Contact
KOR ENG

Methodology

SLDA

Semantic Logic Drift Analysis.
SLDA cross-analyses multiple rounds of documents and dispute material — by time, by party, and by issue — measuring, at the sentence level, how claims move, how expression intensity shifts, how attribution changes, and where mutually incompatible premises collide.

Early Access · 3 / 20 · Free Methodology in closed beta (CBT) · 20 participants No payment step at this time As of 2026.07

The SLDA scoring methodology is currently in closed beta (CBT): we are validating the rubric together with the first 20 participants. Results are not immediate — they take up to three business days. This is not an unfinished product; it is early participation in refining the method.

🔒 The first transformation happens on your device
SLDA is designed in line with the EU General Data Protection Regulation (GDPR) and the principle of Privacy by Design. Original text and real names are never transmitted to our servers. All first-stage transformation — labeling and anonymization — is performed entirely on your own computer. What the server receives is only the already-labeled result.
Inquire about SLDA analysis Check submission status
English service status

The automated SLDA methodology is currently tuned for Korean-language semantics. The English-language analysis service is under development, and direct application is not available on the English site at this time.

Same material, same result — whoever submits it

SLDA does not judge whether a matter is true or false, right or wrong, or who will win. It scores by a rubric we developed ourselves, decomposing the features of words, syllables, and sentences through semantic analysis. When the analysed data is identical, the result is the same no matter who requests it.

  1. Fixed rubric

    Each item is judged against a predefined scoring table. Rules, not the analyst's judgment, set the score, so the same material yields the same result whenever it is analysed.

  2. Gate method

    A claim that does not address the designated issue cannot earn points on internal coherence alone. However well-formed a sentence is, if it departs from the issue it does not pass that item.

  3. Citation-unit normalisation

    Repeated identical strings are counted as a single citation unit. Repeating the same expression is not inflated into new evidence.

  4. Pre-analysis normalisation

    Submitted labeled material is normalised for analysis (format consistency, notation unification). The original content is not altered, so evidence traceability and reproducibility are preserved.

  5. Evidence traceability

    Every judged value is recorded alongside the original sentence it came from. You can trace back to the source why a given score was produced.

SLDA reports the amount, distribution, and amplitude of change.
It does not report the cause, truth, or intent of that change.
It counts and displays; it does not judge right or wrong, win or loss, advantage or disadvantage. "Why did it change here?" is left to the reader.

One engine, three models

The measurement axes — drift, premise collision, attribution — are a single engine. On top of it sit models (packages) by material type, tuning only the observation rules to each context.

Litigation (LIT) · legal filings

Reads multi-round briefs and opinions separately by side (plaintiff / defendant / non-party — never mixed). Registers only measured values; does not judge win/loss or advantage.

Expression-intensity trend Repetition count of the same conclusion Count of evidence substitutions Premise-collision registry Attribution shift

Round by round, observed values only

Only the values observed in each round's document are stacked vertically. No interpretation or conclusion is attached — only how the values moved is left on record.

Round 1 · brief
first claim registered
intensity 0.41 conclusion repeated ×1 premise collisions 0
Round 2 · brief
same conclusion, evidence substituted
intensity 0.56 repeated ×2 substitutions 1 collisions 1
Round 3 · reply
attribution-label shift observed
intensity 0.63 repeated ×3 substitutions 2 label changes 1 collisions 2

Values above are display examples. SLDA counts and displays; what to call that movement, it does not judge.

We do not receive the original text at all

This is not a method of receiving personal data and then deleting it. It is a method of not receiving the original in the first place. Masking and labeling are performed by the user with their own AI, and the site receives only the result.

International standards

SLDA's data-processing design follows the GDPR (EU General Data Protection Regulation) principles of data minimisation, purpose limitation, and storage limitation, and is built toward alignment with the APEC CBPR (Cross-Border Privacy Rules) framework as a target standard.

Core principle: first-stage transformation (labeling, anonymization) occurs only on the user's own computer. The server receives only the completed labeled result; original text, real names, and identifiers never reach the server by any path.

Step Action Location of original text
1. Masking & labeling User masks real names and sensitive content; labels each sentence User's device only
2. Submission Only the labeled result is submitted to the site Original stays on user's device
3. Analysis SLDA engine scores the labeled data Original never transmitted
4. Report Scored result returned; labeled data discarded after analysis Original stays on user's device