Don't Tell Me What I Already Know

A senior engineer asks for a task-conditioned digest on migrating her team’s retriever to a new vector index — Task A from Chapter 14’s world, implementation reading. The system returns fifteen HIGH-relevance units. She knows eleven of them cold: she built the current index, she has tuned HNSW parameters for two years, she wrote the internal wiki on product quantisation. Four units are new — the undocumented Windows workaround, a changed default in the new version, two benchmark caveats — buried at positions 9, 11, 13, and 14 beneath what she could have recited. The digest is relevant and useless in proportion: every known unit consumed attention to deliver zero information.

Chapter 14 would score that digest highly. Every unit is HIGH relevance under the declared task — correctly, by Chapter 14’s charter. This chapter exists because relevance is not enough. Its question is the first genuinely personal one in the book, kept deliberately narrow:

Given information that is relevant to the declared task, what would actually add something to this person’s current knowledge?

Not preference, taste, identity, or broad profiling. Knowledge — uncertain, evidence-dependent, prerequisite-sensitive — and only knowledge.

Two dimensions, not one scalar

The chapter refuses the tempting formula novelty = relevance − knownness, because subtraction invites a single misleading number. Instead, two preserved dimensions:

TASK RELEVANCE
Does this information matter for the declared purpose?
(Chapter 14's judgement, inherited intact)

        ×

USER KNOWLEDGE
What evidence do we have that this person already knows it?

        ↓

PERSONAL NOVELTY
Would surfacing this unit probably add useful information?

Chapter 14’s relevance stands untouched — the known-but-relevant unit stays HIGH relevance, and the novelty layer decides presentation, not worth. Each unit then carries a knowledge state, and the state vocabulary matters: NEW, KNOWN, UNCERTAIN, STALE, PREREQUISITE. The last three do the chapter’s real work. UNCERTAIN marks the vast middle where evidence is thin. STALE marks demonstrated-then-lost knowledge — the concept mastered two years ago and failed today. PREREQUISITE marks units that must surface regardless of knownness because retained novel material depends on them. A binary known/new filter cannot express any of these, which is why binary filters fail.

Seen ≠ read ≠ understood ≠ remembered

UNCERTAIN is first-class because the evidence for knowledge is radically unequal, and the chapter gives every knowledge entry provenance:

KnowledgeState {
    information_unit
    state            NEW / KNOWN / UNCERTAIN / STALE / PREREQUISITE
    confidence
    evidence         EXPLICIT / DEMONSTRATED / EXPOSURE / INFERRED
    elicitation_context
    observed_at
    user_confirmed
}

Here KNOWN carries the narrowed technical sense fixed in PASS A: evidence that the specified knowledge is currently accessible under specified conditions — never “the person knows X” as a durable trait. The five states are operational, not psychological: NEW means no accessibility evidence; KNOWN means current-accessibility evidence above threshold in the recorded elicitation context; UNCERTAIN means thin or conflicting evidence; STALE means previous accessibility evidence that may no longer justify suppression; PREREQUISITE means load-bearing for retained novel material regardless of other states. Retrieval success in one context predicts nothing about later availability elsewhere — currently retrievable ≠ durably available — so EXP-15 pairs every T0 accessibility estimate with a T1 delayed-retrieval check, and suppression policy reads the pair, never the instant alone.

Four evidence tiers with deliberately unequal authority. EXPLICIT (“I already know this,” or “I don’t”) outranks all. DEMONSTRATED (the user correctly used or explained the concept) is strong but decays — demonstration two years ago is STALE until refreshed. EXPOSURE (content containing the unit was shown) is weak: opened is not read, read is not understood, understood is not remembered. INFERRED (related behaviour suggests possible knowledge) is weakest, always provisional, never sufficient alone for suppression. The chapter’s strongest boundary follows directly:

Seen ≠ read ≠ understood ≠ remembered. Exposure especially must not equal knowledge.

The literature supports exactly this shape while stopping exactly where the chapter stops. PersonaX (Shi et al., Findings ACL 2025, pp. 5764–5787, verified via ACL Anthology) shows long behavioural histories compressible into multiple textual personas that improve recommendation agents (AgentCF +3–11%, Agent4Rec +10–50%, from 30–50% of behaviour data) — evidence that user models can be built from long histories, and simultaneously the chapter’s warning exhibit: those profiles model interests and preferences, not knowledge. Behaviour-as-interest is the wrong variable for this job, and PersonaX’s sophistication makes the misapplication tempting rather than justified. Knowledge tracing is closer to the real problem: DPKT (Yang et al., Scientific Reports 2025, 15:11475, verified) models programming learners’ uncertain, changing mastery with LLM-extracted difficulty features, dynamic updates, and attention over concept relations — an explicit uncertain-knowledge-state precedent. But its results come from educational interaction settings (problem attempts, submission records), so the chapter treats it as mechanism precedent — uncertain states, dynamic updates, difficulty-awareness — not proof that general knowledge is inferable from browsing history. The gap between tutoring interactions and ambient reading is where this chapter’s conservatism lives.

The prerequisite rule: known does not mean deletable

The crucial mechanism. Consider three units: A (“vector databases use ANN indexes,” KNOWN, high confidence), B (“HNSW’s efSearch controls recall/latency,” NEW), C (“this benchmark changes efSearch without holding construction parameters fixed,” NEW). Naïve filtering deletes A. Usually harmless — until a retained novel unit depends on a supposedly known prerequisite whose state is merely probable. Then deletion strands the reader: C is incomprehensible without the A the system hid.

Suppression therefore requires a conjunction, never a single flag:

relevant + probably known + NOT required to interpret
retained novel material
        ↓
SUPPRESSIBLE

not known → delete. And presentation follows confidence, not binary fate:

high-confidence known      → candidate for suppression
medium-confidence known    → compress / offer collapsed
uncertain                  → retain
prerequisite               → retain or make expandable

The model grows more conservative as confidence falls — the safe direction. This graduated policy is also the inheritance Part III needs: when the Personal AI compresses aggressively, the prerequisite rule is what stops it deleting load-bearing foundations.

The chapter’s own asymmetry, joining Chapters 4 and 11: FALSE NOVEL (showing the known, wasting attention) versus FALSE KNOWN (suppressing the needed, removing information). The second is costlier, so ties break toward retention. A novelty layer that occasionally bores is functioning; one that strands is failing.

Adversarial users, not predictable clicks

EXP-15 freezes the Chapters 13–14 stack (corpus, events, units, tasks, relevance judgements, budget) and varies only the knowledge model: A no model (task digest); B exposure-derived (“previously encountered”); C behaviour/inference-derived; D explicit user-declared; E hybrid uncertain model with provenance and confidence. Judged against an independently constructed human knowledge reference — never recommendation improvement, but: known units removed, novel units surviving, prerequisites wrongly removed, false knowledge claims, attention saved, comprehension/task-performance change, user corrections required, Chapter-9 dimensions where transformation occurred. The adversarial users break behavioural inference on purpose: exposure without understanding (five transformer articles opened, attention unexplained); expertise without history (the database engineer with no logs); forgotten knowledge (demonstrated two years ago, failed today); terminology mismatch (mechanism known, new paper’s terms unknown); partial knowledge (HNSW known, this parameter interaction not); prerequisite dependency (known unit load-bearing for a novel one). Condition B should fail exposure-without-understanding by construction; condition C should fail the cold-start expert; only D and E have paths through all six — which is the chapter’s empirical bet, stated before running.

Explicit control, preference fenced out

The model is inspectable from its first deployment, following the architectural point of Wozniak et al. (WWW Companion 2025, DOI 10.1145/3701716.3717734, verified venue/DOI; numeric magnitudes held for evidence audit): user-editable textual profiles outperforming history-only representations supports making personal models correctable rather than treating logs as ground truth:

You think I know: embeddings, cosine similarity, HNSW
You are unsure whether I know: product quantisation, efConstruction
[correct] [forget this assumption] [show evidence]

Correction, forgetting, and evidence display are controls, not feedback forms — the agency groundwork Part III inherits without building the Personal AI early. And the final fence, stated as inequalities the book will reuse:

"I know this" ≠ "I like this"
"I have seen this" ≠ "I know this"
"I often click this" ≠ "I need more of this"
"I rarely click this" ≠ "I already understand it"

Interest modelling is a different job with different variables. Recommender logic — optimise for engagement with the familiar — must not quietly take over a novelty layer whose entire purpose is information gain. PersonaX stays cited for what it proves (interest from history) and fenced from what it doesn’t (knowledge from behaviour).

What this chapter earned

Relevant is not novel: task-relevant units carry uncertain, evidence-graded knowledge states (NEW/KNOWN/UNCERTAIN/STALE/PREREQUISITE with EXPLICIT/DEMONSTRATED/EXPOSURE/INFERRED provenance), suppression requires the three-way conjunction with prerequisite protection, confidence falls toward retention, false-known outweighs false-novel, adversarial users test inference rather than clicks, and the model is editable from birth. What remains open is structural: relevance and novelty are still computed per item, searched per query. The person still subscribes to channels.

Once the system has source units + task relevance + personal novelty, subscriptions can stop meaning “follow this channel.”

References

• Shi, Y. et al. (2025). PersonaX. Findings ACL 2025, pp. 5764–5787. DOI 10.18653/v1/2025.findings-acl.300. Verified via ACL Anthology. Used for interest-from-history capability AND as warning (interests ≠ knowledge).
• Yang, L. et al. (2025). Difficulty aware programming knowledge tracing via large language models. Scientific Reports 15:11475. Verified. Used as uncertain-dynamic-knowledge-state precedent; educational-setting fence.
• Wozniak et al. (2025). Improving LLM-Based Recommender Systems with User-Controllable Profiles. WWW Companion 2025. DOI 10.1145/3701716.3717734. Verified venue/DOI. Used for editable-profile architecture; numeric magnitudes held for audit.
• Memory-book inheritance (durable past vs present selection; temporal validity for STALE): conceptual, pending sibling verification at Part II pass.

Proposed experiment EXP-15: knowledge models under freeze

Status: PROPOSED. Freeze: corpus, events, units, tasks, relevance judgements, budget (Ch 13–14 stack). Vary: knowledge model A–E above, with adversarial user profiles (six breakage cases) and independent human knowledge references. Measures: known-removed, novel-surviving, prerequisites-wrongly-removed, false-knowledge claims, attention saved, comprehension/task change, corrections required, Ch-9 dimensions on transformed outputs. Failure criteria: B/C tie D/E (inference suffices — conservatism unnecessary); E suppresses prerequisites (conjunction broken); false-known rates exceed predeclared tolerance (uncertainty machinery decorative); corrections unusable (inspectability theatre). Artifacts: frozen stack, knowledge references, model traces with provenance, per-case tables. What a positive result would not justify: preference modelling, identity inference, or any trait beyond unit-level knowledge — the narrowest personalisation in the book stays narrow.