The Personal Information Radar

Build the Too Much Information capstone over conferences, subscriptions and wider discovery.

Part II built nine mechanisms and ends with none new. Attention as budget. Events from duration. Units from events. Source importance, task relevance, personal novelty — three different judgements that refuse to collapse. Standing subscriptions instead of channels. A progressive ladder with contracts. An autopsy that stops compression. Timing separated from worth. This chapter composes them into one system and asks the integration question Part I taught us to ask: do independently earned pieces still help once composed, or do they silently rewrite each other?

INFORMATION WORLD
        ↓
event segmentation          Ch 12
        ↓
information units           Ch 13
        ↓
source importance           Ch 13
        ↓
task relevance              Ch 14
        ↓
personal novelty            Ch 15
        ↓
semantic subscriptions      Ch 16
        ↓
resolution ladder           Ch 17
        ↓
preservation gate           Ch 18
        ↓
timing policy               Ch 19
        ↓
PERSONAL INFORMATION RADAR

The chapter’s explanatory device is waste: each stage removes a different failure mode, which is why the Radar is not seven implementations of “relevance”:

13  source redundancy          (the undifferentiated pile)
14  task irrelevance           (important to no present purpose)
15  already-known information  (relevant but learned)
16  channel dependence         (invisible because unpublished nearby)
17  unnecessary detail         (full text where a glance decides)
18  destructive compression    (detail removed that decisions needed)
19  badly timed attention      (right item, wrong moment)

Remove any stage and its waste returns in full — the composition argument in one table.

The no-overwrite contract

The integration’s load-bearing rule: downstream stages may not rewrite upstream judgements. Task relevance HIGH plus novelty LOW does not become relevance LOW — the unit stays HIGH-relevance-and-known, and the novelty stage decides presentation (suppress, collapse, retain-as-prerequisite) on recorded grounds. Resolution SUMMARY plus timing DIGEST does not let timing recompress to GLANCE for queue-fitting. Every stage’s evidence persists separately in the decision record:

RadarDecision {
    source_importance      (Ch 13 judgement + agreement class)
    task_relevance         (Ch 14 judgement + reason)
    knowledge_state        (Ch 15 state + provenance + confidence)
    novelty                (derived, with derivation shown)
    subscription_relation  (Ch 16 match + scope confidence + UNKNOWN where evidence is thin)
    evidential_validity    (current/reliable/independent/safe-to-act-on — Chapter 16's fence, scored not assumed)
    resolution             (Ch 17 rung + contract reference)
    preservation_profile   (Ch 18 per-rung vector)
    delivery_route         (Ch 19 route + reason)
}

Every field preserves its uncertainty instead of collapsing to a crisp value — confidence, agreement class, scope confidence, and UNKNOWN travel with the judgements that produced them, because a decision record that launders doubt into fields is a confidence trick with good schema. Presented as conceptual contract, not code: each field inspectable, each with its provenance, none collapsible into a late-stage scalar. The moment one scalar quietly dominates — a final “priority score” multiplying everything — composition dies and attribution with it. EXP-20’s conflict cases exist to detect exactly that death.

This contract is also Part III’s inheritance mechanism: the Personal AI will need to explain why something appeared as it did, and the RadarDecision is the explanation substrate — every downstream presentation choice traceable to the stage evidence that authorised it.

Ablations that report damage

EXP-20 climbs A–H (chronological subscriptions → semantic retrieval → source+task filtering → novelty → subscriptions → adaptive resolution → preservation gate → timing policy) on the frozen Part II world with staged reveals, and every rung reports both gains and harms across shared dimensions: useful novel information surfaced, important information missed, irrelevant and duplicate attention consumed, decision-critical omissions, preservation violations, harmful interruptions, time consumed, manual corrections. The expected story is non-monotonic and therefore informative: novelty saves attention but increases false suppression; the gate restores safety at compression cost; timing lowers interruptions but delays awareness. A victory march would teach nothing; priced trade-offs teach deployment — which rung earns its keep under which task economics.

Component conflicts adjudicate compositionality: HIGH-relevance known non-prerequisites compress away while known prerequisites retain context; novel contradictions with low scope confidence never foreground as clean contradictions; GLANCE-safe low-urgency items digest while DETAIL-demanding urgent items interrupt. If the pipeline resolves all five correctly with stage evidence intact, it is compositional. If late stages systematically override — everything urgent becomes DETAIL, everything novel becomes NOW — one scalar has eaten the pipeline, and the chapter reports the failure as the finding. Attribution discipline from Chapter 10 applies unchanged: each rung evidences its composite stage, never causal proof of components, with frozen candidate pools, generators, models, budgets, and judges wherever the stage is not supposed to alter them. And the chapter closes the Part with the verdict the hostile review demanded: the Radar is first testable as a controlled closed-world selection pipeline; robust open-world discovery additionally requires corpus incompleteness, distribution change, stale and conflicting evidence, adversarial sources, and unseen relationship structures — none of which the frozen world supplies.

Inspectability as the output

The Radar’s final form is an inspectable mediation system, demonstrated in the trace every delivered item carries:

Why am I seeing this?
Because: relates to Project X; contradicts assumption Y;
appears novel to you; relation passed evidence threshold;
SUMMARY is the lowest safe resolution; queued, not
interrupted — urgency was low.
Show: [full detail] [source] [why novel]
[why contradiction] [change this policy]

Correction paths attach per field — dispute the relation, the novelty, the resolution, the route — so each correction trains its stage rather than vanishing into a global thumbs-down. That per-stage correctability is what makes the system governable rather than merely personal, and it is the strongest thread Part III inherits. Two integration verdicts close the Part. First, purpose-branching: SCAN/DECIDE/REFERENCE tasks may take the lowest sufficient safe resolution, but LEARN tasks obey a different contract — additional context, active reconstruction, retention/transfer measurement — because semantic preservation is not learning preservation, and the lowest rung that permits a correct immediate answer is not thereby the right rung for durable understanding. Second, regime honesty: EXP-20 is initially a closed-world mechanism test, full stop — open-world Radar claims additionally require corpus incompleteness, distribution shift, new relation types, stale and conflicting evidence, adversarial sources, source-quality variation, and unseen relationship structures, none of which the frozen world supplies. And output here depends on multiple experimental instruments at once: where a gating evaluator fails validation, the corresponding conclusion is INCONCLUSIVE, never merely lower-confidence — PASS B imported, not re-argued.

Part III changes the object of adaptation again: the Radar decides what information reaches the person, how much of it, and when — but still presents it through a generic interface. What if the interface itself becomes personal?

References

No new sources. Composes Chapters 11–19 evidence (Simon; LongVALE/Chapter-Llama/LVBench; Lee LLMVS; QEVA; Prompts-to-Summaries; Zhao sentence-conditioning; PersonaX; DPKT; Wozniak; SOLAR; Ou & Lapata; NexusSum; Retkowski; PACO; Mujahid; ProMemAssist; Krisam; Lam) under ledger states CL-11–CL-22.

Proposed experiment EXP-20: Radar ablation ladder

Status: PROPOSED. Frozen world + staged reveals (extends EXP-16 protocol through the full stack); conditions A–H; shared harm/benefit dimensions; conflict-case battery; correction-effort accounting. Failure criteria: H ties C–E on net utility (composition unworthwhile); late-stage overrides detected (scalar capture); corrections unusable per-stage (governability theatre); misses concentrate in contradictions (the valuable discovery lost). Artifacts: world freezes, decision records, per-rung harm/benefit tables, conflict resolutions. What success would not justify: preference modelling, interface personalisation, cognitive-state inference — Part III’s problems, explicitly not premises.