National Sun Yat-sen UniversityDADH2026Kaohsiung31 October 2026
Download PDF

Scaling Interpretation

Agentic Knowledge Distillation for Computational Intellectual History

Javier Cha

The University of Hong Kong

HKADH, Hong Kong Association for Digital Humanities

Schmidt SciencesSchmidt Sciences

Supported by the Humanities and AI Virtual Institute (HAVI), a program of Schmidt Sciences.

DADH2026National Sun Yat-sen UniversityKaohsiung31 October 2026playingheaven.ai/talks/2026-10-31/dadh

Computational Hermeneutics?

Until recently, following and interpreting an argument was assumed to be an exclusively human ability.

Word2Vec and Conceptual History

Word2Vec-style embeddings are useful for conceptual history (Begriffsgeschichte), such as tracing the changing meaning of the word “evolution” before and after Darwin.

  1. 1850s
  2. 1900s
  3. 1990s
spreadsowseedsowsscatter circulatednewspapers televisionradiobbc broadcast (1850s) broadcast (1900s) broadcast (1990s)
Redrawn from the published HistWords figure. English historical corpora, decade bins. nlp.stanford.edu/projects/histwords

竹之始生，一寸之萌耳，而節葉具焉。……今畫者乃節節而爲之，葉葉而累之，豈復有竹乎？故畫竹，必先得成竹於胸中。

Su Shi, “Record of Wen Yuke Painting the Curved Bamboo of Yundang Valley,” 1079; excerpt

When bamboo first grows, it is but an inch-long sprout, yet joints and leaves are complete in it. … Today’s painters make it joint by joint and pile it up leaf by leaf; how can there still be bamboo? Therefore, in painting bamboo, one must first have the complete bamboo in one’s breast.

Translation: Gemini 3.8 Flash (High)

故竹畫

Schematic

Detecting words is not reading an argument.

Word counts and word vectors recover patterns of vocabulary. Neither follows how Su Shi develops an argument across the passage.

A Short History of the Transformer

  1. 2017

    “Attention Is All You Need”

    Transformer architecture, developed for machine translation

  2. 2018

    BERT

    Contextual language representation for classification and question answering

  3. 2019

    BERT in Google Search

    Queries interpreted in context to improve search results

  4. 2020

    Retrieval-Augmented Generation (RAG)

    Retrieve relevant passages, then generate from them

  5. 2022

    ChatGPT

    Conversational generative AI

  6. 2023–2025

    Agentic AI

    Models use tools, work across files and applications, and carry out multi-step tasks

Humans and Machines Read Differently

  1. Humans read text by paying close attention to language and nuance—and with experience, historical knowledge, and judgement.

    Interpretation

  2. Machines work with numerical representations learned from data. Across many layers and thousands of dimensions, these representations become complex enough to produce interpretation-like outputs.

    Interpretation-like output

Similar outputs do not imply the same process.

Still, this approximation in interpretive ability is too good to ignore.

Chatbots vs. Agents

Chatbot

responds to a prompt

Agent

can also

  • work with files and structured records
  • use tools
  • run code, where permitted
  • inspect intermediate results
  • decide what operation comes next, within a bounded task

An agent is not a mysterious autonomous machine. It works within a bounded environment: it can use permitted tools, inspect the result of one operation, and decide what to do next.

  • SOURCEthe essay, in Classical Chinese
  • SKILLwhat to attend to, what to return
  • RULESrules of evidence it must follow
  • HOOKScode that runs around each model call
  • TASKanalyze this essay section by section

An orchestrator runs the agents in sequence, passing each step’s records to the next.

How agents can be run

  • GUInot recommended
  • CLIgood for development tasks
  • Applicationhow it should be done and deployed

Pivot 機: Agentic Knowledge Distillation

Pivot organizes specialized AI agents to analyse historical texts and construct successive layers of interpretation, from individual passages to essays, author profiles, and comparative corpora.

Schematic

  1. 1PassageOriginal text and source
  2. 2SectionAnalysis of an argumentative unit
  3. 3EssaySynthesis of the essay’s arguments
  4. 4Author ProfileInterpretation based on a selection of writings
  5. 5Broader ComparisonsComparison across authors and texts

Every interpretation remains linked to its sources.

Each layer is open to revision and depends on what was selected. Secondary scholarship is distilled and linked to every layer.

Making Tacit Knowledge Explicit

SKILL.md is a plain-text instruction file that the agent reads when it works. The model has not permanently learned these instructions.

Writing instructions for an agent forces us to articulate parts of scholarly practice that usually remain tacit.

skills/primary-sections/SKILL.mdExcerpt

Give each term’s meaning: what it means in these lines, or null when they do not settle it.

The same word can mean something else elsewhere, and that a term occurs establishes the term, never a position.

undetermined: one item for each domain you cannot judge from these lines.

A term’s occurrence is not a position, and a confident reading does not remove a real alternative.

taxonomy v2Excerpt4 of 10 in §1

  1. D1Knowledge, Language, and Reasoningcentral
  2. D2Cosmology, Nature, and Creativitycentral
  3. D3Moral Cultivation and Ethical Life
  4. D4Literature, Arts, and Aestheticscentral
  5. D5Classics, Philology, and Interpretation
  6. D6Governance, Institutions, and Political Economy
  7. D7History, Temporality, and Historical Judgement
  8. D8Scholarly Communities and Intellectual Transmissionincidental
  9. D9Ritual, Religion, and Cultural Practice
  10. D10Emotions and Affect

Ten Domains of Intellectual Inquiry

Co-created with agents over multiple iterations, with feedback from test runs on increasingly larger samples of texts.

  • 名實知行言意

    Knowledge, Language, and Reasoning

  • 天人理氣造化

    Cosmology, Nature, and Creativity

  • 修身心性出處

    Moral Cultivation and Ethical Life

  • 文詩書畫

    Literature, Arts, and Aesthetics

  • 經學訓詁考證

    Classics, Philology, and Interpretation

  • 經世理財郡縣

    Governance, Institutions, and Political Economy

  • 史勢時

    History, Temporality, and Historical Judgement

  • 師友道統學派

    Scholarly Communities and Intellectual Transmission

  • 禮祭祀佛老

    Ritual, Religion, and Cultural Practice

  • 情喜怒哀樂

    Emotions and Affect

Agentic Knowledge Distillation in Practice

Specialized AI agents develop interpretations from passages to essays and scholar profiles in a manner that preserves links to the underlying sources for verification and revision.

From a passage to a comparative interpretation, and backFive panels telescoping left to right: a passage as one column of characters; the section, the few columns of one argumentative unit with the passage's column picked out and a lapis bracket over them; the essay, all its columns with the section's run washed in lapis; the author profile, a selection of essays as rectangles of varying size with this essay's picked out; the comparative interpretation, this author's tiles beside another's with a dashed lapis loop across both, open. At the last step a lapis line returns beneath the panels to the passage: Every interpretation remains linked to its sources..PASSAGESECTIONESSAY§1§2§3§4§5AUTHOR PROFILECOMPARISONSEvery interpretation remains linked to its sources.
  1. Original text and source
  2. Analysis of an argumentative unit
  3. Synthesis of the essay’s arguments
  4. Interpretation based on a selection of writings
  5. Comparison across authors and texts

1Select and Segment the Text

pivot › read_text PV.SRC.000484

文與可畫篔簹谷偃竹記, line 4

  1. §1Epistemological Theory of Ink Bamboo Painting
  2. §2Philosophical Allusions and Methodological Transmission
  3. §3Epistolary Banter and Exchange of the Yundang Scroll
  4. §4The Culinary-Poetic Interaction in Yundang Valley
  5. §5Sudden Bereavement and the Commemorative Function of Humor

Historical Source

故畫竹，必先得成竹於胸中

Therefore, to paint bamboo one must first have the complete bamboo in the breast.

Su Shi 蘇軾, 1079, working translation; the essay divided into five argumentative units by the plan step

2Apply an Agent Skill

pivot › start_run analyze-primary work=PV.SRC.000484 from=sections

primary-sections

Agent Skill

Analyse up to four planned sections of one classical Chinese essay. For each section give its key terms, structural pivots, key arguments, logic and flow, support and evidence, classical allusions, and which broad domains of the taxonomy it bears on, citing the author’s lines by id.

SKILL.md, revised for taxonomy version 2; workflow analyze-primary, step sections

3Construct a Section Interpretation

Model Output, Test Reading

The author sets out a rule for painting bamboo, calls it Yuke’s teaching, and goes on to say why a person who knows the reason for such a rule may still fail to carry it out.

Position on 成竹, affirmed, as stated. Held back: he does not claim to have done it himself, and the lines give no account of how the fully formed bamboo is got.

Section §1, arguments and a position; Claude Sonnet 5.5, a test reading under taxonomy version 2, not yet a Pivot record; excerpts

4Examine the Domains

Domains

  • D1Knowledge, Language, and Reasoning centralLine 5 separates knowing the reason in the mind from being able to do it.
  • D2Cosmology, Nature, and Creativity centralLine 4 has the painter start from the bamboo already formed in his chest.
  • D4Literature, Arts, and Aesthetics centralAn account of how a painter of bamboo is to work, to a quick, straight drive of the brush.
  • D8Scholarly Communities and Intellectual Transmission incidentalLine 5 gives the account as Yuke’s teaching to the author.

成竹 Fully formed bamboo內外不一 Inner and outer not one心手不相應 Mind and hand not answering each other

The same test reading’s domains, with their salience, and its terms; the claims are excerpts

5Construct Larger Interpretations

Model Output, Test Reading

Thesis, from the five section records:

The essay joins an account of how bamboo is painted with an account of a friendship, and lets the second carry the commemoration.

Central: Literature, Arts, and Aesthetics; Emotions and Affect. Developed: D1, D2, D7, D8.

成竹 Fully formed bamboo不學 Not learning意 / 法 Idea and method戱笑之言 Jesting words

Essay synthesis; Claude Sonnet 5.5, a test reading under taxonomy version 2, not yet a Pivot record; excerpts

6Construct Larger Interpretations

pivot › read_records PV.PRF.000009

“The profile is provisional and conditioned on exactly these essays.”

Corpus Scale

164 selected essays and 761 section records

Pivot, Original Pivot reading, 13 October 2026, under taxonomy version 1; not the whole of Su Shi’s surviving work

7Construct Larger Interpretations

Recorded Debate Outcome

What are 王安石 (Wang Anshi) and 蘇軾 (Su Shi)’s positions on what is 道 (dao)?

Compared in a Pivot Debate over both authors’ distilled records; outcome PV.MDB.000002, provisional. The next slide.

No compare-authors run exists for the pair; the comparison was conducted as a Debate

8Construct Larger Interpretations

pivot › find_links PV.ESS.000483

Source Return

  1. PV.ESS.000483the essay synthesis
  2. PV.SEC.002053 §1the section record
  3. PV.SRC.000484, line 4故畫竹，必先得成竹於胸中

Each record keeps its route back to the passage; inspection, correction and revision happen at every link

Experiment 1: Wang Anshi and Su Shi on Dao

Model
  • Gemini 3.8 Flash (High)
  • Gemini 3.8 Flash (Medium)
  • DeepSeek V4 Flash
  • Kimi K3
  • GLM-5.3
Question

What are 王安石 (Wang Anshi) and 蘇軾 (Su Shi)’s positions on what is 道 (dao)?

Participants
  1. First王安石 (Wang Anshi)
  2. Second蘇軾 (Su Shi)
  3. positions read from the dossier by the model, not the authors speaking

Rounds 2 Comparison between interpretations

Dossier
  • PV.ESS.000678與祖擇之書, Wang Anshi
  • PV.PRF.000007Profile, Wang Anshi
  • PV.ESS.000363答張文潛書, Su Shi
  • PV.SRC.000843Bol, This Culture of Ours, excerpt
  1. Model selected
  2. Question submitted
  3. Dossier assembled
  4. Debate begun: PV.DEB.000002, run PV.RUN.002528

Recorded Debate Output

Position read for 王安石 (Wang Anshi)

Wang Anshi conceives of Dao (道) as grounded in practical governance, statecraft, and administrative order (「治教政令」、「實用和實踐」). True literature (「聖人之所謂文」) is ontologically identical with sage statecraft, policy decrees, and institutional execution […]

[PV.SRC.000843] [PV.ESS.000678] [PV.PRF.000007]

Position read for 蘇軾 (Su Shi)

Su Shi conceives of Dao (道) not as an authoritarian monopoly or rigid statutory code, but as a living, immanent dynamic of nature and balance (「中庸」) […] literature is co-extensive and contemporaneous with Dao (「與道俱」), emerging organically from moral character rather than enforced conformity.

[PV.SRC.000843] [PV.ESS.000363]

They differ onThe relationship between Literature (文) and Dao (道)

Wang Wang views true literature (「聖人之所謂文」) as coterminous with the actual administrative machinery and moral self-admonition of statecraft, subordinating rhetorical composition to governance [PV.SRC.000843] [PV.ESS.000678]

Su Su views literature as co-extensive and unfolding with the Way (「與道俱」), maintaining that creative, unforced literary engagement allows individuals to independently attain sagehood [PV.SRC.000843]

Source return

  1. PV.ESS.000678與祖擇之書Wang Anshi’s letter, as its essay synthesis reads it
  2. PV.ESS.000363答張文潛書Su Shi’s letter, as its essay synthesis reads it
  3. PV.PRF.000007Profile · Wang Anshithe author profile across his selected essays

Debate outcome PV.MDB.000002, a provisional synthesis written by Gemini 3.8 Flash (Medium): draft, 28 September 2026; Pivot’s validation left it requires review; stale since 1 October, when a new text version changed sections it rests on. The participants’ own models are not in the public record.

Experiment 2: From Kaifeng to Kaesŏng

Early in the twelfth century, a few Koryŏ princes, relatives of the regent, and scholar-retainers visited the Northern Song during the afterlife of the New Policies 新法. The sources reveal little about what they saw, whom they met, or which texts informed them.

講書無逸篇

《高麗史》 Koryŏsa

A scholar-official “expounded the ‘Against Luxurious Ease’ chapter of the Documents.”

The record preserves the event, not the lecture.

The chapter’s own themes

  • royal diligence

  • agrarian labour

  • consumption

  • the burdens of government

Implications, inferred

  • fiscal and military policy

  • extraction

  • administrative excess

The task is not to generate the missing lecture, but to bound the range of arguments that could plausibly have informed it, read with commentarial traditions and Koryŏ debates.

Scaling Interpretation

While still in infant stage, agentic knowledge distillation promises to allow intellectual historians to examine larger bodies of evidence in a way that remains attentive to the possibilities, limits, and contingencies of what we can infer about the past world views.

Thank You

Research material

Citations

Permalink: https://playingheaven.ai/talks/2026-10-31/dadh/

Landscape suits the presentation.