PaperlayerView exact revision on arXiv

Source-linked AI summary

Sentence Specificity Scores for Collaborative Technical Documentation: A Domain-Transfer Study

Rocker D'Antonio, Thomas Benton Townsend, Dimitrios Michael Manias

arXiv:2610.01046v1cs.CLcs.LG

TL;DR

Sentence-specificity scores may reflect different textual evidence across general and technical domains, leaving their interpretation and decision value open questions. The study audits two predictors across documentation corpora and tests ranking on fixed generated revisions. Score ordering and decision gains vary by predictor and candidate set.

  • Problem

    Sentence-specificity scores may reflect different textual evidence in technical and general prose, leaving their interpretation and decision value in technical documentation open to evaluation.

    Source: §I · p. 1

  • Method

    The study audits cross-domain score behavior and tests post-generation ranking on fixed, score-blind revisions against choosing the first candidate.

    Source: §I · pp. 1–2

  • Results

    Corpus order and same-sentence rankings vary by predictor; SpeciTeller’s direction-valid selection gains were +11.7 points for Gemma and +5.0 points for GPT-OSS-120B, whose primary single-score intervals all include zero.

    Source: §VIII · p. 9

  • Takeaways & Limitations

    Scores should be characterized within the document family and tested on the decision they will inform.

    Source: §VIII · p. 10

    Evaluation covers fixed-candidate direction validity, not author adoption, contextual checks, or downstream documentation and collaboration outcomes.

    Source: D. Three-Stage Evidence and… · p. 9

Abstract

from arXiv · show

Collaboration depends on shared context, and technical documentation is one way that context persists across people and AI teammates. Specificity, the amount and exactness of detail expressed in language, shapes what information documentation captures and how precisely that information is communicated. This work audits sentence-specificity scoring artifacts on technical documentation and tests whether scores applied only after generation help choose among fixed LLM-generated revisions. Across Wikipedia and three technical-documentation corpora, the fixed general-domain predictor SpeciTeller and the pinned post-publication author-repository implementation of Ko et al.'s target-adapted predictor produce different corpus orders and same-sentence rank agreement from -0.066 to 0.510. Strict filtering and token-length adjustment change these patterns without reconciling them. In the Gemma set, SpeciTeller ranking raises direction-valid selection from 71.7% to 83.3% (+11.7 points; 95% source-case bootstrap interval +1.7 to +21.7); in the GPT-OSS-120B set, SpeciTeller ranking raises direction-valid selection from 51.7% to 56.7% (+5.0 points; 95% source-case bootstrap interval -6.7 to +16.7), and every primary single-score GPT-OSS-120B interval includes zero. These findings tie score interpretation and decision value to the predictor and candidate set.

I. INTRODUCTION

Technical documentation preserves context across people and AI teammates, but specificity scores may reflect different evidence in technical and general prose. The study audits score interpretation, local correspondence, and bounded post-generation selection, finding that both score behavior and decision value depend on the predictor and candidate set.

I. INTRODUCTION

  • Technical documentation gives shared expertise a durable form, preserving procedural and operational context beyond an immediate exchange,,.Source: §I · p. 1
  • Sentence-level specificity scores may reflect different textual evidence in technical and general prose, so identical scores need not represent the same information.Source: §I · p. 1
  • The audit separates score interpretation, local correspondence, and value for a named decision rather than treating one form of evidence as a substitute for another.Source: §I · pp. 1–2
  • SpeciTeller is applied without adaptation, while the pinned post-publication author-repository predictor performs its documented target adaptation; no new transfer-learning model is trained.Source: §I · p. 2
  • 83.3% versus 71.7% for first-candidate selection (+11.7 points) in Gemma; GPT-OSS-120B is 56.7% versus 51.7% (+5.0 points), with every primary single-score interval including zero.Source: §I · p. 2

A. Specificity and Scalar Sentence Modeling

B. Domain Shift and Predictive Stability

General-domain and technical texts differ in style and structure, motivating evaluation of how specificity predictors behave across domains. The study compares corpora, applies consistent preprocessing, and tests whether filtering changes the resulting patterns.

B. Domain Shift and Predictive Stability

  • Domain-shift research documents model variation across domains and calibration contexts; this study distinguishes contextual score interpretation from usefulness in a specified decision,,.Source: B. Domain Shift and Predictive Stability · p. 2

C. Technical and Domain-Specific Text

  • Procedural texts often use ordered steps, conditional branches, and goal-oriented structure, motivating domain-aware specificity analysis rather than treating one signal as documentation quality,.Source: C. Technical and Domain-Specific Text · p. 3

A. Domain Selection

B. Corpora

  • The corpora include Wikipedia, GitHub documentation, Ansible documentation, and Python 3.12 reference material, representing expository, procedural, configuration-oriented, and definitional text.Source: B. Corpora · p. 3

C. Preprocessing

  • Preprocessing removes executable code before segmentation but retains inline code and technical identifiers; a strict sensitivity subset removes markup, URLs, command- or path-dominated strings, fragments, short strings, and low-alphabetic-content strings.Source: C. Preprocessing · p. 3

D. Corpus Statistics

The evaluation distinguishes corpus-level score patterns, sentence-level correspondence, and candidate-selection value. It compares a fixed general-domain predictor with a pinned post-publication target-adapted comparator and reports a reproduction check that differs from published results.

D. Corpus Statistics

  • Corpus statistics are reported after preprocessing; descriptive measures use sentences, while bootstrap uncertainty resamples documents, and corpus sentence counts are not size-matched.Source: D. Corpus Statistics · p. 3

A. Questions and Tests

  • Interpretation tests compare distributions, corpus orders, and same-sentence ranks before and after filtering and length adjustment; correspondence tests associations and directional behavior rather than accuracy.Decision value tests whether ranking fixed, score-blind alternatives selects a revision matching the requested specificity direction more often than choosing the first candidate.Source: A. Questions and Tests · p. 4

B. Predictors and Comparator Identity

  • SpeciTeller is used without parameter changes, while the unavailable 2019 checkpoint leads the study to treat a pinned later author-repository implementation as a separate comparator.Because its Python augmentation cannot be seeded, target adaptation is repeated three times per corpus; the repetitions are not different models.Source: B. Predictors and Comparator Identity · p. 4
  • Spearman’s ρ ranged from .262 to .614 versus published .676–.750, Kendall’s τ from .184 to .444 versus .487–.555, and MAE from .139 to .164 versus .107–.114 across three domains.The post-publication objective moved mean/standard-deviation distribution regularization from the labeled source batch to the unlabeled target batch.Source: B. Predictors and Comparator Identity · p. 4
  • GranuScore measures semantic and referential granularity rather than specificity, and its native higher-is-coarser direction makes agreement with higher-is-more-specific scores appear as negative correlation.Source: B. Predictors and Comparator Identity · p. 4

C. Human-Reviewed Candidate Selection

The candidate-selection evaluation fixes score-blind alternatives before ranking and tests whether post-generation policies select revisions that match requested directions under blinded review. Separate analyses characterize score correspondence, preprocessing sensitivity, and length-adjusted corpus differences.

C. Human-Reviewed Candidate Selection

  • After three score-blind alternatives are fixed for each source request, policies select by score change and blinded review checks whether the selected revision moved in the requested direction.The possible workflow adds factual and semantic checks before ranking and leaves the final decision to an author; this extension is proposed, not evaluated.Source: C. Human-Reviewed Candidate Selection · pp. 4–5, Figure 2 · p. 5
  • Each candidate set contains 180 ratings from one blinded evaluator, while each policy selects once for each of 60 sources; paired source-level bootstrap samples estimate gain over the first-candidate baseline.Candidate scores are computed only after alternatives are finalized; ties select the earliest candidate, and candidate-set comparisons pair the same sources.Source: C. Human-Reviewed Candidate Selection · pp. 4–5

D. Cross-Measure Comparison

  • Same-sentence Spearman correlations compare rankings within each corpus, while raw scores remain on their native scales; document-cluster bootstrapping preserves sentences grouped within documents.The analysis also reports native-scale quantiles and effective range, defined as (p95 − p05)/(U − L), plus boundary concentration.Source: D. Cross-Measure Comparison · p. 5

E. Preprocessing Sensitivity

  • The strict-filter sensitivity analysis compares existing scored rows with a uniformly filtered subset, records removal reasons, and checks that retained and excluded counts reconstruct the original manifest.Paired document-cluster bootstrap intervals describe sensitivity to document composition.Source: E. Preprocessing Sensitivity · p. 5

F. Length Standardization

  • Length standardization reweights corpora to a shared token-count distribution using counts supported by at least 100 rows from 20 documents per corpus, with separate original and strict supports.Spline-adjusted corpus gaps are checked across sentence lengths; character count is not jointly controlled.Source: F. Length Standardization · pp. 5–6

G. Proxy Feature Alignment

The section examines visible cues associated with specificity scores, then tests score behavior through controlled edits, human judgments, and post-generation candidate selection. Results show that candidate-selection gains and cross-corpus rankings depend on the predictor and setting.

G. Proxy Feature Alignment

  • Length, lexical rarity, and technical-token ratio are audited as transparent cues associated with scores, not as alternative specificity measures.The analysis also decomposes technical-token cues into identifiers, paths or flags, version and numeric tokens, parameter assignments, and token-shape complexity.Source: G. Proxy Feature Alignment · p. 6

H. Controlled Edit Behavioral Probing

  • Controlled edits probe score responses across low, median, and high specificity bands, with directional consistency defined separately for increasing, reducing, and neutral edits.The 60 author-crafted pairs also receive GranuScore, whose expected native deltas run opposite the specificity-edit direction.Source: H. Controlled Edit Behavioral Probing · p. 6

I. Human Pilot Annotation

  • A 40-sentence-per-corpus pilot uses three non-expert annotators and five-point labels; score correspondence is assessed against individual and pooled labels.Matched LLM judges apply the same rubric to the same 80 pilot sentences; the non-expert ordinal labels are treated as descriptive correspondence.Source: §I · p. 6

VI. RESULTS

  • Results are presented first for decision value, followed by interpretation and then correspondence.Source: §VI · p. 6

A. Decision-Level Candidate Selection

  • In Table II’s exploratory, unadjusted intervals, only the Gemma SpeciTeller individual-score gain interval excludes zero, while all primary GPT-OSS-120B and between-set gain intervals include zero.The Gemma SpeciTeller gain concentrates in add-specific requests.Source: A. Decision-Level Candidate Selection · p. 6

B. Predictor-Dependent Corpus Patterns

  • The two specificity sources reverse Wikipedia–GitHub and Wikipedia–Ansible corpus contrasts, while technical-corpus same-sentence rank agreement ranges from negative to moderate.Python shows notable variability across comparator repetitions, and scale-use diagnostics also differ by source.Source: B. Predictor-Dependent Corpus Patterns · p. 7

C. Related Granularity Measure

The section compares a related granularity measure and tests how filtering, length adjustment, visible cues, controlled edits, and human annotation shape interpretation. These analyses reveal persistent predictor differences and qualify what the adjusted gaps and correspondence results support.

C. Related Granularity Measure

  • GranuScore retains a distinct corpus order, correlates negatively with SpeciTeller in every corpus, and changes sign in its comparator relationship in Python.Filtering, unit-only sensitivity, and exact-token standardization preserve GranuScore’s corpus ordering.Source: C. Related Granularity Measure · p. 7

D. Preprocessing Sensitivity

  • 35.6%–70.9% of rows remain under strict filtering; Wikipedia stays highest and Python lowest, while all six Wikipedia-to-technical mean and median gap changes are positive.GitHub and Ansible exchange middle positions, so the complete four-corpus ordering does not remain unchanged; filtering also changes the comparator’s pattern without reconciling the predictors.Source: D. Preprocessing Sensitivity · p. 7, Table IV · p. 8

E. Length-Controlled Comparisons

  • Exact-token standardization reduces every Wikipedia-to-technical SpeciTeller gap, but all remain positive; length explains part, not all, of the gaps, with persistence varying by token count.Overlap qualifies adjusted estimates; comparator orderings and contrasts also remain unreconciled after standardization.Source: E. Length-Controlled Comparisons · p. 7, Table V · p. 8

F. Feature Alignment Across Domains

  • Length and TF–IDF associations are stronger in Wikipedia than in technical corpora, where the technical-token ratio remains positive but is not one uniform code-like signal.Source: F. Feature Alignment Across Domains · p. 8

G. Controlled Edit Behavioral Probing

  • Across 60 author-crafted pairs, both measures respond locally; SpeciTeller moves as intended for all GitHub add- and de-specific edits and most Ansible edits.Both measures’ intervals include zero for irrelevant rewrites; one Ansible edit with a duplicated clause and a time string raises SpeciTeller from .277 to .595.Source: G. Controlled Edit Behavioral Probing · p. 8

H. Human Pilot Annotation

  • The selected pilot is descriptive: SpeciTeller has positive rank correspondence with pooled labels in both corpora, but estimates characterize this sample and rubric, not accuracy or general human validity.Annotator agreement is low to moderate and corpus-dependent; correspondence across other score sources varies by corpus, annotator, source, and judge.Source: H. Human Pilot Annotation · p. 8

VII. DISCUSSION

The discussion separates score interpretation, correspondence, and decision value, finding that results depend on the predictor, corpus, and decision. The conclusion recommends characterizing scores within the document family and decision they will inform.

A. Interpretation

  • Filtering and exact-token adjustment changed score differences but did not reconcile predictor-specific corpus orders, same-sentence rankings, scales, or cue associations.Token length explains part, but not all, of the variation; corpus means and thresholds therefore retain predictor- and corpus-specific interpretations.Source: A. Interpretation · pp. 8–9

B. Correspondence

  • Both predictors and GranuScore responded as expected to controlled edits, but SpeciTeller’s response was smaller and less consistent for Ansible additions than GitHub additions.GranuScore was related but non-interchangeable, and the selected pilot showed corpus-dependent non-expert agreement and no consistent GPT-OSS-120B correspondence advantage.Source: B. Correspondence · p. 9

C. Decision Value

  • Positive correspondence did not guarantee selection improvement: the Gemma gain supported a bounded ranking decision, while every primary single-score GPT-OSS-120B interval included zero.The study did not pre-specify a minimum practically important gain; the observed benefit was direction- and candidate-set-specific.Source: C. Decision Value · p. 9

D. Three-Stage Evidence and Context-Aware Reuse

  • The three-stage evidence structure distinguishes what a score represents, whether it behaves and corresponds as expected, and whether it improves a named decision.The evaluated decision covered fixed-candidate post-generation ordering for direction validity; factual and semantic checks, author choice, and task outcomes remain unevaluated.Source: D. Three-Stage Evidence and… · p. 9

VIII. CONCLUSION

  • SpeciTeller gained +11.7 points for Gemma and +5.0 points for GPT-OSS-120B, whose primary single-score intervals all included zero; corpus ordering also remained predictor-specific.The target-adapted comparator reversed the Wikipedia–GitHub and Wikipedia–Ansible contrasts on original rows, while SpeciTeller retained positive standardized Wikipedia-to-technical gaps.Source: §VIII · pp. 9–10
  • The authors identify technical-documentation-specific training, calibration, and multidimensional representations as future directions, alongside larger expert review and executable tasks.Pilot correlations from Qwen and direction-aligned GranuScore were preliminary and corpus-dependent.Source: D. Three-Stage Evidence and… · p. 9, §VIII · p. 10

IX. REPRODUCIBILITY

The reproducibility materials specify how to reconstruct corpora and rerun analyses, while the authors bound interpretation to the studied predictors, corpora, and decision setting. Ethical review was exempt, and the authors report their use of AI assistance.

IX. REPRODUCIBILITY

  • The repository supplies reconstruction and rerun commands, fixed configurations, manifests, schemas, checksums, and validation checks; experimental work used specified Spark and workstation hardware.The README gives stage-level entry commands, and appendices reproduce generation and rubric prompts.Source: §IX · p. 10
  • Without sentence-level gold labels, scores remain uncalibrated to a shared scale, and comparisons characterize relative behavior rather than comparative accuracy.Coverage includes one general English corpus, three English technical-documentation corpora, and one pinned post-publication target-adapted comparator—not the paper-era 2019 checkpoint.Source: LIMITATIONS · p. 10
  • Decision evaluation used one evaluator, two review sessions, two corpora, two generators, and fixed source cases; candidate-set and session effects are inseparable.The pilot also used selected sentences, three non-expert annotators, and one score per judge and sentence; author choices and task outcomes were not evaluated.Source: LIMITATIONS · p. 10
  • Proxy features are visible lexical and token-shape cues, not semantic or model-internal explanations; length adjustment also omits other differences and leaves confounding.GranuScore is a construct-sensitive comparison, sampled edits retain formatting remnants, strict filtering changes samples unequally, and Python has the weakest common support.Source: LIMITATIONS · p. 10

ETHICAL CONSIDERATIONS

  • The study qualified as exempt from formal IRB review, and the authors state that they reviewed and verified AI-assisted work and assume responsibility.Adult annotators participated voluntarily, consented, and received nominal nonmonetary compensation.Source: ETHICAL CONSIDERATIONS · p. 10, AI DISCLOSURE · p. 10

APPENDIX A ILLUSTRATIVE CONTROLLED EDIT EXAMPLES

Controlled edits probe specificity scores against intended changes, while pilot correlations compare predictors, granularity, and reader judgments. Results vary across corpora and score sources.

APPENDIX A ILLUSTRATIVE CONTROLLED EDIT EXAMPLES

  • GranuScore moved in its expected native direction: adding specificity lowered its mean by 6.43 in Ansible and 2.52 in GitHub, while de-specifying raised it by 3.64 and 5.38, respectively.Source: Appendix A · p. 13

A. Add-Specific Example from Ansible

  • For an example showing host uptime in days/hours/minutes/seconds (assuming facts were gathered), the edited version adds “02d 23h 58m 59s” and raises the score from 0.277 to 0.595.The score change is +0.318.Source: A. Add-Specific Example from Ansible · p. 13

B. De-Specify Example from Ansible

  • The edited instruction’s score falls from 0.960 to 0.127 (∆ = −0.833), removing the specific metadata-file path and linked verification reference while retaining a generic metadata-directory instruction.Source: B. De-Specify Example from Ansible · p. 13

C. Irrelevant Rewrite Example from Ansible

  • The irrelevant rewrite reorders “different variables” and “a certain order” without materially changing the sentence’s content or score.The displayed score changes from 0.016 to 0.013 (∆= −0.003).Source: C. Irrelevant Rewrite Example from… · p. 13

A. Pilot Annotation Instructions

  • Pilot annotators used a five-point specificity scale, treating concrete identifiers and operational details as cues for greater specificity.Instructions contrasted “Restart the service” with a command naming nginx, systemctl, and host web-03.Source: A. Pilot Annotation Instructions · p. 13

B. Unified Pilot Correlations

  • On selected rows, Qwen–SpeciTeller rank agreement exceeded Qwen–comparator agreement in both corpora, while rubric-judge correspondence varied across annotators.The ST-minus-comparator pooled contrasts were near zero in Ansible and positive in GitHub; all eight Qwen-minus-specificity-source intervals included zero.Source: B. Unified Pilot Correlations · p. 13

C. Binned Correspondence

Appendix analyses examine binned correspondence, surface proxies, normalization, granularity sensitivity, and score-distribution diagnostics. The results reveal predictor- and corpus-specific patterns.

C. Binned Correspondence

  • The binned correspondence analysis partitions SpeciTeller scores into within-corpus quintiles and compares them with pooled pilot labels normalized to [0, 1].The pilot uses a small sample and an ordinal scale.Source: C. Binned Correspondence · p. 13, Table VIII · p. 14

APPENDIX C ROBUSTNESS AND SURFACE PROXY CHECKS

  • Technical corpora do not share one dominant surface proxy: GitHub aligns most with technical tokens, Ansible is mixed, and Python reference text tracks token shape and version or numeric tokens.The analysis omits Wikipedia because it is the general-domain reference corpus.Source: Appendix C · p. 14, Table X · p. 14
  • Normalization preserves controlled-edit directions, while within-corpus z-score and min–max transformations preserve Spearman ranks; raw corpus mean differences remain learned-scale behavior.Directional consistency is identical under raw, z-score, and min–max views.Source: Appendix C · p. 14, Table XI · p. 14

APPENDIX D GRANUSCORE SENSITIVITY AND VERIFICATION

  • GranuScore preserves Wikipedia-minus-technical ordering under strict filtering and token standardization, although Ansible changes most when rows without referential units are excluded.No-unit rates range from 0.70% to 11.57%; the scorer matches its official examples within maximum absolute error 3.28 × 10−6, with stable row identifiers supporting reproduction.Source: Appendix D · p. 15

APPENDIX E OUTPUT-DISTRIBUTION DIAGNOSTICS

  • Across corpora, SpeciTeller effective ranges span .814–.973, while comparator upper-bound mass is at most .024%; all 18 pre-specified scale-use comparisons differ.For Python Adapt-2, lower-bound shares reach 42.26% and 83.54% at the outer 5% and 10% thresholds.Source: Appendix E · p. 15

APPENDIX F CORPUS AND COMPONENT PROVENANCE

The appendices document comparator and model provenance, candidate-selection examples, and sensitivity boundaries. They also report important differences in the evaluated candidate sets and predictor outcomes.

APPENDIX F CORPUS AND COMPONENT PROVENANCE

  • The pinned comparator uses a different objective from the 2019 experimental artifact, whose checkpoint was not released; its augmentation is unseeded.The appendix reports a released-domain diagnostic against published Twitter results and states that the pinned implementation is used as a separate comparator.Source: Appendix F · pp. 15–16
  • The 80-sentence pilot used full coverage, identical inputs and rubric across four judges, and one blind sentence per fresh conversation.The DGX archive’s accepted provenance waiver identifies missing audit records as completeness gaps, not evidence of tampering.Source: Appendix F · p. 16

A. Candidate-Selection Example

  • The candidate-selection example is illustrative, not representative: all three neutral Gemma rewrites received blinded rating 3, while SpeciTeller chose the candidate with the smallest absolute score change.The first-candidate baseline selected Candidate 1; SpeciTeller selected Candidate 3.Source: A. Candidate-Selection Example · p. 16

APPENDIX G DETAILED SENSITIVITY BOUNDARIES

  • The strict condition retains 35.6%–70.9% of rows, and controlled-edit specificity can co-vary with length, lexical realization, modality, or proposition content.Unit-only GranuScore sensitivity preserves ordering but changes Ansible most.Source: Appendix G · p. 16
  • For Gemma, SpeciTeller’s gain concentrates in add-specific cases at +33.3 points over 18 paired sources, compared with zero for 20 neutral cases.Removing each session’s first 30 ratings preserves the estimated directions with less precision; every comparator interval includes zero.Source: Appendix G · p. 16
  • Within GPT-OSS-120B, SpeciTeller’s direction-valid selection gain over the first-candidate baseline is +5.0 points [−6.7, +16.7], and every primary single-score interval includes zero.GPT-OSS-120B direction-validity is 51.1% versus Gemma’s 76.1%; separate sessions leave candidate-set and session effects inseparable.Source: Appendix G · p. 16

APPENDIX H LLM ROLES AND PROMPT TEMPLATES

The appendix separates generator, rubric-judge, and fixed-text scoring roles. It records prompt constraints and input boundaries for reproducibility.

APPENDIX H LLM ROLES AND PROMPT TEMPLATES

  • Table XVII distinguishes generators, rubric judges, and fixed-text scoring; rubric-judge results concern the 80-sentence pilot, not generated candidate sets.A model name appearing in both roles does not imply self-judging.Source: Appendix H · p. 17, Table XVII · p. 17

A. Exact Generator Prompt G

  • Generator prompts request one grammatical rewrite preserving meaning, factual commitments, polarity, modality, and communicative intent while making only the requested change.They prohibit assuming hidden context or inventing unsupported facts.Source: A. Exact Generator Prompt G · p. 17
  • The add-specific and de-specify templates request more or less specificity while preserving meaning and factual correctness; the neutral template requests approximately unchanged specificity.Each template supplies the original sentence as its input.Source: A. Exact Generator Prompt G · p. 17
  • Generator requests exclude corpus identity, identifiers, prior outputs, proxy values, predictor scores, rubric scores, human labels, and review information.The stated blindness rule also excludes prior attempts and author edits.Source: A. Exact Generator Prompt G · p. 17

B. Exact Rubric-Judge Prompt R

  • All four rubric judges use the same fixed five-point scale, defining specificity through concrete details such as commands, parameters, identifiers, and operational information.They judge only the supplied sentence without assuming surrounding context.Source: B. Exact Rubric-Judge Prompt R · p. 17
  • Each judge call contains one pilot sentence in a fresh conversation and allows one attempt; judge outputs do not score, select, or evaluate generated candidates.The response schema restricts the score to an integer from 1 to 5.Source: B. Exact Rubric-Judge Prompt R · p. 17

C. Role Boundary

  • Candidate generation is score-blind, and predictor and GranuScore values are computed only after candidates are fixed.Separate runs and inputs prevent models serving both roles from self-judging or self-selecting.Source: C. Role Boundary · p. 17

Related papers

Loading 2610.01046v1…