Important to Whom?

Chapter 13 ended with an inheritance that looks like a gift and behaves like a problem. We know what the source itself says is important — the units, the agreement fractions, the minute-4 qualifier rescued by global context. And yet: the implementer does not need the literature-review units. The reviewer does not need the dependency versions. The adopter deciding on migration needs the limitation units above all. Same source, same units, same budgets — different survivors. The source digest was never the reader’s digest. It was the control condition for this chapter.

The progression:

CH 13

source → information units → source-level importance

                    ↓ add purpose

CH 14

task/query + information units
      ↓
task relevance
      ↓
task-conditioned digest

One new conditioning variable enters: purpose. Nothing about the person — no knowledge model, no history, no inferred traits. The task is declared, in words, and everything downstream is a relation to that declaration. Chapter 15 will subtract what the person already knows; this chapter must happily return units the user knows, because known-or-not is not its question.

The central sentence

Task relevance is not an intrinsic property of an information unit. It is a relation between an information unit and a declared purpose.

One technical talk, three digests. Task A — “decide whether to adopt this technique” — retains limitations, assumptions, cost and comparison units; drops reproduction minutiae. Task B — “reproduce the experiment” — retains configuration, versions, parameters, and the undocumented workaround; drops the motivational framing. Task C — “survey competing methods for a literature review” — retains positioning, comparisons, and citations; drops implementation detail. The source has not changed. The units have not changed. The relation changed, so the digest changed. Any system that stores a single relevance score per unit has mistaken a relation for a property, and every downstream policy built on that score inherits the error.

From queries to tasks

The literature enters through query-focused summarisation and the chapter immediately generalises beyond it. Prompts to Summaries (Barbara & Maalouf, arXiv:2506.10807 — verified: v1 June 2025, revised v3 February 2026, still arXiv/CoRR, cited here strictly as a mechanism precedent) segments video into scenes, describes them with batched VidLM captioning that scales to hours on a single GPU, scores scene importance with an LLM judge conditioned on an optional natural-language query, and propagates scores to frames via temporal-consistency and uniqueness metrics — competitive on SumMe/TVSum and on the Query-Focused benchmark without training data, plus a VidSum-Reason query-driven dataset as baseline. The architecture is exactly this chapter’s shape (segment → describe → judge importance under an optional query), which is why it is useful and why its evidential status must stay labelled: preprint mechanism precedent, not peer-reviewed evidence of reader benefit.

The peer-reviewed companion: Zhao, Ma, Yu and Gongyang’s sentence-driven query-focused summarisation (Displays, February 2026; ScienceDirect S0141938226000612, verified via publisher indexing) — richer natural-language sentences conditioning selection beyond predefined keyword concepts. Together they license the chapter’s narrow technical claim: natural-language purpose statements can condition selection over precomputed units. Neither licenses the stronger claims this chapter refuses (below).

But a query is not yet a task. “Tell me about performance” is an input string; “decide whether this is fast enough to justify migration from our existing system” is a purpose with a decision objective, a comparison target, constraints, and required evidence. The chapter therefore moves from query-focused summarisation (the entry mechanism) to task-conditioned selection (the general abstraction): relevance judged against a declared purpose structure — objective, target, constraints, evidence needs — with the raw query preserved alongside it so the conditioning stays auditable. That abstraction matters beyond video: it is the same declared-purpose object Part III’s Personal AI will condition interfaces on, introduced here in its simplest form.

Inspectable relevance, prohibited mysticism

One prohibition, stated as law: the chapter never says “the AI determines what is important.” The licensed sentence is always:

The system estimates relevance under a declared task.

And every estimate shows its work in a fixed, inspectable form:

UNIT
Benchmark improved 18%.

TASK
Can we reproduce this work?

RELEVANCE
medium

WHY
The result matters, but implementation details and
experimental configuration are more directly relevant.

Unit, task, judgement, reason — the reason citing the task’s evidence needs, checkable by the reader, disputable, loggable. A relevance score without a reason is a scalar mysticism the book has banned since Chapter 9; a relevance judgement with a reason is policy input the reader can override. Manual corrections of relevance (promote this, drop that) are first-class outputs of EXP-14, not failure noise — they are how declared tasks get refined into better declarations.

Adversarial task conflicts

The experiment’s sharpest cases are the conflicts — units whose source importance and task relevance point opposite ways:

• Source-central, task-irrelevant: the paper’s main theoretical contribution, immaterial to the deployment question. A task-conditioned system must demote what Chapter 13 promoted. If it cannot — if source importance leaks into every digest — conditioning is decorative.
• Source-minor, task-decisive: the minute-37 aside — “Windows clients require an undocumented workaround” — inside a fifty-minute database-architecture talk, for the task “can we deploy this in our Windows-heavy organisation?” A source-level digest reasonably omits it; the task-conditioned digest must surface it, or the chapter fails its headline example.
• Known-but-relevant: technique X requires dependency Y; the task (implement X) rates it HIGH; the user already knows it. Chapter 14 returns HIGH anyway — correctly, by its charter. Relevance is not novelty. Chapter 15 exists because this case is routine, and this chapter’s willingness to “fail” it is what keeps the ladder’s rungs separate:
• Task-ambiguous queries: “tell me about performance” under-specified between adoption, reproduction, and review readings. The system must surface the ambiguity (or condition conservatively across readings) rather than silently picking one.

Design of the comparison

EXP-14 freezes everything except the task — same source, same Chapter-12 events, same Chapter-13 units, same output budget, same underlying models where possible — across three genuinely different declared tasks (adopt / reproduce / review). Conditions: A the Chapter-13 source digest (control — does conditioning change anything?); B keyword matching against the task; C embedding similarity between task text and units; D generic LLM relevance judgement; E structured task-conditioned judgement with explicit reasons and required-evidence fields. Human judges score against the declared task, never generic quality: task-relevant unit recall, irrelevant retentions, critical omissions (false omission weighted heavier where the missing unit changes the task outcome — the Chapter 4/11 asymmetry again), redundancy, task-completion accuracy, time, and the inherited Chapter-9 preservation dimensions per digest. The relevance evaluator itself must first survive adversarial framings — high source importance with irrelevant task, minor detail with decisive task relevance, shared vocabulary with wrong task, low lexical overlap with correct task relationship — validated before it scores the experiment, per the book-wide instrument gate. Failure criteria: E ties D (structure adds nothing over generic judgement); E cannot demote source-central/task-irrelevant units (source importance leaks); E drops the Windows-workaround unit (the headline case fails); task-ambiguous queries silently resolve to one reading (ambiguity mishandled). Artifacts: frozen units, task declarations, relevance traces with reasons, per-task tables. What a positive result would not justify: anything about what the person needs — known HIGH-relevance units stay HIGH here by design, and their subtraction belongs to EXP-15.

What this chapter earned

Retention value changes with declared purpose, so relevance is modelled as a unit–task relation with inspectable reasons — query-focused mechanisms as entry point, task-conditioned selection as the abstraction, source digest as control, ambiguity surfaced, known-but-relevant returned HIGH without apology. What remains open is the person: of the task-relevant units, which would actually add something to this particular reader?

A good fifteen-point summary still assumes the same fifteen points matter to everyone — and now, worse, the same fifteen points matter to everyone with the same task.

References

• Barbara, M. & Maalouf, A. (2025–2026). Prompts to Summaries: Zero-Shot Language-Guided Video Summarization with Large Language and Video Models. arXiv:2506.10807 [cs.CV] (v1 Jun 2025; v3 Feb 2026). Verified via arXiv listing. Used strictly as mechanism precedent (segment → describe → LLM-judge importance under optional query → propagate); preprint status labelled; no reader-benefit claims taken.
• Zhao, K. et al. (2026). Query-focused video summarization driven by sentences. Displays (ScienceDirect S0141938226000612, Feb 2026). Verified via publisher indexing. Used as peer-reviewed companion for sentence-level NL conditioning beyond keywords.
• Lam et al. CHI 2026 Just-In-Time Objectives (lead, full read pending): noted as the inferred-objective counterpart this chapter deliberately does not use — declared tasks only; inference arrives with error bars or not at all.

Proposed experiment EXP-14: task variation under full freeze

Status: PROPOSED. Per the design above (A–E, three tasks, identical budgets, task-scored judging, omission weighting, preservation inheritance). The Ch-13 digest as control is the experiment’s keystone: any conditioning method must beat unconditioned source importance on task-scored measures while preserving profile fields. Task-conflict units (source-central/task-irrelevant; source-minor/task-decisive; known-but-relevant; ambiguous) required in the source set.