PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Individual Scientific Taste with Privileged Hindsight Distillation

Hanwei Fan, Yanqiao Wang, Yiqi Zhu, Xuanyu Lei, Peng Li, Yang Liu

arXiv:2610.19719v1cs.AIcs.LG

TL;DR

Scientific taste varies across researchers, while existing criterion-based assessments and community signals do not directly capture how individuals translate shared assessments into preferences. The paper combines a shared evaluation space, a population mapping learned with retrospective evidence, and sparse-feedback individual adaptation. It reports that researcher-specific preferences can conflict on the same proposal pairs, including patterns fixed criterion reweighting misses.

  • Problem

    Criterion-level evaluations and aggregated community signals do not directly capture how an individual translates shared assessments into preferences.

    Source: §1 · pp. 1–2

  • Method

    The framework adapts a population preference mapping to researchers, using Privileged Hindsight Distillation to train shared assessment and population learning with later evidence while keeping inference proposal-only.

    Source: §1 · p. 2, §3.1 · p. 3, §3.2 · p. 4

  • Results

    Researchers show sharply different preferences on the same proposal pairs, and conditional mappings recover conflicting preferences that fixed criterion reweighting misses.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    A population mapping can provide a transferable prior for individual adaptation, while researcher-specific conditional mappings capture preferences beyond fixed criterion reweighting.

    Source: §1 · p. 2, §5 · p. 9

    The shared criteria are neither fully objective nor exhaustive of scientific judgment.

    Source: §1 · p. 2

Abstract

from arXiv · show

Scientific taste shapes which research directions researchers pursue, and it varies across individuals. Existing work evaluates research directions along multiple evaluation criteria or models scientific taste using community feedback signals such as citations and reviews, but neither directly captures how individuals translate assessments under the same criteria into different preferences. We model scientific taste as a researcher-specific mapping from criterion-level assessments in a shared, evidence-grounded evaluation space with proposal-visible evidence, to relative preferences. A population-level mapping learned from retrospective supervision provides a cold start when individual feedback is scarce. We introduce Privileged Hindsight Distillation (PHD), which uses later-emerging scientific evidence during training to improve proposal-time criterion assessment and the population mapping, while exposing only proposal-visible information at inference. We evaluate our approach primarily in quantum research, where structured public reviews are largely unavailable, and use pairwise proposal comparison as a direct interface for observing individual taste. With only 15 comparisons per researcher, our method captures sharply different tastes, including opposing preferences over the same research-direction pairs when they trade off across evaluation criteria. In an interactive study, personalized rankings also surface proposals associated with greater subsequent user engagement, connecting learned taste to downstream research exploration.

1 Introduction

Scientific taste shapes research choices, yet researchers may value the same criterion assessments differently, and direct feedback is often scarce. The paper models these individual preferences using shared proposal-time assessments and a population mapping learned with retrospective evidence.

1 Introduction

  • Researchers can disagree about the same work, while criterion assessments and aggregated ecosystem signals do not directly capture how an individual translates criteria into preferences.Direct feedback is also scarce.Source: §1 · p. 1
  • A shared, evidence-grounded evaluation space separates criterion assessment from researcher-specific mappings to relative preferences.Later-emerging evidence may supervise population-level learning but is unavailable when proposals must be judged.Source: §1 · p. 1, Figure 1 · p. 2
  • PHD uses later-emerging evidence during training to refine proposal-only student artifacts, improving shared evaluation and population-level taste learning while preserving proposal-only inference.The population mapping is then adapted using sparse direct feedback.Source: §1 · p. 2
  • Pairwise proposal comparisons provide a lightweight way to observe relative preferences, and the evaluation focuses primarily on quantum research with additional real- and simulated-researcher experiments.The paper reports disagreement especially when research directions trade off across criteria, and personalized rankings surface proposals users subsequently inspect, refine, or pursue.Source: §1 · p. 2
  • The framework separates criterion assessment from valuation and reports that conditional researcher-specific mappings recover conflicting preferences on the same proposal pairs where fixed criterion reweighting misses them.The authors also note that the criteria are neither fully objective nor exhaustive of scientific judgment.Source: §1 · p. 2

2 Related Work

Prior work studies scientific judgment through impact signals, proposal evaluation, privileged-information learning, and sparse preference learning. This paper focuses on learning individual scientific preferences from sparse, sequential feedback.

Modeling Scientific Taste

  • Prior scientific-taste work uses impact and ecosystem signals or studies aggregate proposal preferences, whereas this work targets researcher-specific preferences learned from sparse feedback.Related proposal studies include TASTE, taxonomy-based comparisons [21], and explicitly encoded researcher preferences [22].Source: §2 · p. 3

Research Artifact Evaluation

  • Research-artifact benchmarks assess criteria such as novelty and feasibility, while evidence-grounded methods support assessment with retrieved literature; this work organizes proposal evidence in a shared evaluation space.Prior holistic LLM judgments do not always predict executed-research quality and have limitations in novelty assessment.Source: §2 · p. 3

Privileged Information Learning

  • Privileged-information learning trains with information unavailable at deployment, and PHD applies this separation to retrospective scientific evidence and a proposal-time agentic workflow.The related work also describes distillation of privileged information into deployment-restricted policies [35] [36].Source: §2 · p. 3

Sparse Preference Learning

  • Sparse preference research covers few-shot and population-informed learning, heterogeneous preferences, disagreement preservation, and sequential feedback; this work combines individual scientific taste with sparse, sequential feedback.Related methods include structured preferences [45] [46], perspectivist learning [47] [48] [50], and replay for continual learning [51].Source: §2 · p. 3

3 Method

The framework represents proposals in a shared evaluation space, learns a population preference mapping, and adapts it to individual researchers. PHD improves the first two components using retrospective evidence during training while retaining proposal-only inference.

3 Method

  • The method has three components: shared evaluation-space learning, population taste distillation, and individual taste adaptation from sparse sequential feedback.PHD improves the first two components.Source: §3 · p. 3

3.1 Shared Evaluation Space and Scientific Taste

  • A researcher-independent criterion assessor maps proposal-visible information into shared assessments, while each researcher maps assessments and supporting evidence to relative preferences.Researchers may translate the same criterion representation into different preferences.Source: §3.1 · p. 3

3.2 Privileged Hindsight Distillation

  • PHD iteratively rolls out a proposal-only student, identifies failures, lets a hindsight-informed teacher propose artifact revisions, and accepts revisions only after leakage screening and held-out validation.At inference, only final accepted artifacts and proposal-visible information remain.Source: Figure 2 · p. 4, §3.2 · pp. 4–5

3.3 Evaluation Space Learning

  • Evaluation Space Learning uses a six-criterion shared space, with a hindsight-informed reference rubric and a proposal-only student rubric whose revisions are retained based on held-out MAE.The criteria are generality, novelty, importance, parsimony, depth, and algorithmic complexity.Source: §3.3 · p. 5

3.4 Population Taste Distillation

  • Population Taste Distillation models pairwise preferences from criterion differences and proposal-visible evidence, combining fixed criterion weights with context-dependent rules and memory.The population mapping is learned from population training pairs and serves as the prior for individual adaptation.Source: §3.4 · p. 5

3.5 Individual Taste Adaptation

  • Individual adaptation starts from the population mapping, refits researcher-specific criterion residuals from sequential feedback, and retains conditional rules only when cross-fitted replay improves prediction without breaking correct preferences.Held-out query labels and privileged hindsight are not used during individual adaptation.Source: §3.5 · pp. 5–6

4 Experiments

Across five experiments, the method is tested on sparse individual comparisons, shared-component robustness, downstream proposal discovery, privileged-hindsight ablations, and adaptation behavior. It performs best among the individual-taste methods on real and simulated researchers, with gains reported across backbones and in downstream rankings.

4.1 Learning Individual Scientific Taste

  • PHD performs best on real and simulated researchers while keeping the backbone frozen, using sparse researcher feedback for individual-taste prediction.Each of seven quantum researchers supplied 15 support comparisons; the study also included 100 simulated researchers.Source: §4.1 · p. 6, Table 1 · p. 7

4.2 Robustness and Sensitivity Analysis

  • Across all four backbones, PHD improves criterion estimation and population preference prediction, while individual results outperform PP and CIPHER; alternative population priors also yield persistent differences after identical adaptation.The prior sensitivity result indicates that sparse individual feedback does not fully overwrite the population initialization.Source: §4.2 · pp. 7–8, Table 2 · p. 7, Table 3 · p. 8

4.3 Downstream Evaluation

  • Learned individual taste improves proposal discovery as candidate pools evolve, with gains persisting under cross-domain transfer at varying magnitudes.The interactive evaluation compares personalized rankings with alternative methods using real and simulated researchers.Source: §4.3 · p. 8, Table 3 · p. 8

4.4 Role of Privileged Hindsight

  • Proposal-visible refinement improves criterion assessment but not population preference prediction; privileged hindsight adds a smaller gain to criterion metrics and a larger gain to population preference prediction.This contrast points to hindsight's greater importance for mapping criterion differences to population preferences than for criterion assessment itself.Source: §4.4 · pp. 8–9, Table 4 · p. 9

4.5 Understanding Individual Taste Learning

  • Conditional rules yield modest gains for real researchers and Sim.-Q, residual adaptation suffices for Sim.-Ctrl., and replay becomes beneficial as researcher-specific preferences emerge.Held-out RW-SP remains non-monotonic as feedback accumulates; removing PHD-derived shared structure further reduces RW-SP.Source: §4.5 · p. 9, Table 5 · p. 9

5 Conclusion

The paper concludes that a population-level judgment offers a useful starting point but does not fully capture individual scientific preferences. Its appendix details how hindsight-derived revisions are diagnosed, made student-safe, audited, and validated before acceptance.

5 Conclusion

  • PHD improves proposal-time assessment and population learning while preserving proposal-only inference, but the study is limited by sparse real-expert feedback and a finite criterion set.The conclusion also reports opposing preferences over the same proposal pairs and greater subsequent engagement for personalized rankings than the matched Single Call baseline.Source: §5 · p. 9

A.1 Stage-Specific Failure Diagnosis

  • The analyzer prioritizes large errors, evidence inconsistencies, constraint violations, and recurring failures, then packages task-specific student judgments and reference comparisons for the teacher.The analyzer identifies failures but leaves their causes and corrections to the privileged teacher.Source: §A.1 · p. 13

A.2 Privileged Teacher Editing

  • The privileged teacher uses hindsight to diagnose selected failures and proposes rubric, memory, or conditional-rule revisions without replacing student judgments with privileged answers.Revisions must remain executable using proposal-visible information; teacher-only diagnostic explanations are not passed into student-facing artifacts.Source: §A.2 · p. 14

A.3 Artifact Leakage Audit

  • The leakage auditor rejects student-facing revisions that expose retrospective outcomes or require unavailable proposal-stage features, while allowing hindsight to inform general reusable instructions.Conditional rules must use proposal-visible properties and avoid identifying a specific proposal.Source: §A.3 · p. 14

A.4 Gated Validation and Artifact Acceptance

  • A revision must pass leakage review, improve held-out validation performance, and avoid material behavioral regressions before replacing the incumbent artifacts.Rejected revisions are rolled back, and the final evaluation set remains untouched by selection and acceptance decisions.Source: §A.4 · p. 14
  • For evaluation-space assessment, acceptance requires leakage risk below 0.15, validation MAE improvement greater than 0.01, correct-to-wrong flips no higher than 5%, and positive net repair.For population taste prediction, acceptance requires improved StrictPair, a McNemar test with p<0.05, and a paired bootstrap 95% confidence interval excluding zero.Source: §A.4 · p. 15

B Evaluation Space Learning

This section specifies reference and proposal-time criterion assessment, then describes how PHD converts recurring errors into reusable student-side rubric and memory updates. The student remains limited to proposal-visible information at deployment.

B.1 Reference Criterion Annotation

  • Reference annotations score six criteria independently on an absolute 0–3 scale, grounding each score in criterion-specific evidence that may include privileged hindsight.The criteria are generality, novelty, importance, parsimony, depth, and algorithmic complexity.Source: §B.1 · pp. 15–16

B.2 Seed Proposal-Time Rubric

  • Seed Rubric B operationalizes the same six criteria for proposal-time assessment using only the proposal and literature available at evaluation time.It marks evidence gaps rather than silently choosing a stronger adjacent score, and PHD may refine its score boundaries while preserving the information boundary.Source: §B.2 · pp. 16–17

B.3 Learning Rubric B and Student-Safe Case Memory

  • PHD turns recurring assessment failures into rubric revisions or student-safe case memories, removing hindsight-dependent facts before either artifact reaches the student.Memory entries use proposal-visible triggers and reusable scoring lessons; examples include distinguishing a new workflow from a new primitive and judging parsimony by functional coherence, not component count alone.Source: §B.3 · pp. 18–19

B.4 Evaluation-Space Learning Case Study

The held-out Quantum Hamiltonian Descent case study illustrates how PHD changes proposal-time criterion scores. The learned calibration improves aggregate test-set MAE, while the individual example also shows that a transferable rule can be over-applied.

B.4 Evaluation-Space Learning Case Study

  • Seed Rubric B mistakes the Hamiltonian formulation for a wholly new primitive, counts implementation components without their shared role, and treats unproven quantum advantage as little complexity contribution.The reference assessment gives intermediate novelty, parsimony, and algorithmic-complexity scores for the proposal’s stated features and resource comparisons.Source: §B.4 · pp. 19–20
  • 0.183 versus 0.216 absolute-score MAE on the held-out test split: the validation-selected PHD student state outperforms the w/o PHD student.The comparison evaluates revised rubric and memory jointly, not the causal contribution of either artifact alone.Source: §B.4 · p. 20
  • The held-out proposal’s MAE decreases from 0.667 with seed Rubric B to 0.500 after learning, although generality shifts from a correct 3 to 2.The learned novelty boundary corrects the score from 3 to 2, parsimony moves from 0 to 1, and algorithmic complexity remains underestimated.Source: §B.4 · pp. 19–20

C.1 Hindsight Reference Panel

The reference panel turns privileged hindsight into population-level supervision, while the pairwise formulation represents preference through differences in shared evaluation-space features. Individual adaptation then updates a researcher-specific mapping using sparse feedback.

C.1 Hindsight Reference Panel

  • Three model families independently judge each proposal pair using the same hindsight evidence bundle; their majority vote supplies the population reference label.Judge probabilities, rationales, confidence, and decisive factors remain teacher-side metadata and do not change that label.Source: §C.1 · pp. 20–22

C.2 Pairwise Instantiation of Population Taste

  • Pairwise population taste uses relative evaluation-space features for proposals A and B, with proposal-visible conditional rules augmenting criterion weights.The resulting preference probability is modeled from a utility over the relative features and proposal-visible evidence.Source: §C.2 · p. 22

D Individual Taste Adaptation Details

  • Individual adaptation freezes the shared assessor and population state, then updates personal state from revealed pairwise feedback without hindsight evidence or held-out labels.A population-centered residual is fit first; personal conditional rules are considered only for recurring errors and replay-checked before retention.Source: D Individual Taste Adaptation Details · p. 22

D.1 Replay-Based Rule Acceptance

Replay-based rule acceptance uses cross-fitted predictions on accumulated support comparisons to screen candidate personal rules before deployment.

D.1 Replay-Based Rule Acceptance

  • Leave-one-out cross-fitted replay refits the individual residual and candidate rule on n−1 comparisons, then evaluates the held-out comparison.The replay tracks corrected and broken StrictPair predictions and the change in negative log-likelihood.Source: §D.1 · pp. 22–23
  • A personal rule is accepted only with a nonzero coefficient, sufficient activations, at least one net correction, no broken predictions, and improved negative log-likelihood.The activation threshold is max(4, round(k/3)); deployed coefficients aggregate out-of-fold fits.Source: §D.1 · p. 23
  • The replay ablation bypasses activation, correction, breakage, and likelihood gates while retaining the nonzero-coefficient check, isolating replay-based acceptance.It keeps the rule proposer, coefficient fitting, support sequence, and downstream evaluation unchanged.Source: §D.1 · p. 23

D.2 A Case Study of Divergent Individual Taste

In a case study of two researchers, the complete personalized mapping recovered more conflicting preferences than either cold start or residual-only adaptation, but accuracy did not rise for both researchers.

D.2 A Case Study of Divergent Individual Taste

  • The two researchers judged the same ten proposal pairs using identical frozen criterion features, with five pairs eliciting opposing preferences.This setup tests whether personalized mappings recover conflicting choices rather than reproduce a shared population prediction.Source: §D.2 · p. 23
  • On five conflicting pairs, the complete mapping recovered three opposing preferences, versus one for residual-only adaptation and none for cold start.U04 rose from 4/10 to 8/10, while U05 fell from 7/10 to 6/10 after adaptation.Source: §D.2 · p. 23
  • Improved recovery of opposing choices did not guarantee higher accuracy for every researcher.U05's accuracy decreased after adaptation even as the complete mapping recovered more conflicting preferences overall.Source: §D.2 · p. 23

E Data Construction and Evaluation Protocols

The evaluation combines reconstructed quantum-research proposals, paper-disjoint population comparisons, and sparse-feedback researcher sessions, supplemented by simulated cohorts.

E Data Construction and Evaluation Protocols

  • The shared proposal corpus is curated from SciRate quantum-algorithm papers, with broad topical coverage and a classic lookup set.Retained papers are stratified by topical cluster, release time, and scites after noisy clusters are removed.Source: E Data Construction and Evaluation… · p. 23
  • Paper-derived proposals preserve proposal-time claims and plans while rewriting results and other post-hoc outcomes as objectives or hypotheses.Synthetic strengthen, weaken, and revision variants are excluded from absolute criterion assessment.Source: E Data Construction and Evaluation… · p. 24
  • Population evaluation uses curated, paper-disjoint splits and temporally comparable, informative proposal pairs with controlled proposal degree and winner imbalance.Pairs are not duplicated across splits, and no underlying paper crosses train, validation, and test.Source: E Data Construction and Evaluation… · p. 24
  • The seven-researcher study tests within-person held-out preference prediction, not the population distribution of scientific taste.Each researcher provides 15 sequential support comparisons and 10 held-out query pairs; query pairs are excluded from adaptation, though proposals may recur across the split.Source: E Data Construction and Evaluation… · p. 24
  • Seven real researchers are complemented by 100 simulated researchers under the same 15-support and 10-query protocol.The simulated cohorts comprise 50 publication-history-derived quantum profiles and 50 controlled profiles; they are reported separately and are not treated as human evidence.Source: E Data Construction and Evaluation… · p. 24
  • Baseline configurations vary backbones, input representations, and target-researcher feedback, with external population judges also tested as alternative priors.Detailed implementation settings are provided separately.Source: Table 9 · p. 25

F Baseline Configurations and Metric Details

This appendix specifies the alternative population priors and individual-taste baselines, along with how confidence is mapped into the RW-SP evaluation weighting.

F Baseline Configurations and Metric Details

  • All personalization methods use 15 support comparisons and 10 held-out query pairs; RW-SP reports individual performance using regret-weighted StrictPair.Support memory uses arXiv abstracts, while queries use full proposal text unless otherwise specified.Source: F Baseline Configurations and Metric… · p. 25, §F.2 · p. 25

F.1 Alternative Population Priors

  • External scientific judges are converted into preference priors and evaluated under the same adaptation protocol; Auto-J, MLR-Judge, and SciJudge use distinct scoring or voting procedures.MLR-Judge averages ensemble OverallAssessment scores, while SciJudge combines order-swapped votes using protocol-assigned preference values.Source: §F.1 · p. 25

F.2 Individual-Taste Baselines

  • The evaluated baselines include direct history prompting, researcher profiles, and methods that select demonstrations or induce preference notes.SC includes support comparisons in context; PP summarizes them, while SynthesizeMe selects demonstrations and CIPHER retrieves preference notes.Source: §F.2 · pp. 25–26
  • Other individual-taste baselines use latent preference models, iterative prompt refinement, shared preference bases, or parameter-efficient adaptation.The section specifies implementation settings for VPL, FERMI, LoRe, and FSPO.Source: §F.2 · p. 26

F.3 Confidence Mapping for RW-SP

  • RW-SP converts choice-relative confidence into an orientation-specific probability: confidence for a selected A, or one minus confidence for a selected B.The metric uses this probability rather than the raw choice-relative confidence annotation.Source: §F.3 · p. 26
  • A probability of 0.5 receives zero regret weight, while probabilities of 0 or 1 receive unit weight.If probabilities are unavailable, the evaluation falls back to the hard preference label; if all regret weights are zero, it uses unweighted StrictPair mean.Source: §F.3 · p. 26
  • For aggregated persona judgments, orientation-specific probabilities are averaged; these probabilities weight evaluation but are not soft training targets.Individual residual and rule fitting use observed binary preference labels.Source: §F.3 · p. 26

G Interactive Proposal Exploration and Deployment Evaluation

The deployment evaluation integrates personalized rankings into interactive proposal exploration and compares them with a matched in-context baseline. It measures engagement in shared competitive displays, while treating immediate interaction as distinct from eventual scientific outcomes.

G Interactive Proposal Exploration and Deployment Evaluation

  • The web system uses the learned taste mapping and accumulated feedback to personalize proposal rankings during successive exploration interactions.Figure 3 illustrates the interface and an example interaction process.Source: G Interactive Proposal Exploration and… · p. 26, Figure 3 · p. 27
  • The deployment study evaluates 20 real researchers across quantum research, computer science, statistics, and medicine, using interaction trajectories as behavioral evidence.Researchers can inspect, refine, or pursue displayed proposals.Source: G Interactive Proposal Exploration and… · pp. 26–27
  • Offline simulations add 100 profiles—50 quantum-publication-derived and 50 controlled trade-off profiles—and are reported separately from observed human engagement.The simulated cohorts broaden the range of preference configurations.Source: G Interactive Proposal Exploration and… · p. 27
  • Both ranking policies receive the same proposal pool, backbone, visible information, and interaction history, isolating the comparison of explicit taste mapping against direct in-context personalization.The SC baseline conditions directly on the complete interaction history.Source: G Interactive Proposal Exploration and… · p. 27
  • Source-blind competitive-pair interleaving places disagreeing selections together, randomizes within-pair order, and withholds which policy contributed each proposal.Proposals selected by both policies appear once and receive no policy-specific credit.Source: G Interactive Proposal Exploration and… · p. 27
  • Competitive-Pair Engagement Gain compares average engagement between policy-contributed proposals within competitive pairs; win rate and ranking NDCG are additional measures.Engagement scores range from no substantive engagement to adoption; NDCG treats these scores as graded relevance.Source: G Interactive Proposal Exploration and… · p. 28
  • Because engagement is observed under a shared interleaved display, NDCG is descriptive and competitive-pair outcomes remain the primary policy comparison.Engagement reflects immediate interaction with candidate directions, not eventual scientific outcomes; the deployment evaluation therefore addresses proposal exploration rather than long-term impact.Source: G Interactive Proposal Exploration and… · p. 28

H Agentic Workflow and Prompt Specifications

This section specifies the agentic workflow and prompts for learning evaluation-space assessments, population preferences, and individual taste, while separating language-model proposal and diagnosis tasks from externally controlled deterministic acceptance and numerical estimation. It also describes how hindsight may guide training while deployment remains limited to proposal-visible inputs.

H Agentic Workflow and Prompt Specifications

  • Deterministic operations handle retrieval, rule parsing, validation, acceptance, and other workflow steps; language-model calls use temperature zero unless otherwise stated.Personal properties must be independently answerable for one proposal, separate a nontrivial subset of support interactions, and avoid topic labels, pair memorization, hindsight outcomes, and hidden researcher information.Source: H Agentic Workflow and Prompt… · p. 28, §H.3.2 · p. 33, §H.3.3 · p. 33

H.1.1 Proposal-Time Criterion Assessor

  • The deployable criterion assessor is the only language-model call in the forward perception path; it uses retrieved student-safe memory and Rubric B to score proposals on five criteria from 0–3.It uses only proposal-visible information, the supplied rubric, and retrieved memory, and returns per-criterion confidence (0–1), proposal-grounded comments, evidence gaps, and constraint violations.Source: §H.1.1 · pp. 28–29

H.1.2 Privileged Criterion Teacher

  • A privileged criterion teacher uses student errors against Rubric-A references and teacher-side evidence to create reusable, proposal-visible assessment lessons, without copying hindsight-dependent facts into them.Lessons undergo schema and leakage checks, deduplication, conflict resolution, and held-out validation; updates are retained only if validation improves.Source: §H.1.2 · p. 29

H.2 Population Taste Distillation

  • Population Taste Distillation freezes the evaluation-space assessor and learns a population preference mapping from hindsight-panel supervision, while deployment uses only proposal-grounded features and evidence.The comparator selects A or B solely by the sign of its utility; conditional-rule activations are deterministic inputs, and orientation swapping is reserved for robustness and validation.Source: §H.2 · p. 29, §H.2.1 · pp. 29–30

H.2.2 Privileged Population Teacher

  • A privileged population teacher translates recurring disagreements into small, independently testable policy edits, abstaining when hindsight has no proposal-visible proxy.Candidate changes are screened for leakage and evaluated independently by a deterministic validation gate.Source: §H.2.2 · pp. 30–31

H.2.3 Conditional-Rule Proposer

  • The conditional-rule path proposes falsifiable rules evaluable on one proposal, with activations based on evaluation features or registered text extractors and pairwise effects computed from activation differences.Tier-2 extractors assess a present-tense proposal property with evidence and confidence; their descriptions, models, and prompts are frozen at registration. Workflow reflection uses aggregate statistics, cannot inspect evaluation examples, and cannot approve its own revision.Source: §H.2.3 · p. 31, §H.2.4 · p. 32, §H.2.5 · p. 32

H.3.1 Evaluation-Space Personal Rule Proposer

  • Individual conditional-rule proposals are considered only after continuous weight residuals have been fitted and sufficient support reveals a repeated error pattern; they must not be disguised global weight changes or one-example anecdotes.The proposer targets a repeated pattern in remaining cross-fitted errors and states the effect direction, while rule strength is estimated statistically outside the language model.Source: §H.3 · p. 32, §H.3.1 · pp. 32–33

H.3.2 Proposal-Text Personal Property Proposer

  • If no evaluation-space personal rule survives replay, a fallback proposer may identify a reusable, present-tense proposal-text property outside the six-criterion space that generalizes across topics and avoids hindsight signals.A rewrite agent may use rejection diagnostics and the same support history to propose a substantively different property, but receives no additional labels.Source: §H.3.2 · p. 33, §H.3.3 · p. 33

H.4 Deterministic Components

  • Nearest-neighbor case retrieval, error ranking, and logistic fitting of w_pop and Δw_u use no language-model prompt.These operations are listed as prompt-free deterministic components.Source: §H.4 · p. 34
  • Proposal-order swapping and StrictPair computation are deterministic operations with no language-model prompt.Source: §H.4 · p. 34
  • Predicate parsing and rule-activation computation are performed without language-model prompts.Source: §H.4 · p. 34
  • Coefficient fitting, reliability updates, and cross-fitted replay are prompt-free deterministic operations.Source: §H.4 · p. 34
  • Policy atomization, leakage screening, validation-gate decisions, rollback, and held-out evaluation use no language-model prompt.Source: §H.4 · p. 34
  • Language-model agents propose or diagnose candidate textual artifacts, while acceptance and numerical parameter estimation remain externally controlled.Source: §H.4 · p. 34
Loading 2610.19719v1…