Soo Hur — Research

CRISP · crisper

A Research Sequence and Its Workbench

Concepts and workflow in brief · 24 July 2026

CRISP divides source-based research into five stages; crisper is the workbench that lets you work in that sequence, handing over rules and tools together.

Everything here turns on a single question — may this claim be used as evidence? Ordinary knowledge-management tools are good at connecting material but have no place to record the answer; crisper puts that place inside the file itself.

1 Five stages, and the axis that precedes them

0

Center

중심

Set the research question, split it into sub-questions. Not a stage but an axis running through all of them.

C

Collect

수집

Gather prior scholarship and primary sources. The original is never altered.

R

Refine

정제

Make the text readable and citable — correction, page-addressing.

I

Investigate

분석

Break it small and observe. One passage = one card.

S

Signify

해석

Cut across the cards to answer the question.

P

Present

발표

Two standards — coherence (toward the question), communicability (toward the audience).

Stages are implemented as folders. Because the folder names are the stage names, where you currently are is visible in the file path.

2 Observation · Commentary · Interpretation — the firewall

If a judgment slips into the record disguised as description, there is no way to retrace it later. So the act of assigning meaning is split into three levels, each written in a different place.

Observation

descriptive · reproducible

What is in the source, written so that anyone would write the same thing. Subject to verification.

Commentary

meaning-making · fragment

What this one passage means. Assertible, but the evidence is confined to the passage in front of you.

Interpretation

meaning-making · question

The arrangement between passages — contrast, sequence, absence. No single fragment can carry it alone.

Interpretation is not the sum of the commentaries. Objects such as an absence ("this name occurs 0 times in the corpus") or a trajectory (a step-change in dating) do not exist at the scale of the fragment; they are produced only in interpretation.

A one-line example. A concept term first appears in a text of 1926, while the name of the thinker who coined it appears only in 1931. That much is observation. "Was the name deliberately avoided?" is interpretation, and it is written elsewhere.

3 The workbench — rules and tools as one bundle

Folder conventions, card format and status markers (the rules) together with editor, conversion tools and version control (the tools) are handed over as a single bundle, which the user receives by copying it whole.

It is not an application. We first ran the whole process by hand, without any app. Exactly one point of friction turned up, and a link resolved it. No evidence emerged that a program was needed, so it stayed as files and rules.

Floor
editor · AI tools · version control
swappable
Workbench
folder conventions · card format · the two states · metrics
the identity
Work
refined texts · cards · manuscripts
each one's own

The workbench must survive having the floor swapped out. Had it been built as a program, floor and workbench would be one piece and could not be separated.

4 Two states — the heart of the workbench

collated: false # does the excerpt match the original? — filled by AI confirmed: false # may this card be used as evidence? — filled only by a person

collated

collation · delegable

Do the quotation and page match the original?

Open the original and check; anyone arrives at the same answer. It is needed because this actually goes wrong often.

confirmed

confirmation · not delegable

May this card serve as evidence?

AI may observe, compare, and even name patterns, but it does not confirm.

Why they are split. Collation is labor that can be delegated; confirmation is not. Bound into a single field, the delegable work queues up behind the bottleneck of human judgment. Split, the AI goes as far as it can and the person is left holding judgment alone.

The same structure doubles as a record of AI use. Instead of a self-reported "how much AI did you use," what the machine did and what the person did are inscribed separately in the file.

5 Metrics — a display, not a gate

An unmet requirement does not block the next stage. Instead, the location and the quantity of what remains unsettled stay visible. Counting cards and confirmations per sub-question yields a table, and the blanks in that table tell you what to do next.

A blank is a finding, not a failure. A sub-question with no cards at all means that no material has yet been brought in for that direction — and now that fact is a number. A low confirmation count says the same thing: confirmation is not homework in arrears but the act immediately preceding citation, so cards not yet cited remaining unconfirmed is normal.

6 Several questions — one body of material, many workbenches

When the question changes, the sub-questions change and everything above them changes with them. So a new workbench is set up for each question. What has been gathered and what has been corrected, however, pass to the next question unchanged.

Per question — closes, incommensurable

0 · I · S · P center / investigate / signify / present

Excerpts, interpretation and composition belong to that question. A different question cuts the material in a different place.

While work is under way the two questions know nothing of each other. No burden is placed on the researcher to keep other projects in mind — because a rule that depends on human memory is a rule that will not be kept. The connection is made afterwards by reading file names alone. The naming convention already carries source and page, so it costs nothing.

question-A/investigate/YiDonhwa1931-p97_elan.md question-B/investigate/YiDonhwa1931-p97_self-and-other.md └── the same "site" = source + page ──┘

The resulting map shows three things. Sites with heavy overlap — either the passage really matters, or this is where I always look. Sites visited once — starting a new question here means reading my own earlier cards rather than someone else's. Empty stretches — pages no one has read yet.

7 Where it sits, and what it does not do

Ordinary knowledge managementThe workbench
Core connection confirmation
Direction of growth more links, a denser graph — breadth the same material standing as evidence in one question after another — depth
Weak point a link joins the false as neatly as the true; no place to record whether it was verified this axis appears only once the workbenches are separated