← Language From First Principles

From Channels to Semantic Channels

Subscribe to ideas, questions and relationships rather than whole creators or feeds.

A researcher follows fourteen conferences, nine labs, thirty authors, and four preprint feeds on her topic — retrieval evaluation. In March, the result that overturns her architecture claim appears in none of them: a reinforcement-learning paper, from authors she has never heard of, using none of her keywords, demonstrating a mechanism that contradicts her scaling assumption. She finds it in June, by accident, cited in a thread about something else. Three months of work rest on the old assumption. Her subscriptions worked exactly as designed — they delivered what their channels published — and failed at exactly what she needed: telling her when something she cares about changed, regardless of where it was published.

That failure is structural to channel subscriptions. Following a creator is a coarse contract: you receive everything they emit (most of it irrelevant to you) while everything outside your channels stays invisible (including the contradiction that matters). Chapters 13–15 built the machinery to judge units — source importance, task relevance, personal novelty — but every judgement so far runs against a presented pile. This chapter turns the system persistent: the user declares relationships worth discovering, and new information qualifies when the world changes, whether or not anyone is looking.

The new object:

SEMANTIC SUBSCRIPTION

anchor
    what claim / concept / project / question am I tracking?

relationship
    what kind of change am I interested in?

conditions
    what would make the candidate qualify?

novelty requirement
    must it add something I probably do not already know?

evidence threshold
    what evidence must accompany the candidate?

exclusions
    what should not count?

A worked instance:

ANCHOR
Our current claim: "Method X is the best-performing
approach under constraint Y."

SUBSCRIBE TO
CONTRADICTS · UPDATES · REPLICATES · IMPLEMENTS

CONDITIONS
new evidence; relevant benchmark;
same or comparable constraint

NOT
generic articles mentioning Method X
duplicates · restatements

The stream is now organised around relationships to something the person cares about rather than publishers or channels. That is the chapter’s conceptual contribution, and it extends Chapter 8 cleanly: Chapter 8 asked how a candidate relates to what is currently being read; Chapter 16 asks what relationships should remain active even when no one is looking.

Subscriptions are not recommendations

The distinction deserves explicit treatment because recommender logic will otherwise annex the chapter:

A recommender asks: what item should this person consume next? A semantic subscription asks: what change in the information world would satisfy a relationship I have explicitly declared worth knowing about?

The difference is enormous in practice. A user can subscribe to anything that contradicts my current architecture — content they would never click historically, that no engagement model would ever surface, that exists precisely to disturb rather than please. Or to any independent replication of this result, any implementation making this theoretical technique practical, any update invalidating this benchmark. These are standing informational relations the user owns — declared, editable, scoped — not predictions of next consumption. Optimising for clicks would suppress exactly the contradictions and replications the subscription exists to catch.

SOLAR (Yuan et al., Findings EMNLP 2025, pp. 10146–10169, verified via ACL Anthology) is useful here in both roles — precedent and contrast. As precedent: a two-stage pipeline (sequential ID-based recommender generating candidates, LLM reranker with weak supervision producing serendipity-aware pseudo-labels, plus domain-adaptive SUN instruction tuning) improving both accuracy and serendipity over strong baselines on three real-world datasets — evidence that optimising only for familiar relevance misses useful unexpected material, and that unexpectedness can be pursued without abandoning accuracy. As contrast: SOLAR remains a recommendation setting — user histories, preference judges, serendipitous items — while this chapter defines a different object, a user-owned standing relation. The serendipity literature contributes one importable idea, fenced: some useful information is valuable precisely because the user would not have searched for it — so the system reserves capacity for low-expectedness items with credible relations, task relevance, and novelty. Unexpectedness alone never qualifies; surprise is not value, and the fence prevents “serendipity” becoming a licence for noise. Quantitative definitions of serendipity remain unsettled in that literature, and the chapter claims nothing about them.

The layering: watch, match, qualify — not deliver

Three separations keep the subscription honest:

SUBSCRIPTION
What relationship am I asking the system to watch for?
        ↓
MATCH
Does this new information instantiate that relationship?
        ↓
QUALIFICATION
Is it relevant + sufficiently novel + sufficiently evidenced?
        ↓
AVAILABLE INFORMATION
        ↓
DELIVERY / INTERRUPTION  →  Chapter 19

Matching is Chapter 8’s readout applied to streaming candidates; qualification reuses Chapters 14–15 (task relevance, novelty requirement from the subscription itself). Neither stage notifies. A contradiction discovered at 03:00 becomes eligible information — available in the morning’s structures, not a 03:00 alert. Timing and intrusiveness belong to Chapter 19 absolutely; this chapter that notifies is a chapter that has trespassed. Likewise compression aggression is Chapter 17’s: qualification decides whether information is eligible, never at what resolution it arrives. And one validity fence stands over all of it: a matching relationship does not establish that the matched evidence is current, reliable, independent, or safe to act upon — stale, conflicting, duplicated, low-quality, or possibly adversarial evidence must face explicit conditions in EXP-16 (or be queued as stated debt), because relationship ≠ evidential validity, a separation Chapter 9 already supports.

Cross-domain discovery: where RELATE lineage pays

The subscription’s hardest test is the item no channel logic can find. The user’s project concerns context compression. A new paper mentions none of her keywords, authors, or venues — but demonstrates a mechanism that implements the same relation she tracks, in another field. Keyword alerts miss it. Creator subscriptions miss it. Pure embedding retrieval may rank it poorly (shared mechanism, divergent vocabulary). A relationship-aware subscription — anchor plus relation plus conditions, matched on relational structure rather than topical overlap — can surface it. That is the RELATE lineage made operational without claiming RELATE solves it: relation-first matching over streaming candidates, with the readout’s limits (Chapter 8’s fences) inherited intact.

The mandatory failure case guards the victory. Subscribe to CONTRADICTS claim X, then present: same topic but a different question, a different population, a different period, different assumptions, an actual contradiction, a partial contradiction. Naïve disagreement-detection calls all of them contradiction. Chapter 16 punishes that heavily — per-relationship-type scoring (DUPLICATES, SUPPORTS, CONTRADICTS, UPDATES, EXTENDS, IMPLEMENTS, REPLICATES reported separately, never averaged) requires scope preservation: a contradiction outside the relevant scope is not a contradiction of the subscribed claim. Chapter 9’s preservation requirement returns as relationship hygiene — labels must preserve the scope they assert, and the qualification stage checks it.

Design of the comparison

EXP-16 runs on the frozen Part II world with temporal staging: at T0 freeze projects, questions, units, knowledge models, and subscriptions; reveal new material at T1, T2, T3. Conditions: A creator/channel subscriptions; B keyword alerts; C embedding-similarity alerts; D generic LLM relevance alerts; E semantic subscriptions (relation + task relevance + novelty). The evaluation unit is informational change caught: relationship-specific recall, false alerts, duplicates, genuinely new discoveries, contradictions missed, cross-domain catches, evidence sufficiency, subscription corrections. Broken out by relationship type throughout — no single accuracy number. Failure criteria: E ties D (relations add nothing over generic relevance); E catches contradictions but floods duplicates (matching without qualification); scope-violating “contradictions” score as hits anywhere (adjudication broken); cross-domain items caught only when keywords accidentally overlap (relation matching decorative). Artifacts: staged freezes, subscription pack, change keys, per-relation tables. What a positive result would not justify: timing, resolution, or that anyone should be interrupted — eligibility only.

What this chapter earned

Information interests as standing subscriptions to semantic relationships — anchor, relation, conditions, novelty requirement, evidence threshold, exclusions — evaluated by change-caught rather than clicks, layered as watch/match/qualify with delivery and compression owned downstream, fenced from recommender logic, with cross-domain discovery as the RELATE-lineage prize and scope preservation as its guardrail. The system is no longer a filter over today’s pile. It is an ongoing interface to a changing world.

The result should not be one irreversible summary; readers need multiple levels of resolution.

References

  • Yuan, Z. et al. (2025). SOLAR: Serendipity Optimized Language Model Aligned for Recommendation. Findings EMNLP 2025, pp. 10146–10169. DOI 10.18653/v1/2025.findings-emnlp.538. Verified via ACL Anthology. Used as precedent (beyond-familiar relevance without abandoning accuracy) AND contrast (recommendation setting vs standing relations); serendipity-quantification claims not taken.
  • Ch 08 relation readout + discovery vocabulary; Ch 14–15 relevance/novelty; Ch 09 scope/preservation: reused as subscription machinery, not re-explained.

Proposed experiment EXP-16: staged-world subscription comparison

Status: PROPOSED. Per the design above (T0 freeze, T1–T3 reveals, A–E, change-caught scoring per relationship type, scope-preservation adjudication, cross-domain cases required). Subscription-correction effort measured as first-class cost (uneditable subscriptions are unownable).