is not knowable from the outside and asserting it is the error, not the analysis. The claim is structural: once the suspicion is available to everyone about everyone, the test the norm depends on stops returning a usable answer, and it stops returning one regardless of whether anybody is actually behaving badly.

That is why these disputes do not resolve. Not because one side is dishonest, but because the mechanism for declaring them resolved was left out, and the substitute test cannot be applied.

What replaced it, which is nothing

No better norm has arrived. It is worth saying that plainly rather than gesturing at one.

The obvious substitute is counting something, and every version of counting has a worse failure than the one it replaces. Counting people invites the vote stuffing RFC 7282 was written to defeat. Counting holdings makes the objection the norm could not evaluate into the thing that decides, which does not solve the problem, it formalises it, and later eras on this site show at length what that produces. Counting hardware measures capital deployment rather than agreement.

What is actually left is the arrangement described elsewhere on this site, where merging a change is not the same as anyone running it. Users choosing which software to run is the residual decision procedure, and it is slow, noisy, hard to read while it is happening, and impossible to appeal. It is also, and this is the honest concession, remarkably difficult to capture, because there is no office to take.

Both of those are the same property. A system that cannot be captured quickly cannot decide quickly either.

What was expanding, what was contracting

What expanded was the number of people entitled to participate in the argument. The standards culture the norm came from was open in principle and small in practice: attendance cost money and time, and the participants mostly knew each other. A public network with an open repository and a public list has no such filter, and by the middle of the following decade the number of people who could credibly claim standing in a protocol dispute had grown by orders of magnitude.

What contracted, over exactly the same period, was the capacity to conclude one. Every mechanism the original norm used to end an argument depended on a small group, a shared professional culture, and a person willing to make an unpopular call. All three thinned as participation grew.

Who could tell at the time? The people who had read the standards documents closely enough to notice that the office was missing, which was a very small number, and who mostly raised it as a procedural quibble rather than as a prediction. The rest of the field took a slide from 1992 as a description of how the thing already worked, and did not check what the second half of the sentence had been resting on.