The Semantic Browser

Part I has built five mechanisms and one instrument. A router that chooses representations or declines. A record that stabilises structure. A sidecar that enriches without interrupting. A readout that types relationships. A profile that measures what survived. Each was earned against its own failure, in isolation, with its own experiment. None has yet met the others.

That meeting is this chapter — and the chapter’s governing suspicion is that the meeting might go badly. Components that pass alone can fail together: the sidecar’s useful items become distraction beside a demanding diagram; the preservation gate’s checks add latency that kills scanning; three individually excellent panels compose into noise. Integration is not victory laps. It is the experiment where the book discovers what its machinery costs.

The architecture under test is deliberately simple — composition, not invention:

SOURCE
  ↓
SEMANTIC RECORD
  ↓
REPRESENTATION ROUTER
  ├── VIEW
  ├── SIDECAR
  │     ↓
  │  RELATION READOUT
  ↓
PRESERVATION GATE
  ↓
USER

Read evidentially, the same pipeline carries a licensing chain:

SOURCE
   ↓
SEMANTIC RECORD
   ↓
REPRESENTATION / RELATION MECHANISM
   ↓
INSTRUMENTED EVALUATION
   ↓
PRESERVATION GATE
   ↓
DOWNSTREAM TASK

Nothing here is new. The record is Chapter 6’s, the router Chapter 5’s, the sidecar Chapter 7’s, the readout Chapter 8’s, the gate Chapter 9’s profile applied as policy. This chapter introduces no concepts, coins no terms, and proposes no mechanisms. Its only job: can these independently earned pieces operate together without the system becoming slower, noisier, less faithful, or less useful than the fixed page it was supposed to improve? One attribution caveat governs the whole ladder: each rung localises gains only to the extent successive conditions differ in the declared mechanism — where implementation necessarily changes several things at once, the result evidences the composite stage, never proof of an individual causal component. An ablation ladder is evidence about attribution only where adjacent conditions differ in the declared mechanism — that constraint, not the ladder itself, is the doctrine. And one licensing rule runs through every stage, inherited from the repaired Chapter 8: no stage may inherit a stronger evidential claim than the stage before it has earned — encoded ≠ decodable ≠ relation-classifiable ≠ retrievable ≠ useful downstream, with each arrow requiring its own evidence. The full protocol for every rung — CHANGED, FROZEN, COMPOSITE yes/no, excluded confound per baseline, evaluator gate — is governed by planning/experimental-method.md; the chapter states the requirement, the method owns the machinery.

The ablation ladder: attribution before verdict

A fixed-page-versus-everything comparison would answer the wrong question. If the full browser wins, we learn nothing about which components earned the win; if it loses, we learn nothing about what to fix. EXP-10 is therefore a ladder, each rung adding one mechanism so failures attribute:

A  fixed source page

B  source + representation routing

C  B + semantic record (mediated rendering)

D  C + sidecar enrichment

E  D + relation-aware readout

F  E + preservation gate (full browser)

No full permutation matrix is needed — six rungs suffice to localise the common failures. If C beats B, the record constrains generation as Chapter 6 hypothesised. If D beats C on discoveries but loses on task time, enrichment’s distraction price is quantified rather than debated. If E beats D on contradictions found but not on completion, typed relations help finding while costing attention. If F beats E on fidelity while losing SCAN speed, the gate’s latency is priced. Each comparison is a chapter hypothesis re-tested in composition, which is exactly what “earned in isolation” obliges.

The chapter’s standing principle: a component can pass in isolation and still fail in composition. The sidecar that surfaced useful items in EXP-07 may reduce task performance here through distraction. The gate that improved fidelity may make reference lookup worse through latency. The router’s best view plus the sidecar’s best items may exceed attention budget where either alone fit. These are not disappointments; they are the integration effects this chapter exists to discover, and a ladder that cannot reveal them is decoration.

Four tasks, because mediation costs differ

The full system is tested across four task types — SCAN, LEARN, IMPLEMENT, REFERENCE — because mediation may earn its cost on one and lose on another, and the book wants that texture, not a single verdict:

• SCAN: triage a set of pages under time pressure; keep-or-close decisions with false skips counted. Mediation must be fast or it loses to the title list.
• LEARN: grasp a technical topic well enough to transfer to a novel case. Enhancement and sidecar have their best shot here; time budgets are generous.
• IMPLEMENT: execute a procedure from documentation (the retry-policy descendant). Diagram, table, and record-mediated views face their proving ground; errors are behavioural, not just wrong answers.
• REFERENCE: look up a specific value, exception, or condition quickly. The fixed page’s home turf — findability and latency dominate, and any mediation overhead shows immediately.

A browser excellent for IMPLEMENT but slower for REFERENCE is not a failed result. It is a routing policy in waiting: heavy mediation where structure is complex, raw source where lookup is simple. The task split converts integration findings into deployment policy, which is more useful than a trophy.

Composed measures, no new ontology

Every measure below is inherited — selection from Chapter 4, understanding from Chapter 5, enrichment accounting from Chapter 7, relation classes from Chapter 8, profile fields from Chapter 9 — plus the overhead terms composition introduces:

selection accuracy, false skips/opens, comprehension,
transfer, task time —
   Ch 04/05 instruments, unchanged

useful / irrelevant / duplicate enrichment surfaced and used —
   Ch 07 categories, per-category

contradictions found, extensions found, duplicates identified —
   Ch 08 vocabulary as outcome classes

claim / relation / qualifier preservation, unsupported additions,
emphasis shift, traceability —
   Ch 09 profile, full vector, no scalars

interaction overhead, latency, generated-element count,
manual corrections, preference —
   composition's own price tags

No new measurement ontology is introduced. That restraint is itself a test: if Part I’s instruments cannot evaluate Part I’s composition, the Part’s methodology is incomplete. Preference stays last and separate, per the standing rule — a browser loved and unfaithful is a failure the ladder must be able to show.

The fail-closed gate: UNCHANGED as system fallback

The preservation gate is policy, not just scoring: a candidate transformation whose preservation evidence is insufficient does not ship with a warning — it does not ship. The fallback path:

candidate transformation
      ↓
preservation evidence insufficient
      ↓
USE SOURCE / UNCHANGED

This closes the loop Chapter 4 opened. UNCHANGED began as a routing option for unsuitable pages, became a successful router output in Chapter 5, and is now the system-wide default-deny posture: mediation renders only what has earned permission, per transformation, per view. A browser that shows the source where its machinery is uncertain is not a browser with a gap — it is a browser with a conscience, and the gate’s refusal rate is reported as a first-class outcome (a gate that never refuses is either perfect or decorative; perfect is unlikely).

Latency gets the same honesty: the gate’s checks cost time, the record costs an extraction pass, the sidecar costs retrieval. EXP-10 prices each rung’s latency against its gains per task type, so the REFERENCE-task finding — if mediation loses there — arrives with the cost breakdown that explains it. Part I’s closing account, in three states — ESTABLISHED/PRECEDENT: dynamic representation, sometimes-recoverable relation information, dimensionally measurable preservation; BOOK ARCHITECTURE: source-linked semantic intermediate with routing, enrichment, relation readout, and fail-closed preservation composition; UNRESOLVED: strong supervised alternatives, ontology-vs-model contribution, open-world relation discovery, downstream benefit. That is the honest line the rest of the book inherits.

What this chapter earned

Nothing new, on purpose — and that is the result. The Semantic Browser composes record, router, view, sidecar, readout, and gate into one inspectable pipeline with raw-source escape, per-component attribution via the ablation ladder, task-conditioned performance across SCAN/LEARN/IMPLEMENT/REFERENCE, composed inherited measures, and a fail-closed gate that makes UNCHANGED the system fallback. Whether the composition helps is EXP-10’s empirical question, with component-level failure localisation built in rather than hoped for. Part I’s mechanisms are now one system with one evaluation programme.

Even if the browser can choose a better representation, enrich it, relate surrounding information and measure what survived, it has only solved the problem for information already in front of us. The harder problem is that most potentially useful information never reaches the page at all.

References

No new sources. This chapter composes Chapters 1–9’s evidence bases (Cao CHI 2025; Chen ACL 2026; Mayer CTML constraints; foraging/scent; visual-abstract evidence; DiagramEval; Deng proactivity survey; RELATE inheritance) under the claim-ledger states CL-01–CL-10. Any citation below repeats a Part I source for the composition argument; the Part I evidence summary records the full base. Part I’s final scientific position: bounded representation and relation precedents are SUPPORTED BY PRIOR WORK and RELATE; the record-mediated architecture is PROPOSED BY THIS BOOK; whether it beats strong supervised alternatives and improves real downstream tasks is UNTESTED.

Proposed experiment EXP-10: ablation ladder across four tasks

Status: PROPOSED. Corpus: 20–30 real pages from dense technical material (the author’s own books where rights permit, else equivalent technical sources), frozen with human ledgers for preservation scoring. Conditions A–F per the ladder above; every rung fully logged (record dumps, router traces, sidecar transcripts, readout labels, gate decisions incl. refusals). Tasks: SCAN (triage sets, keep/close vs relevance keys, false-skip primacy), LEARN (study then transfer to novel case), IMPLEMENT (execute a procedure; behavioural correctness), REFERENCE (targeted lookup; time-to-value). Measures per the composed list; profile scored full-vector on every rung’s outputs; overhead (latency, element count, corrections) per rung per task. Failure criteria: full browser (F) underperforms fixed page (A) on net task utility (integration invalidates composition — kept as the Part’s most important negative); a middle rung dominates F on some task (identifies the cost centre); gate refusal rate near zero with imperfect profiles (gate decorative); REFERENCE losses without latency attribution (measurement gap). Artifacts expected: page/ledger packs, task instruments, rung logs, per-task per-rung tables, attribution analysis. What a positive result would not justify: that every page should become generative, or that human understanding is solved — a prototype communication primitive with priced costs, nothing more.