EDA Desk - notice
=================

edakit.js is a JavaScript port of three command-line tools, and the helpers they import, from the
"exploratory-data-analysis" agent skill in k-dense-ai/scientific-agent-skills
(skills/exploratory-data-analysis/scripts, commit 49c6e97775eaa18ba791bebe23162a70ae601c18, skill
metadata version 1.2):

  tabular_profile.py, missingness_leakage_audit.py, distribution_sensitivity.py,
  and the parts of _tabular.py, _common.py and _capabilities.py they call.

The skill is published under the MIT License, Copyright (c) 2025 K-Dense Inc.; the licence text is in
LICENSE-exploratory-data-analysis-skill.txt. The port keeps the original messages, bounds, sampling
rule (BLAKE2s-prioritised heap), tokens, formulas and JSON layout.

How it was checked. The Python scripts were run in-process under CPython 3.12.13 on fuzzed CSV/TSV
files (valid and malformed: quoting, CR/LF/CRLF line ends, blank lines, BOMs, invalid UTF-8, Unicode
digits and case folding, missing codes, huge and subnormal values, 500-5,200-row files that exercise
the bounded samples, role and column arguments that are missing or duplicated) and compared with
edakit.js run by Node on the same bytes and arguments. On the shipped edakit.js (sha256 defe3a0b...)
5,000 fuzzed runs matched in exit code, stderr and stdout, including 2,085 refusals (exit 2) and 201
runs where the Python stops with an uncaught exception (exit 1: overflow in pow or fsum), which the
page reports as a crash rather than inventing a report. An earlier 2,500-run batch on the previous
revision also matched. datetime.fromisoformat (C implementation, 3.12 behaviour) was compared on
generated strings: one difference in the first 60,000 (a zero-second UTC offset carrying
microseconds is plain UTC in CPython) was fixed, then 350,000 more matched with no difference.

One stated difference. The three skewness fields call pow() and log1p() in CPython, which defer to
the platform libm; macOS libm is not correctly rounded for pow(x, 2) in about 0.2% of inputs or
log1p in about 0.03%, and V8's Math functions differ again. The page computes x**2, x**3, v**1.5 and
log1p correctly rounded (BigInt arithmetic). In the 5,000-run batch 8 skewness values differed from
CPython-on-macOS in the last digits (all within 1e-12 relative); everything else was byte-identical.
The Python output itself varies by operating system in these fields.

Known limits of the port: a file that is not valid UTF-8 is refused before any row is read (the
Python decodes lazily and can process rows before the bad bytes in files larger than its read
buffer); Unicode "printable" and case-folding tables follow the browser's Unicode version, which can
differ from CPython 3.12's (Unicode 15.0) for newly assigned characters; CPython 3.14 parses
"24:00" in time values and 3.12 does not - the page follows 3.12.

fixkit.js (the resplit by entity or group, the plain-CSV rewrite of semicolon exports, the missing-code
suggestions and the since-last-audit counts) is this page's own code, not part of the skill. The
resplit hashes each entity or group value with BLAKE2s and assigns it by the current splits' row
shares; it ignores time.

The example datasets (clinic visits, bioreactor runs) were generated for this page with a seeded
random generator. They describe no real people and no real plant.

The skill asks that, if it materially contributes to a manuscript, report, presentation or code
release, the following paper be cited (record checked at arxiv.org on 2026-09-27: v2, revised
2 September 2026):

  Kassis, T., Agarwal, V., He, Y., Patel, D., & Brueckner, A. M. (2026). Scientific Agent Skills:
  A Library of Procedural Knowledge for Research Agents. arXiv:2609.00065.
  https://doi.org/10.48550/arXiv.2609.00065