Two doors: conversation and ledger

Research does not arrive in well-formed records. It starts as an argument: a hunch, a table someone recomputes three ways, an objection that turns out to be a misreading, and an hour of going back and forth before anyone can say what the question is. By the time there is a hypothesis precise enough to be worth registering, most of the intellectual work has happened somewhere that has no schema. Structure describes the end of that process, and the process spends most of its time at the beginning.

The two ways to get it wrong#

A system that demands structure at the first keystroke gets one of two things, usually both.

It gets nothing, because the work moves elsewhere. If the only way in is to declare a hypothesis, a plan and a predicted outcome, everyone thinks in a chat window and arrives afterwards to file the conclusion. What the system holds is then a tidy summary of research that happened somewhere it could not see, which is the situation the record was supposed to improve on.

Or it gets theatre. The forms are filled in, in the order the system expects, after the fact: the hypothesis written once the result is known, the plan reconstructed from what was run, the predicted outcome the observed one. Every field is complete and the whole is a reconstruction. That is worse than the first failure, because it looks like compliance while asserting an order of events that did not happen.

The opposite mistake is equally available. A record loose enough to accept anything — paste in the transcript, call it a result — supports nothing. If a conversation counts as a finding, nothing in the store has a scope, an author answerable for it, or a boundary between what was shown and what was guessed.

Two doors, one place#

The way out is not a compromise between the two. A record that is moderately strict is strict enough to annoy and loose enough to be untrustworthy. The way out is two entrances into the same place: a conversational door, wide enough for a half-formed idea, and a strict door for anything that is to be reused, with an explicit act of crossing between them.

Through the wide door go Threads: public discussion inside a Room, where people and their agents argue, propose, object and abandon things at the cost of a sentence. Nothing there is a claim about the world. It is the working surface, and it has the property that matters most at that stage: being wrong is cheap.

Through the strict door go publications: a claim with its scope, its evidence, its author and its exact version, admitted only if it is well formed. Nothing arrives there by accident. Somebody decides that one sentence out of the argument is now a claim, takes responsibility for it, and puts it where other work can rest on it.

The crossing is the interesting part, and it is deliberate on purpose: the one moment where a judgement is made explicitly, that this is settled enough to be built on, at this scope, by this person. With a single wide door that judgement is made implicitly, by everything counting equally. With a single strict door it is made before the thinking, which is to say not made at all.

The paradox agents introduce#

There is a reason this shape looks more attractive now than it would have a few years ago. The structure that makes a record useful — named roles, typed values, explicit premises, an exact version for every reference — is close to unbearable for a person to maintain by hand, not because people cannot understand it but because filling it in correctly every time is clerical work of a kind research does not reward. The same structure is unremarkable for a machine: a schema that reads as bureaucracy to a person reads as an interface to an agent.

So the trade-off that forced earlier systems to choose between adoption and auditability may not be fixed, and the strict door can be stricter than a person would tolerate, provided the wide door stays genuinely wide. That is the hypothesis. If agent authoring produces well-formed records that are subtly wrong more often than useful ones, the trade-off comes back exactly as it was.

What each door costs#

The wide door costs volume and exposure. A Thread accumulates far more text than anyone will read, and the useful parts are not marked. It is also public, which is the point — a private working surface reproduces the invisible work the record was meant to expose — but thinking aloud in view of others is a real cost some will decline to pay. And a conclusion that lives only in a Thread has no scope, no status and nothing pointing at it.

The strict door costs effort and shape. Writing a record is more work than saying a thing, so fewer things get said. The schema also decides what is expressible: a result that does not fit the available shape gets bent until it does, or stays in the Thread. And a record can be well formed, correctly attributed, properly scoped, and wrong, because admission looks at the structure and never at the science.

Most of it never crosses#

The honest summary is that the majority of what matters in a Thread does not become a record, and no arrangement of doors changes that.

A constructed example. A Thread runs for two weeks about whether a comparison is fair. Somebody points out that the baseline was tuned on the same split; the design changes quietly in response; four of five proposed variants are dropped, three for good reasons and one because nobody had time. What crosses the strict door is a single finding with its scope and its evidence. The objection that reshaped the whole comparison stays prose, and the dropped variants leave no trace at all.

That is a loss, and naming it is better than pretending the structured record is the research. What two doors buy is not complete capture. It is a visible boundary: a reader can see what was claimed, see where the argument that produced it happened, and know which of the two they are reading.

In Substrate#

Both doors open into the same Room, and the same people and agents use both. Records are author-curated: admission checks structure, attribution and permission, never the science.

• The wide door. A Thread is public discussion in a Room — messages, replies, and the cards the work leaves as it happens ●Threads: Live. See Threads.
• The crossing is drafted where you can see it. A connected agent renders a draft record in your own chat for review, sending nothing ●Local publication preview: Live, and you publish once you agree. See Publish with your agent.
• And it can be crossed by hand. The Publish page writes the message for your agent, with a structured form as the alternative for anyone not using one ●Publish page: Live.
• The strict door refuses out loud. A refusal names its code and says whether anything committed and whether to retry ●Refusal envelope: Live, so a rejected draft is something you can fix rather than a silence. See HTTP API.
• Crossing once is enough. An exact version is linked into another Thread or Room with a stated purpose rather than copied ●Reuse into Threads: Live.

What does not cross stays prose. An open question raised in a Thread lives as text inside a checkpoint rather than as a record of its own, with an author and typed links to whatever would answer it ○Question records: Idea.

Open question#

Whether a wide conversational door and a strict machine-facing one really give both adoption and a record worth reading is not something argument settles, and neither is how much of a Room’s discussion ought to cross. The research agenda keeps both, with the evidence that would count.