KUPAS MASTER: Distilling the Tacit Expertise of Master Practitioners into Agent-Ready Experience Corpora

Changmian Wang; Yuchao Ma; Xuchao Lu; Chen Zhang; Ping Sun; Jiazheng Wang; Shan Wang; Xuanwen Chen

cs.AISep 29, 2026

KUPAS MASTER: Distilling the Tacit Expertise of Master Practitioners into Agent-Ready Experience Corpora

Authors: Changmian Wang, Yuchao Ma, Xuchao Lu, Chen Zhang, Ping Sun, Jiazheng Wang, Shan Wang, Xuanwen Chen, +8 more

Organizations: Shanghai Kupas Technology Co., Ltd. and Tongji University

Abstract

Experienced professionals know more than just facts and conclusions. They know which cues matter, why a judgment is reasonable, and which action to take. Routine work records often leave out this tacit knowledge, making it difficult for Large Language Model (LLM) agents to use professional experience effectively. We introduce KUPAS MASTER, an experience engineering platform built around nine-layer cognitive corpus construction. It turns heterogeneous work records and practitioner interviews into traceable, reusable experience corpora for agents. Six case elements preserve the task process: context, cues, judgment, action, boundaries, and outcomes. Nine-layer cognitive corpus construction organizes tacit experience along nine extraction dimensions and stores the resulting assets in six libraries: rules, constraints, best practices, negative examples, corner cases, and skills. Semantic alignment, individual experience distillation, organizational consolidation, and cross-review preserve source evidence, conditions of use, and unresolved disagreements. The platform packages these assets into callable skills with explicit inputs, steps, dependencies, and stopping conditions, connecting experience collection to task execution and evaluation feedback. Using authorized samples from 20 randomly selected practitioners, the platform processed 1,576 source files into 23,024 individual experience records and 13,113 organizational assets. The evaluation spans multiple professional domains. Under common task inputs and scoring criteria, the base model, raw corpus retrieval-augmented generation (RAG), and KUPAS MASTER agent scored 70.63, 79.75, and 89.58, respectively. The KUPAS MASTER agent improved on raw-corpus RAG in all seven scoring dimensions. The platform provides a practical path from individual tacit experience to organizational knowledge and agent capabilities.

Figures & tables

Figure 1 : System overview of KUPAS MASTER and its architecture for using professional experience.

Figure 2 : The seven-step KUPAS MASTER workflow, grouped into data collection, experience construction, and agent evaluation. Solid arrows show processing order. The dashed loop returns evaluation findings to corpus revision. Each stage retains source references, conditions of use, review status, and version identifiers.

Element	Content	Distinctions to retain
Context	Task, role, object, environment, time, and available resources.	Conditions observed in this case versus conditions assumed for reuse.
Cues	Signals, measurements, statements, and changes noticed during work.	Direct observations versus reported or inferred observations.
Judgment	Interpretations, supporting reasons, alternatives, and uncertainty.	Practitioner accounts versus model-generated explanations.
Action	Selected operations, order, parameters, and stopping conditions.	Planned, attempted, and completed actions.
Boundaries	Prohibitions, scope, missing prerequisites, and referral conditions.	Mandatory constraints versus preferences or usual practice.
Outcome	Immediate results, later verification, and unresolved effects.	Expected, reported, and independently verified outcomes.

Table 1: Case elements and recording requirements. Facts, judgments, and execution states remain distinct, and missing fields are explicit.

Figure 3 : The nine extraction dimensions, grouped by function. Case elements record the task process, and reviewed statements can become six types of experience asset. Dimensions and asset types have a many-to-many relationship: a skill may combine attention, decision, habit, and constraint information.

ID	Dimension	Extraction question	Representative objects
L1	Knowledge	Which facts and concepts were used?	Entities, terms, relations, and domain materials.
L2	Attention	Which observations received priority?	Diagnostic cues, ignored distractions, and shifts of focus.
L3	Judgment	What supports this interpretation?	Conditional judgments, thresholds, alternatives, and uncertainty.
L4	Decision	How was an action selected?	Prerequisites, selection criteria, and stopping conditions.
L5	Association	Which other cases or concepts were relevant?	Analogies, recalled precedents, and links across cases.
L6	Anticipation	What consequences were expected?	Predicted effects, time horizons, and follow-up checks.

Table 2: Extraction questions and representative objects in the nine-layer framework. Outputs become assets after evidence and applicability review.

Figure 4 : Shared dependencies and cross-layer links among the nine extraction pipelines. The arrow from L1 to the L2–L9 group denotes a shared anchor map for object identity and source tracing. Each pipeline also reads its own required source records. Other arrows denote evidence references, candidate additions, or review constraints supported by sources.

Figure 5 : Many-to-many functional correspondences between brain systems and the nine-layer framework, based on neuroscience studies. The upper drawing shows the right lateral hemisphere; the lower drawing projects medial and deep structures. Letters match the table rows, and arrows indicate approximate locations. References: N1 [ 11 ] ; N2 [ 13 ] ; N3 [ 14 ] ; N4 [ 15 ] ; N5 [ 16 ] ; N6 [ 12 ] ; N7 [ 17 ] ; N8 [ 18 ] ; N9 [ 19 ] ; N10 [ 20 ] .

Asset	Reusable content	Common dimensions
Rules	Scoped conditions and recommended judgments or actions, including exceptions.	L3, L4
Constraints	Prohibitions or required prerequisites, their basis, and allowed alternatives.	L7, L9
Best practices	Supported successful procedures and conditions for considering reuse.	L1, L4, L5, L6
Negative examples	Inappropriate actions or adverse outcomes, context, and reviewed corrections.	L3, L4, L7, L9
Corner cases	Unusual contexts where a routine rule may fail or need adjustment.	L2, L3, L5, L7
Skills	Callable procedures with inputs, outputs, dependencies, checks, and evidence.	L2, L3, L4, L8, L9

Table 3: Asset types, reusable content, and common source dimensions. Assets can combine information across dimensions.

Figure 6 : Four representations in corpus construction. Source materials (S0) become case narratives (S1), which are analyzed for decision elements (S2) and converted into structured candidate records (S3). Dashed links preserve source traceability across stages.

Operator and input	Transformation and output	Reject or leave unresolved when
Term alignment: wording and context	Propose a standard term while retaining aliases, units, and source spans.	Multiple interpretations remain plausible, or a unit conversion lacks support.
Claim extraction: case narrative	Separate observations, judgments, and planned actions, each with its own source.	The extracted claim lacks supporting content.
Condition preservation: conditional statement	Recover triggers, recommendations, prerequisites, and exceptions as a candidate rule.	Compression changes the action, drops an exception, or alters a prerequisite.
Conflict classification: comparable asset pair	Check scope overlap and action compatibility to distinguish agreement, different conditions, and conflict.	Scope overlap is unknown or evidence is insufficient to resolve the conflict.

Table 4: Core transformation operators and reasons to reject an output or leave it unresolved. Every operator records input and output versions and source locations.

Figure 7 : Four outcomes of organizational consolidation and their evidence requirements. Agreement retains independent support. Context-dependent alternatives remain separate. Outcome comparison requires comparable cases. Unresolved contradictions remain inactive until review. Consolidated assets retain both supporting evidence and conditions of use.

Figure 8 : Skill selection and execution. Retrieval retains active, authorized assets with true or unknown applicability. Invocation requires satisfied prerequisites and action checks. Missing information prompts clarification and another state check; a rejected action may lead to escalation. Actions and escalations are logged. Mandatory constraints apply independently of retrieval ranking.

Item	A: Base model	B: Raw-corpus RAG	C: KUPAS MASTER
Approach	General analysis and an evidence checklist.	Cites work injury insurance regulations and proposes mediation steps.	Calls the mediation assessment skill and combines rules with corner cases.
Boundaries	Does not clearly separate mediation from formal injury determination.	Identifies some procedural and risk boundaries.	States that mediation does not replace formal determination and distinguishes accident scenarios and risks.
Output	General principles and evidence suggestions.	Legal grounds and process advice.	A structured process and an output file.
Comparison	Rated lowest on all 8 items.	Ties with C on 1 item.	Rated best or tied for best on all 8 items.

Table 5: Three-configuration comparison for workplace injury mediation. All configurations assess an accident during a detour to collect a child after work, the evidence needed, and the order of mediation steps.

Figure 9 : The three-configuration comparison. Direct inference provides a model-only reference. Raw-corpus RAG measures the benefit of access to source records. The KUPAS MASTER agent evaluates the complete experience-guided configuration. All three answer the same tasks under common input requirements and scoring criteria.

Item	Configuration and scale
Sample	20 randomly selected authorized practitioners, covering finance and accounting, community governance, engineering quality, safety oversight, healthcare, mediation, and emergency management.
Tasks	177 questions, independently answered by A, B, and C, for 531 responses.
Difficulty	8 easy, 83 medium, and 86 hard questions.
Configurations	A: base model; B: raw-corpus RAG; C: KUPAS MASTER agent.
Scoring	A shared seven-dimension rubric applied to independent answers to the same tasks.
Run traces	Individual records for 5 practitioners and 53 questions, covering retrieval, library calls, and skill execution.

Table 6: Experimental setup and data composition.

Method	Composite score	Score pass rate (%)	Evidence sufficiency and accuracy	Boundaries and compliance	Time (s)
A: Base model	70.63	87.57	62.18	74.52	16.2
B: Raw-corpus RAG	79.75	98.87	77.27	79.19	34.0
C: KUPAS MASTER	89.58	100.00	88.32	88.23	57.1

Table 7: Overall performance of the three configurations. Higher scores and pass rates are better; lower latency is better. Scores and latency weight practitioners equally; pass rates weight the 177 questions equally, with a pass threshold of 60.

Figure 10 : Composite scores for the 20 practitioners. Equal weighting across practitioners gives means of 70.63, 79.75, and 89.58 for A, B, and C, with

C-B=+9.83

and

C-A=+18.96

. C exceeds both baselines for every practitioner.

Figure 11 : Mean scores in seven dimensions, weighting the 20 practitioners equally. C scores range from 87.7 to 91.2. For every practitioner, C exceeds B in every dimension.

Figure 12 : Score gains by dimension. Orange shows the gain from raw-corpus retrieval (

B-A

); green shows the additional gain from experience assets and skills (

C-B

). Labels above the bars give the total (

C-A

). All dimensions weight the 20 practitioners equally.

Asset type	Runs	Observed role and example tasks
Rules	46	Express judgment conditions as rules, thresholds, and requirements, including flood dispatch decisions and engineering quality escalation criteria.
Best practices	33	Supply established procedures and steps that turn experience into a task process.
Corner cases	35	Add exceptions, high-risk situations, and business boundaries that routine rules may miss.
Constraints	28	Identify prohibited actions or required prerequisites for procedural and compliance checks.
Negative examples	18	Provide recurring errors and inappropriate responses to help recognize risky decisions and exceptions.
Skills	46	Organize retrieved experience into steps, checks, responsible roles, and structured deliverables.

Table 8: Asset and skill use in 53 C runs for 53 questions. Counts are by run and can overlap.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 13 : Building, using, and evaluating a hypertension practitioner assistant.

ID	Analysis focus	Supporting evidence
E1	Overall improvement in task outcomes	Configuration scores, paired questions, resampling intervals, and below-threshold scores in the random authorized sample.
E2	Added value from structuring the same information	Overall gain of the experience-and-skill configuration built from the same sources.
E3	Useful, traceable experience representation	Asset fields, source relationships, automated checks, and separately recorded human review.
E4	Consolidation into organizational assets	Individual-to-organizational asset counts, library versions, and consolidation records.
E5	Skills in task execution	Skill calls, file generation, and task scores.
E6	Boundary handling and reliability	Boundary score gains and associations between asset use and boundary-judgment rows.

Appendix

Table 9 : Ten aspects of platform performance and supporting evidence (20 practitioners, 19 platform industry labels, 177 questions, and 531 responses).

Figure 14 : Distribution of 531 scores for 177 questions, with 177 scores per configuration. A/B/C question-weighted means are 70.9/79.9/89.6, pass rates are 87.57%/98.87%/100.00%, and fractions scoring

\geq 80

are 7.3%/58.8%/99.4%. Counts below the score threshold are 22/2/0.

Figure 15 : Latency and score, with one point per practitioner and configuration. Mean A/B/C times are 16.2/34.0/57.1 s, rounded to 16/34/57 s in the legend. Ratios of mean times are

C/A=3.53

and

C/B=1.68

Figure 16 : The six libraries contain 13,113 assets: 3,550 rules, 2,892 best practices, 1,955 skills, 1,948 corner cases, 1,623 constraints, and 1,145 negative examples. All six types are present for 17/20 practitioners. The cumulative horizontal positions use a linear scale up to 200 assets and a logarithmic scale above 200; segment widths therefore do not represent type proportions.

Figure 17 : The uploaded corpus contains 1,576 files, 1.73 GB, and 13 formats. Individual file counts range from 2 to 378, covering documents, spreadsheets, structured data, images, and recordings. MB labels are rounded to whole decimal megabytes; 0 MB denotes less than 0.5 MB, not an empty corpus.

Figure 18 : Judgments in three-configuration comparison reports. Left: wins including ties across 123 rows per category in 62 reports. Right: outcomes across 217 retained reports and 1,345 rows, with 1,109 C-only wins (82.5%) and 214 C ties (15.9%).

Figure 19 : Scale and score gains (

n=20

). Left: the Spearman correlation between organizational asset count and

C-B

is 0.07. Right: the correlation between raw chunk count and

B-A

is 0.61.

Figure 20 : Question-weighted composite scores by difficulty. A/B/C means are 75.5/82.1/90.8 for easy questions (

n=8

), 69.0/79.3/89.3 for medium questions (

n=83

), and 72.3/80.3/89.8 for hard questions (

n=86

). C exceeds both baselines and scores above 89 in all groups.

Explore similar work

Jul 17, 2026cs.CL

SkillCorpus: Consolidating and Evaluating the Open Skill Ecosystem for Real-World LLM Agents

Agent skills, SKILL files that package reusable procedural knowledge for an LLM agent, are a popular mechanism for extending agent capabilities. Public repositories now host them in large and growing numbers, yet these artifacts are fragmented, redundant, and uneven in quality, and their value in practice is unclear. A core question remains open, namely how to consolidate this open-source SKILL ecosystem into a single usable corpus, and what bounds its benefit on real-world agent tasks. We present SkillCorpus, a framework that aggregates, curates, matches, and evaluates the open skill ecosystem at scale. It filters ~821,000 crawled skills through a multi-stage pipeline into 96,401 skills organised by a 16-class taxonomy and three quality facets (utility, robustness, safety), and pairs them with a fine-tuned retrieval-and-selection stack that matches task-relevant skills. We evaluate end-to-end across three benchmarks (SkillsBench, GDPVal, QwenClawBench), two harnesses, and two open backbones with a frontier robustness check. Integrating SkillCorpus yields consistent gains across all three benchmarks, largest on SkillsBench (+7.5 pp). An operational analysis traces the gains to a coverage boundary and a harness boundary. SkillCorpus is, to our knowledge, the first end-to-end account of when a curated, retrieval-served community corpus improves real agent tasks, and where it does not. The dataset, models, and code will be released upon acceptance.

Yanze Wang, Pengfei Yao, Tianyi Sun +7

1EverMind · 2Shanda Group · 3Peking University

Large Language Model Agents Text Corpora Task-Relevant Information

Ask PDF

Sep 27, 2026cs.AI

CORTEX: A Verified Experience Layer for Generalist Agents

An agent can solve a task today and face the same task under new facts, tools, or governing knowledge tomorrow. Most agent systems can retrieve relevant text or recall prior conversations, but they lack a principled way to decide when a previous solution is still valid, when it must be adapted, and when it should be discarded. We introduce CORTEX (Contextual Orchestration and Reuse of Task EXperience), a general AI systems framework that connects specialized agents through an external layer of verified experience. Each episode records its task conditions, source and tool state, decisive predicates, proof trace, verifier, and outcome. A meta-controller chooses exact replay, checked adaptation, fresh synthesis, or escalation. Accepted episodes can become task patterns and procedural strategies through a challenge-driven development loop. This gives the system an implicit competence layer that can grow without changing model weights. We formalize system contracts for exact replay and source-version separation, and derive when reuse saves computation. A controlled two-domain implementation tests the exact-replay core on 1,000 synthetic cases. Complete-family holdouts test procedural transfer on 1,000 new-family cases across eight clinical and policy splits, with complete fresh-evidence grounding and perfect invariance to irrelevant-field and insertion-order perturbations. The transfer trace exposes the work required for verified strategy execution. These results establish an initial path toward general intelligence through reusable procedures, typed experience, and developmental transfer.

Garapati Keerthana, Manik Gupta

Birla Institute of Technology and Science, Pilani, Hyderabad, India

Reusable Knowledge Ai-Based Framework Autonomous Agents Orchestration Developmental Trajectories

Ask PDF

Aug 31, 2026cs.AI

mimeo: Compiling Public Expert Corpora into Agent Skills and Testing What Transfers

Giving an agent a file about a named expert can supply hard-to-find material, produce a recognizable persona, or change what the agent decides. These are different claims. We test each one. mimeo is an open-source tool that finds a person's public work, checks each extracted quotation against the cached source text, and writes a file an agent can load. Eight logged builds averaged 38 model calls; the check rejects 13.2% of extracted quotations. We tested four expert files with one coding-agent harness. Knowledge access was clearest: mimeo answered all 20 obscure, quotation-heavy questions; no closed-book condition answered more than 10. Keyword search (BM25) over the same pages answered 15-17, a gap this sample cannot resolve. Grounding showed one clear benefit: personas written from model memory misstated a documented position on 1-4 of 20 answers under every grader; the plain agent and mimeo never did. Every persona was easy to spot on short open prompts, and adding task material lowered identification by 18-23 points. mimeo was no more identifiable than a from-memory profile. Judgment transfer remained unresolved because both tests hit their ceiling: every condition found 94-97% of the problems planted in engineering tasks and scored 94-100% on 16 new application scenarios. An AI-judged "sounds like the expert" score changed with the judge: two of four preferred answers based on a model's stereotype, while two found no difference on the same text. That is a caution against relying on a single AI judge. The evidence supports mimeo as a compact, inspectable reference on a person, not as a demonstrated transfer of their judgment. Toolkit and expert profiles: https://github.com/K-Dense-AI/mimeo

Timothy Kassis

K-Dense, Inc.

Expert Demonstrations Coding Agents Experts Text Corpora

Ask PDF