PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Atoms for Scalable Text Analysis in Data Analytics in the Era of LLM-Based Agents

Jinxiang Xie, Zihao Li, Rui Ding, Bartuer Zhou, Shi Han, Dongmei Zhang

arXiv:2610.34860v1cs.CL

TL;DR

In agent-planned analysis, text columns require query-specified semantic processing interleaved with exact operations, while semantic outputs may need reuse across steps. TADA offers two query-driven atoms that expose reusable outputs, and on TABench it shows comparable row-level semantic quality, higher complete-artifact preservation estimates, and fewer semantic-model requests than semantic operators under evaluated native execution paths.

  • Problem

    Agent-planned TADA must interleave query-specified semantic processing with exact operations and decide which semantic outputs later steps can reuse.

    Source: §1 · p. 2

  • Method

    TADA exposes query-driven Tagging and Categorization atoms whose materialized outputs are reusable by subsequent symbolic operations.

    Source: §1 · pp. 2–3

  • Results

    TADA-Agent delivers comparable row-level semantic quality, higher complete-artifact preservation estimates in all five evaluated classes, and at least 41.8 times fewer dedicated semantic-model requests than semantic operators under evaluated native execution paths.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Reusable semantic state is the interface’s default, allowing an agent to compose semantic calls with exact operations without discovering a map-first plan among composite alternatives.

    Source: §1 · pp. 2–3

    TABench covers 40 of 43 frozen requests, and its quality measures depend on a held-out TADA reference and calibrated LLM judge; tested scales do not establish asymptotic cost or latency across providers.

    Source: §7 · p. 15

Abstract

from arXiv · show

Much of the information needed for data analysis is buried in text columns: an analyst may infer customer frustration from complaint text and use it to explain resolution time. We call workloads in which processing text columns is central to subsequent data analysis TADA (\emph{Text Analysis in Data Analytics}). In an agentic setting, the user's query dynamically specifies this processing, and answering it interleaves LLM-based semantic steps with exact SQL or Python operations. This mixed execution raises two design questions: which work should be neural and which symbolic, and which semantic outputs later steps can reuse. We address them for agent-planned TADA workflows with two complementary, query-driven semantic atoms. Each atom takes a sequence of text items and a natural-language query specifying the semantic task: \textbf{Tagging} returns one string per input item, while \textbf{Categorization} returns $M$ category objects over $N$ input items. Tagging results are materialized as a row-aligned column and Categorization results as a category table; SQL or Python then performs exact computation over this exposed state. We realize this interface in TADA-Agent, whose atoms run inside generated Python programs and internally handle inputs beyond a reliable context window. We evaluate TADA-Agent on TABench, a benchmark of 40 complete analysis requests over spreadsheet tables spanning workflow composition, input--output structure, and literature-grounded analytical scenarios. Against a semantic-operator baseline and relative to a held-out TADA reference run, TADA-Agent delivers comparable row-level semantic quality, has higher complete-artifact preservation estimates, and uses over an order of magnitude fewer semantic-model requests under the evaluated native execution paths.

1. Introduction

Agent-planned TADA interleaves query-specified semantic processing with exact operations, raising questions about where the neural–symbolic boundary belongs and which semantic outputs remain reusable. TADA-Agent addresses these questions with two query-driven atoms and evaluates their use in complete analytical workflows.

Interface design

  • TADA exposes query-driven Tagging and Categorization outputs as reusable semantic state, leaving filtering, sorting, and model fitting to symbolic code.Tagging yields row-aligned values; Categorization discovers a shared vocabulary that a subsequent Tagging call can assign to rows.Source: §1 · p. 2
  • Every semantic call returns reusable state by default, making a map-first decomposition the interface’s default rather than adding expressive power over semantic operators.Tagging is functionally a semantic map; semantic-operator systems with a map can compute the same things.Source: §1 · p. 3

Large-scale execution

  • Large-scale Categorization plans chunks and a pilot-selected local retention budget, then consolidates synonymous or overlapping candidates into a final category table.A sufficient condition relates chunk size, retention, and support margins to recovery; a theory-guided bootstrap selects the smallest budget meeting target reliability.Source: §1 · p. 3

TADA-Agent

  • TADA-Agent runs its atom APIs inside generated Python workflows, while lightweight models execute focused calls and a reasoning model plans the analysis.The skill provides API contracts, composition patterns, and examples, but the orchestrator chooses the workflow and symbolic methods.Source: §1 · p. 3

Evaluation

  • At least 41.8 times fewer dedicated semantic-model requests were used by TADA-Agent than by semantic operators under the evaluated native execution paths.The comparisons use a held-out TADA run as a common reference; class-level intervals overlap, so results measure agreement with that reference rather than absolute quality.Source: §1 · p. 3
  • TADA-Agent has higher complete-artifact preservation estimates in all five evaluated classes, while semantic operators better reproduce the reference’s absolute ranking scale.TADA-Agent and semantic operators are close on classification and extraction, and TADA-Agent performs best on constrained generation.Source: §1 · p. 3

1. An agent-era formulation and semantic-atom inter-

Agent-driven TADA centers on deciding which work is neural or symbolic and which semantic outputs remain available to later steps. The paper proposes two reusable output shapes and evaluates their implementation and workflow behavior.

An agent-era formulation

  • The formulation identifies two design decisions: which work an LLM performs versus exact code, and which LLM outputs remain available to later steps.The two recurring reusable output shapes are row-aligned values and corpus-level vocabularies.Source: §1 · pp. 3–4

An agent-era formulation and semantic-atom interface

  • The proposed atoms expose row-aligned values and corpus-level vocabularies through complementary N →N Tagging and N →M Categorization contracts.The contribution is making reusable state the default output of every semantic call, not adding expressive power over semantic operators.Source: §1 · p. 4

TADA-Agent and evaluation

  • TADA-Agent invokes atoms from generated Python and preserves their public contracts across one-call and multi-call execution.TABench contains 40 complete TADA requests spanning workflow, structure, and scenario; evaluation also includes large-scale execution and public extraction benchmarks.Source: §1 · p. 4

2. Related Work

Related work spans semantic operators, analytical agents and pipeline systems, and methods for stable large-scale text processing. TADA’s focus is the output contract for semantic calls in agent-written analytical workflows.

2.1. Semantic Operators and AI Functions

  • TADA differs from semantic-operator systems by making row-aligned columns or category tables the uniform primary outputs, while retaining exact operations in the symbolic runtime.Semantic operators and AI functions already separate model-backed calls from relational execution; TADA’s Tagging is functionally a semantic map, not an increase in expressive power.Source: §2.1 · p. 4

2.2. Analytical Agents and Semantic Pipeline Systems

  • TADA studies what a semantic call should return when an agent-written plan reuses its result in later analytical steps.TABench complements SemBench with complete natural-language analysis requests that require an agent to choose a decomposition and deliver output files; it uses a reference run and LLM judge because many requests lack a unique answer.Source: §2.2 · p. 5

2.3. Stable and Large-Scale Semantic Processing

  • TADA focuses on preserving atom outputs across calls through row-identity reassembly, semantic consolidation of category proposals, and discover-then-assign joint Tagging.Its controlled Categorization analysis fixes canonical assignments to isolate candidate retention, while its end-to-end scale experiment evaluates fixed-taxonomy Tagging.Source: §2.3 · p. 5

3. Two Semantic Atoms and Their Composition

The section defines two query-driven atom contracts and shows how their materialized outputs support workflows composed with symbolic operations. Together, the contracts support familiar semantic operators as well as broader multi-step analytical tasks.

3.1. Tagging and Categorization

  • Tagging maps each input text to one output string and preserves input length and order, with results materialized as a row-aligned column keyed by row identifiers.Its query can specify classification, extraction, or scoring; multiple requested output columns use separate Tagging calls.Source: §3.1 · p. 5
  • Categorization returns category objects with required names and optional descriptions, but row assignment requires a subsequent Tagging call using the discovered categories.The category objects can be materialized as a separate table with fresh category identifiers.Source: §3.1 · pp. 5–6

3.2. From Semantic Operators to Analytical Workflows

  • The atom outputs compose into varied analytical workflows, including loops, hierarchical discovery, temporal comparison, constrained scoring, and regression over semantically constructed features.Table 1 spans nine scenario families and output structures from derived columns and category lists to aggregates, ordered rows, and fitted models; the agent can build workflows without a separate model-backed operator for each analytical verb.Source: §3.2 · p. 6, Table 1 · p. 7

4. TADA-Agent Design and Implementation

TADA-Agent places Tagging and Categorization inside Python workflows, pairing semantic processing with symbolic data operations. Its design uses in-workflow tool calls, separate model roles, and a skill that guides workflow construction.

4. TADA-Agent Design and Implementation

  • TADA-Agent makes Tagging and Categorization available inside its Python execution environment, where generated workflows can compose them with symbolic code.Its design centers on Programmatic Tool Calling, separate model roles, and a text-analysis skill.Source: §4 · p. 7

4.1. Programmatic Tool Calling for Text Analysis

  • Programmatic Tool Calling keeps group-wise semantic loops and intermediate tables inside executed code, returning only the final mapping rather than repeatedly crossing the model–tool boundary.This avoids the repeated round trips and risks of omitting groups or losing group-result correspondence described for Traditional Tool Calling.Source: Figure 2 · p. 8, §4.1 · p. 8

4.2. Separate Model Roles

  • The orchestrator plans and writes the analysis workflow, while narrower atom calls validate structured responses and can use lightweight models for high-volume processing.Atom calls combine a fixed system prompt with a query and text column, returning one of the two prescribed output shapes.Source: §4.2 · p. 8

4.3. Skill-Based Implementation

  • The text-analysis skill supplies API signatures, conventions, composition rules, and examples, while the library handles batching, parallel calls, validation, and recovery behind unchanged public contracts.The orchestrator still decides how to compose the APIs with symbolic code.Source: §4.3 · p. 8

5. Large-Scale TADA Execution

For large text inputs, TADA partitions semantic work into chunks and consolidates local results to recover global categories. Its planning balances chunk size, local candidate retention, and consolidation, while Tagging supports independent and shared-domain requests.

5. Large-Scale TADA Execution

  • TADA uses MapReduce-style semantic execution because independently categorizing chunks can miss globally important topics, produce synonymous labels, or vary in granularity.It retains local category candidates and semantically consolidates proposals into a global result; partitioning, parallel calls, and recovery remain inside the atom implementations.Source: §5 · p. 8

5.1. Planning a Large-Scale Request

  • Chunk size n and local retention budget r are coupled: smaller chunks need a larger buffer, while increasing r improves preservation but raises output and consolidation costs.A clear frequency gap after rank k can permit r = k; nearly tied categories around rank k call for retaining more candidates.Source: §5.1 · p. 9

5.2. Top-k Preservation

  • The top-k preservation bound makes the trade-off explicit: larger chunks reduce sampling variation, while larger r widens the frequency gap at additional consolidation cost.When frequencies around rank k are nearly equal, a larger retention budget may be needed.Source: §5.2 · p. 9
  • A pilot sample supplies support estimates; paired bootstrap propagates their uncertainty, and the planner selects the smallest tested r that meets the target recovery.In TADA-Agent, the batch size is set to n = 1,000.Source: §5.2 · p. 9

5.3. Large-Scale Categorization

  • Large-scale Categorization runs local top-r requests in parallel, then consolidates synonymous names and descriptions and uses pilot support estimates to select the global k categories.The input rows are shuffled and divided into balanced chunks before local categorization.Source: §5.3 · pp. 9–10

5.4. Large-Scale Tagging

  • Independent Tagging processes chunks directly and reassembles row-aligned results; Joint Tagging first discovers a shared output domain through Categorization, then assigns rows to it.A shared domain cannot be determined from any single row alone.Source: §5.4 · p. 10

6. Evaluation

Across TABench, TADA-Agent combines competitive row-level judgments with higher complete-artifact preservation estimates and far fewer semantic-model requests than semantic operators under the evaluated paths. Large-scale tests and public extraction benchmarks further examine the interface’s execution and transfer beyond TABench.

6.1. RQ1: Quality, Artifact Reliability, and Request Cost

  • The evaluation uses 40 complete spreadsheet-table analysis requests and separates row-level quality, artifact reliability, and semantic-model request cost.The reference-based comparisons use a held-out TADA output; seven rule-based cases instead use absolute accuracy and are excluded from the two comparison tables.Source: §6.1 · p. 10
  • On constrained generation, TADA-Agent scores 89.1 versus 78.6; on ranking, its Spearman correlation is 0.801 versus 0.752, while semantic operators better match the reference’s absolute scale.Overlapping intervals are not interpreted as significant differences, and the ranking reference is itself a TADA run.Source: §6.1 · p. 11
  • TADA-Agent has higher complete-artifact preservation estimates than semantic operators in all five TABench classes, while retaining competitive row-level quality and using at least 41.8 times fewer semantic-model requests.The request comparison is implementation-specific and measures physical request fan-out; it does not imply every operator interface must execute per row.Source: §6.1 · pp. 11–12

6.2. RQ2: Large-Scale Execution

  • For full-scale workloads, TADA-Agent has the highest observed accuracy on all three: 0.973, 0.942, and 0.837, with normalized accuracy at or above 1.0.LOTUS results are measured subsets rather than matched full-workload repetitions.Source: §6.2 · p. 13
  • 33 semantic requests serve each 32K-row TADA run, versus LOTUS’s 32,179 per-row calls, or 975× TADA’s planned request count.One sentiment run encountered transient provider failures, increasing observed events and wall time without changing its planned 33-request computation.Source: §6.2 · p. 13
  • When Claude Code must discover a vocabulary before assigning rows, it reaches 0.68 ± 0.01 accuracy on 32K topic data and 0.21 ± 0.08 on enterprise feedback.Each run yields 3–8 categories without a reference counterpart, mainly from different category granularity or catch-all buckets.Source: §6.2 · p. 13
  • For small-gap category distributions, fixed r = 5 recovery falls from 90.9% to 73.9%, while the pilot rule selects r = 10 and achieves 98.5–100% recovery.For the separated distribution, every tested r recovers the top-5 and the planner selects r = 5.Source: §6.2 · p. 13, Figure 3 · p. 14

6.3. RQ3: Transfer Beyond TABench

  • The unchanged Tagging interface yields the highest observed CUAD score and higher BioDEX point estimates under reproduced protocols, with uneven baseline repetition.CUAD average F1 is 0.787 for TADA-Agent, 0.753 for DocETL, and 0.619 for LOTUS; BioDEX point estimates are 0.230 RP@5, 0.246 RP@10, and 0.385 term recall for TADA-Agent.Source: §6.3 · p. 14, Figure 4 · p. 14

7. Discussion, Limitations, and Conclusion

TADA’s reusable semantic state supports multi-step analyses, while composite operators may suit one-shot requests. The evaluation reports competitive semantic quality and better workflow-level artifact preservation, within several stated scope limits.

7. Discussion, Limitations, and Conclusion

  • Reusable semantic state suits workflows that later reuse the same variable, whereas composite operators may be preferable for one-shot artifacts or specialized algorithms.TADA illustrates how domain-specific agents can expose expert-defined semantic contracts while retaining query-driven workflow composition.Source: §7 · pp. 14–15
  • TADA has higher observed complete-artifact preservation, satisfies deterministic output requirements, and invokes the semantic model fewer times, while retaining competitive semantic quality.The ranking intervals overlap and artifact samples are small.Source: §7 · p. 15
  • The two atoms are not complete contracts for pairwise joins, global matching, graphs, or comparison-based ranking; richer outputs may need explicit schemas.The proposed criterion for adding atoms is a recurring reusable output structure, not each new analytical verb.Source: §7 · p. 15, §A.3 · p. 18
  • Large-scale estimates isolate candidate retention under exact assignments, counts, and consolidation; tested scales establish neither asymptotic cost nor cross-provider latency.Request reductions compare evaluated native implementations, not total systems cost, and do not show that semantic operators cannot batch or materialize outputs.Source: §7 · p. 15
  • Evaluation covers 40 of TABench’s 43 requests; it depends on a fixed reference and calibrated LLM judge, and does not isolate the interface in a matched ablation.End-to-end large-scale Categorization with open category discovery is not evaluated.Source: §7 · p. 15

A.1. What is atomic in TADA?

  • Tagging and Categorization expose stable output structures, but scale transparency does not guarantee identical semantic values across one-call and multi-call execution.Atoms may use internal multiple calls and symbolic processing; their public contracts remain unchanged as input size varies.Source: §A.1 · p. 18, §A.4 · p. 18

A.2. Strings, types, and symbolic validation

  • Tagging’s strings can represent varied values, but the symbolic runtime must convert and validate them for downstream operations.The string contract alone does not guarantee validity for filtering, sorting, aggregation, or model fitting.Source: §A.2 · p. 18

A.3. Relationship to semantic operators

  • Tagging and Categorization can express workflows also available to semantic-map operators; their value is greatest when later steps need the exposed state.Composite operators may remain preferable when only the final artifact is needed or specialized pairwise or retrieval algorithms apply.Source: §A.3 · p. 18

B. Atom and Agent Implementation Materials

The appendix specifies atom prompting, runtime integration, and scale-aware execution, then formalizes candidate retention and its operational selection for large-scale Categorization.

B.1. Prompting and model roles

  • The orchestrator plans workflows while atom models perform narrow, high-volume semantic calls; reproducing a run requires recording both model identifiers.Atom prompts combine a dedicated system prompt, atom-level query, and selected text values.Source: §B.1 · p. 18

B.2. Programmatic Tool Calling and the text-analysis skill

  • The text-analysis skill teaches the orchestrator API signatures, input/output conventions, composition rules, and examples for invoking both atoms inside generated Python.Dataframe slices, loop variables, outputs, and exceptions remain in the same execution environment.Source: §B.2 · p. 19

B.3. Scale-transparent execution and recovery

  • Scale-transparent execution handles chunking, validation, recovery, and consolidation inside the library while preserving the public atom contracts.The implementation records run parameters and treats missing rows, malformed outputs, and failed consolidation as execution outcomes.Source: §B.3 · p. 19

C.1. Fixed-taxonomy preservation analysis

  • Under fixed labels, exact chunk counts and local top-r lists, and canonical-count reduction, retaining every full-input top-k category guarantees top-k recovery.The probability bound concerns random partitions and includes ties in its failure event; when r = C, recovery is exact.Source: §C.1 · pp. 19–20

C.2. Theory-guided operational bootstrap

  • Jeffreys-posterior support profiles and balanced-partition simulations guide selection of the smallest r whose lower empirical recovery quantile meets the target.Monte Carlo error is recorded separately, not presented as a deployment guarantee; the sufficient bound is treated as a diagnostic.Source: §C.2 · p. 20

C.3. Scope of the analysis

The retention analysis isolates candidate preservation under fixed-taxonomy assumptions and distinguishes that layer from live Categorization's additional sources of error.

C.3. Scope of the analysis

  • The analysis assumes one canonical label per row and exact consolidation, so its retention results isolate candidate loss rather than all live Categorization failures.Live execution also involves semantic discovery, under-filled lists, synonymous names, and consolidation error.Source: §C.3 · p. 21
  • The bootstrap treats the observed pilot category list as the available tail and pilot Tagging outputs as canonical, leaving unseen categories and taxonomy drift outside its estimate.Tagging error and differences between count-ranked local top-r and LLM-ranked proposals are also reported as limitations.Source: §C.3 · p. 21
  • Evaluation separates known-label retention, fixed semantic assignment, and end-to-end open discovery instead of attributing every failure to n or r.Source: §C.3 · p. 21

D. Neural–Symbolic Boundary Diagnostic

The diagnostic contrasts semantic processing combined with exact code against monolithic LLM execution, while the appendix reports public-benchmark results, TABench construction, and evaluation details.

D. Neural–Symbolic Boundary Diagnostic

  • Tagging plus symbolic code is exact across tested sizes, while monolithic sorting is incomplete, filtering degrades with scale, and aggregation is numerically incorrect.The single-call diagnostic uses templated inputs and raw prompting to illustrate a mechanism, not estimate comparative system performance.Source: D. Neural–Symbolic Boundary Diagnostic · p. 21

E.1. CUAD

  • 0.787 average F1 for TADA-Agent compares with 0.753 for DocETL and 0.619 for LOTUS under the common CUAD evaluation environment.Source: §E.1 · p. 21

E.2. BioDEX

  • TADA's three saved BioDEX runs report mean (SD) 0.2295 (0.0009) RP@5, 0.2465 (0.0095) RP@10, and 0.3849 (0.0104) term recall.Protocol-matched DocETL and LOTUS points are (0.201, 0.196, 0.305) and (0.194, 0.200, 0.334), respectively, in RP@5, RP@10, and term recall order.Source: §E.2 · p. 21

F.1. Taxonomy-derived construction procedure

  • TABench's evaluated 40 cases are a coverage-driven diagnostic sample, with three seed cases excluded because their output structures fall outside the two-atom contract.Its taxonomy divides broad intents into 12 scenario classes; the sample is not claimed to represent real-world request frequencies.Source: §F.1 · p. 22

F.2. Task inventory

  • The 40 scored requests come from 43 workbook-request pairs and cover 10 of 12 scenario classes; every covered class has at least three cases.Source: §F.2 · p. 22

F.3. Contract checks and rule-based answers

  • 100% on all seven rule-based repair cases is TADA-Agent's score; the semantic-operator baseline reaches 100% on five and scores 99.9±0.1 and 98.2±3.2 on two.On the single row-level case, all five methods fall between 99.8 and 100.0 with fully overlapping intervals.Source: §F.3 · p. 22

F.4. Semantic-operator execution and request ledger

  • At least 18,852 semantic-operator requests is a strict lower bound, with failed runs and recovery attempts lacking persistent usage excluded.Full token totals are unavailable because per-request usage was not persistently recorded; wall time is not used because endpoint and substrate latency confound it.Source: §F.4 · pp. 22–23

F.5. Judge protocol and calibration

  • 78.8% of the 33 calibration units with a human-majority verdict matched the judge, with Cohen's κ = 0.68 across six annotators.Annotator agreement was itself moderate: mean pairwise raw agreement was 0.55 and Fleiss κ = 0.32.Source: §F.5 · p. 23

F.6. Error-analysis procedure

The appendix classifies recurring evaluation failures as diagnostic, and documents recorded-run telemetry for the large-scale comparison.

F.6. Error-analysis procedure

  • Recurring failure modes include semantic substitutions, incoherent categories, incomplete row coverage, incorrect post-processing, missing workbook mutations, and tool failures.These categories are diagnostic rather than additional scoring rules.Source: §F.6 · p. 23
  • Table 9 describes judge calibration across 39 stratified judgment units and six annotators, including majority, agreement, and directional-conflict counts.Source: Table 9 · p. 24

G. Large-Scale End-to-End Execution Details

G.1. Per-repetition cost telemetry

The appendix reports per-repetition costs using system-specific telemetry and supplements observed runs with full-scale estimates and protocols for coding-agent repetitions, open-vocabulary mapping, and controlled retention.

G.1. Per-repetition cost telemetry

  • The cost table keeps metering distinct: TADA-Agent and LOTUS use provider-reported physical tokens, while Claude Code counts come from session transcripts.TADA-Agent sentiment repetition 1 had 105 observed request events instead of 33 planned calls, lost its token report, and took 164.6 minutes; other table exceptions include incomplete or non-aggregable telemetry.Source: §G.1 · p. 24

G.2. Full-scale cost estimates for the per-row baseline

  • 32,179 calls are exact by definition for one full LOTUS repetition, with estimated costs of 4.0M sentiment tokens, 6.1M topic tokens, and roughly 36–42 hours.Token and time figures extrapolate measured per-row costs; they are not measurements.Source: §G.2 · p. 24

G.3. Coding-agent repetition artifacts

  • 22–155 assistant API messages across completed coding-agent repetitions reflect widely varying execution, with heavier runs involving script debugging and intermediate-file rereads.A topic repetition exceeded a three-hour budget without output and was replaced by a completed rerun under the same protocol.Source: §G.3 · p. 24

G.4. Open-vocabulary mapping protocol

  • Open-vocabulary scoring maps each discovered label to its closest reference category or none, counting unmapped rows as errors against the full denominator.Unmapped labels include granularity merges and invented catch-all buckets.Source: §G.4 · p. 25

H.1. Controlled distributions

  • The controlled retention experiment uses 32,000 fixed canonical assignments over 20 categories, split into 32 chunks of 1,000, and requests the global top five.A 1,000-row pilot informs selection among local budgets r = 5–10; absent a certified candidate, the planner falls back to r = 10.Source: §H.1 · pp. 25–26

H.2. Connection to the preservation analysis

The planner uses paired-bootstrap estimates to select the smallest local retention budget meeting a reliability target. Controlled fixed-assignment tests characterize recovery and clarify the analysis’s scope.

H.2. Connection to the preservation analysis

  • 98.5% to 100% recommended-point recovery accompanies a planner recommendation of r = 10 across five narrow levels, exact in the two hardest and conservative by one in the other three.The planner never selects too small a budget in these cases.Source: §H.2 · p. 26
  • 98.5% to 100% empirical recovery for the hardest narrow level costs doubling candidate volume when r rises from 10 to 20.The selected budget is the minimum reliable one, not the unconstrained quality-maximizing value.Source: §H.2 · p. 27
  • At r = 8, near-complete representation of competing ranks can still let the true rank-five category lose its small margin; at r = 9–10, all four fractions exceed 99%.The minimum reliable r is the first whose recovery and every larger tested r remain above target.Source: §H.2 · p. 27
  • The plug-in bound is vacuous at both selected operating points, so the bootstrap selects the budget; the bound identifies a rank-sensitive margin and gives a distribution-free certificate when nonvacuous.It also proves exact recovery at r = C and separates candidate-retention error from semantic assignment and consolidation error.Source: §H.2 · pp. 26–27

H.3. Scope and evidence

  • The controlled experiment measures candidate retention under fixed canonical assignments, complementing the paper’s fixed-taxonomy end-to-end Tagging evaluation.Its protocol, population-generating distributions, and aggregate recovery results are reported in this section.Source: §H.3 · p. 27
Loading 2610.34860v1…