From Channels to Semantic Channels

Priya follows fourteen conferences, nine labs, thirty authors and four preprint feeds on distributed consensus. In March, a result appears that undermines the assumption her team’s design rests on, the one about bounded network delay. It is in none of her feeds. It is a paper on distributed databases, from authors she has never heard of, and it never uses the word “consensus”. It shows a mechanism that works without the delay bound she has been treating as essential. She finds it in June, by accident, cited in a thread about something else. Three months of work stood on the old assumption.

Her subscriptions did exactly what they were built to do. They delivered what their channels published. And they failed at what she needed: to be told when something she cares about changed, wherever it was published and whatever words it used.

Everything so far in this part has happened because someone asked. A person put a question, and the system worked out which units of an hour of media, or which items from a pile, mattered to the purpose and to what she knew. But Priya’s interest in her assumption does not switch off between questions. The system could keep watching on her behalf. That raises a question about what kind of thing she could ask it to watch for:

What would it mean to subscribe not to a source, a creator or a feed, but to a relationship in the world?

What people subscribe to now

The objects people subscribe to today are familiar:

a creator    a publication    a topic    a feed    a keyword

Each is a coarse contract. Following a creator means receiving everything that person emits, most of which is irrelevant, while everything outside the channel stays invisible, including whatever contradicts your work. A keyword alert delivers what mentions the word, which is not the same as what bears on your question, and it misses anything phrased differently. A topic is a bucket of things that mention something. None of them says what kind of change in the world would be worth knowing about.

The earlier chapters have built what is needed to say it. Part I established that similarity is not relationship and that relationships can be typed: this supports that, this contradicts that, this supersedes that. If relationships can be named, then a persistent interest can be stated as a relationship too, and not as a topic.

typed relation
      ↓
standing relational query
      ↓
semantic channel

The difference is the one between two questions:

topic subscription  ≠  relationship subscription

A topic subscription asks: what new things mention X? A semantic subscription asks: what new things change my understanding of X in this particular way?

A small object

A semantic subscription needs only four things, and it helps to keep it that small.

ANCHOR       what the subscription is about
RELATION     what kind of change or connection matters
CONDITIONS   when something qualifies
EXCLUSIONS   what should not trigger it

Here is Priya’s, written out:

ANCHOR
this protocol's safety result

RELATION
contradicts, or weakens an assumption of

CONDITIONS
a new, peer-reviewed result
that bears on the bounded-delay assumption

EXCLUSIONS
tutorials
duplicate reporting
pure opinion

This is more expressive than “subscribe: consensus algorithms”, and it does something the keyword version cannot. The paper from March might share almost no vocabulary with the protocol it affects. A relation-first subscription is matched on what the paper does to the anchor, not on whether it uses the same words.

Three more, to show how different these are from topics:

"Tell me when a new result weakens the assumption that this
protocol needs bounded network delay."

"Show me new evidence that contradicts the causal claim in
this paper."

"Tell me when this API's recommended practice is superseded."

None of these is a topic. Each is a standing relation between something the person already cares about and whatever the world produces next, with change over time built in. Stating them this way makes semantic channels look less like a speculative feature and more like what people were already trying to do with alerts and reading lists, with the wrong tools.

Delivery is deliberately not in the object. How and when something should reach her is a separate decision, taken up shortly. The subscription settles only one question: does this new thing qualify?

Subscriptions are not recommendations

Recommendation systems sit nearby, and the difference is worth being exact about.

RECOMMENDATION   the system infers what you may want
SUBSCRIPTION     the person declares what relation should matter

A recommender asks what item this person should consume next, usually from a history of what she consumed. A semantic subscription asks what change in the information world would satisfy a relationship she has explicitly said matters. The difference has teeth. Someone can subscribe to “anything that contradicts my current architecture”: content she would never have clicked, that no engagement model would ever surface, that exists precisely to disturb. A system that learned from her history would be tempted to suppress the very contradictions and replications the subscription is meant to catch. There is also something recommendation research has noticed: some of the most useful information is useful because the person would not have searched for it. A declared relation can reach it. A history of past clicks usually cannot.

Matching is not confirmation

The discipline from Part I carries forward intact. A system may find something that might contradict the anchor. That does not mean it has established a contradiction.

candidate relation  ≠  confirmed relation

So a subscription can fire on a candidate while keeping its uncertainty visible. The result is not “this contradicts your paper”. It is a candidate contradiction, with the evidence for it and a standing: possible, probable, or confirmed by someone. Otherwise semantic channels become machines for asserting things automatically, and every false positive arrives with the authority of a structure.

The same care applies to scope. Suppose Priya subscribes to contradictions of her protocol’s safety claim and the system finds a paper on a different failure model, a different population of nodes, or a different set of assumptions. The wording may look like disagreement. A contradiction outside the relevant scope is not a contradiction of her claim, and the label has to preserve the scope it asserts. And finding a matching relation does not show the matched evidence is current, reliable or independent. A real match to a weak source is still a weak source.

An interest is not a profile

The last chapter built a narrow model of what a person knows. This one should build an equally narrow model of what she wants to watch.

standing subscription  ≠  profile of the person
explicit standing interest  ≠  inferred identity

Priya can subscribe to “new safety failures in this protocol” without the system concluding that she is generally interested in safety engineering, or anything else about who she is. The subscription is a declaration. It does one thing, the thing it says, and it says exactly what it is. This is why it must be declarative and inspectable: she should be able to open a view that says here is what I am currently watching for and see each subscription as she wrote it, change it, and delete it. An opaque inference about her interests is a hidden decision-maker. A list of the relations she asked to be told about is a tool she owns.

How a semantic channel goes wrong

Because the object is simple, its failures are easy to name, and worth naming before anyone builds one.

too broad          fires constantly, until she stops reading it
too narrow         misses meaningful change
wrong relation     surfaces similarity when she asked about contradiction
stale anchor       no longer reflects what she cares about
silent expansion   the system broadens the subscription without permission

The last is the most dangerous. A system that quietly widens what it watches for, because it has guessed she would like more, has turned a declared interest into an inferred one and taken away her ability to say what she wants. A subscription may be offered a refinement. It should never be rewritten without her say. The stale anchor is the quiet cousin of this: people change projects and assumptions, and a subscription that outlives its reason produces noise that looks like diligence. Subscriptions should be easy to correct, and the cost of correcting them should be counted, because one that cannot be edited cannot really be owned.

Where this leaves us

A subscription can be to a relationship and not a source: an anchor, a relation, conditions and exclusions, matched on what a new thing does to the anchor and not on its words. It sits one step beyond a typed relation, as a standing query over changing information. It differs from a topic because it asks what changes understanding, from a recommendation because the person declares it, and from a profile because it says only what it says. Matching yields candidates with evidence and standing, not confirmed assertions. And it stays visible and editable, so the thing being watched for is always something the person can read.

But look at what this settles and what it leaves open. A standing semantic channel can determine that something qualifies for attention. It does not determine when, how, or whether that thing should reach the person.

qualifies
    ↓
candidate for attention
    ↓
delivery decision

A contradiction found at three in the morning is a candidate for attention. Whether it should arrive as a single line or a full account, whether it should wait for the morning, depends on how much detail the situation calls for and what condition the person is in when it comes. The first of those is the next question: when something qualifies, how much of it should reach her first?