kobaltGIT
Joined 1 month ago
Karma
13
Submissions
13
Member since
Sep 19
Recent submissions
Comment · 1 day ago
Thanks for the explanation! Actually, I don’t speak English at all, so I have to use a translator to read this thread and reply to people here. I didn’t realize that using a translator could make my posts look like AI-generated spam. From now on, I will use a simple, literal Google Translate without any editing of the text.
Comment · 1 day ago
Thanks everyone for the lively discussion, and huge thanks to the folks who emailed mods and vouched to clear the automated filter flag on my new account!
A few common points raised across the thread that I’d like to address in one place:
1. On "Alone", AI, and the Craft Anything methodology: Some folks dug into the repo, found the rules and chronicles, and pointed out this is heavily AI-assisted. They are 100% right — and I don't hide it (there is an open public dev chronicle in docs/ detailing every single step). By "alone" I meant no venture backing, no university grant, and no human engineering team.
However, this isn't chaotic one-shot prompting. StatCraft is built as part of the ArtCraft suite, following the philosophy and methodology of Craft Anything (https://github.com/storytold/craft-anything - "understand how any software works, and craft your own"): using evidence-backed, clean-room investigation to replace closed, predatory software with open, reliable tools. We pair an orchestrator with specialized AI agents, constrained by strict compiler invariants and external mathematical oracles. That's also why StatCraft has a native Model Context Protocol (MCP) server: the software itself is designed to live natively in an agent-first world where researchers can interact with deterministic tools via natural language.
2. On "100% Safe Rust" terminology: Fair critique from the Rust community on the phrasing: what I technically meant is #![forbid(unsafe_code)] across the workspace, zero unsafe blocks, and a zero-panic invariant (Result<T, E> everywhere, zero unwrap/expect in non-test code). The compiler ensures memory safety invariants, while domain correctness is enforced via fuzzing and mathematical oracle tests.
3. On Clean-Room & Legal concerns: No IBM binaries were reverse-engineered or decompiled. The .sav and .zsav codecs follow open specifications established by GNU PSPP and the MIT-licensed ReadStat library. The statistical procedures are implemented strictly from academic literature and public formula specifications (such as the SPSS 24 Algorithms Guide), validated against SciPy with <= 1e-6 precision.
4. "Why not just use R/Tidyverse?": R and Python are wonderful for coders. But thousands of students, sociologists, psychologists, and clinicians are trained purely on SPSS's GUI. When their university license expires, they face a $1000+/year paywall from IBM and have legacy .sav files they can't easily open without proprietary software. StatCraft is built for them: a free, native GUI tool that respects their muscle memory, works offline, and never crashes on messy survey data.
If you have old, weird, or corrupted .sav files, or want to audit a statistical procedure against your own SPSS outputs, issues and PRs are very welcome on GitHub.
Comment · 1 day ago
It's not about memory safety per se. It's about not losing an hour of analysis because a malformed .sav file or a zero-variance sample triggered a segfault or silent wrong output.
SPSS crashes on bad data. Researchers work with messy data. Rust's Result type makes "handle every error explicitly" the default, not an afterthought. That's why I chose it.
Comment · 1 day ago
I'm building StatCraft — an open-source statistical analysis package in 100% Safe Rust,
as a clean-room alternative to IBM SPSS Statistics.
I started this because I believe statistical software should not crash on malformed data, should be scriptable by AI agents via MCP, and should not cost thousands of dollars per year.
Current state (11 days in, solo):
- 14 statistical blocks audited against SPSS 24 Algorithms Guide with ≤1e-6 SciPy precision
- Full .sav/.zsav codec, Excel/CSV import, SPSS syntax parser (.sps)
- 37 SPSS-style dialogs, pivot tables, Output Viewer with TSV export
- Built-in MCP server for Claude/ChatGPT/Ollama agents
- Zero unwrap/expect/panic in production; fuzzed on corrupted inputs — 0 panics
What I need help with:
- Testers with real SPSS datasets (the weirder the encoding, the better)
- Rust developers for the expression evaluator (3 known precedence gaps)
- Statisticians to validate output against their SPSS workflows
- Anyone who wants to port a procedure (regression, factor, reliability are in)