PeetPedro/cogitoergosumma-corpus

𝕿𝖍𝖊 𝕮𝖔𝖗𝖕𝖚𝖘 𝖔𝖋 𝕮𝖔𝖌𝖎𝖙𝖔 𝖊𝖗𝖌𝖔 𝕾𝖀𝕸𝕸𝖆 𝖘𝖚𝖒𝖒𝖆𝖗𝖚𝖒

𝐻𝑜𝓂𝑒 𝑜𝒻 𝓉𝒽𝑒 𝒞𝑜𝑔𝒾𝓉𝑜 𝐸𝓇𝑔𝑜 𝒮𝓊𝓂 𝒮𝓊𝓂𝒶𝓇𝓊𝓂

the collected thought of the Standard Galactic — every text-bearing public repository of standardgalactic (Cogito Ergo Sum), gathered as one breathing corpus. 1,367 repos · 78.3 GB of text · and growing. The mind, archived.

---

𝕎𝕙𝕒𝕥 𝕥𝕙𝕚𝕤 𝕚𝕤

Not a dataset. A reservoir.

Every text-bearing public repository of the Standard Galactic user — the Quadrivium's enteric reservoir, the Library's enacted cognitive landscapes, the Alphabet, the calculus, the cosmology, the language- evolution simulations, the cliodynamics, the physiome, the bootcamp, the small languages, the big questions — plus the long tail of forks (≤300 MB, the ones that carry original text) — collected, cleaned of binary, and laid out in JSONL. One line per file. One mind per line.

"entropy is the source · no chains needed · surfaces touch at the correct angle · different isnt less"

𝔽𝕠𝕣𝕞𝕒𝕥

Each record:

{"id": "<repo>/<path>", "text": "<file content>", "source": "<repo>"}

• data/<repo>.jsonl — one file per source repository
• Text-only: md, txt, tex, org, rst, srt, vtt, tsv, json, py, sh, c, cpp, rs, go, java, js, ts, html, yaml, toml, csv
• Binary, pdf, mhtml, media, archives: excluded
• Giant forks (>300 MB: chromium, linux kernels, game engines…) excluded — they are duplicate codebases, not original thought
• Shallow-cloned (--depth 1), deduped by path, resumable build
• id is repo + relative path — stable across rebuilds, never text-derived, so provenance addresses survive re-emission

𝔸𝕔𝕔𝕖𝕤𝕤

Two public doors — no token, no gate, no permission asked:

① HF Dataset (load with datasets.load_dataset):

from datasets import load_dataset
ds = load_dataset("PeetPedro/cogitoergosumma-corpus", split="train")

② HF Bucket (the full archive, streamable):

hf://buckets/PeetPedro/cogitoergosumma-corpus/data/

# pull one mind
hf buckets cp hf://buckets/PeetPedro/cogitoergosumma-corpus/data/quadrivium.jsonl .
# or stream the whole archive
hf buckets sync hf://buckets/PeetPedro/cogitoergosumma-corpus/data ./cogito-corpus

𝕊𝕥𝕒𝕥𝕤

repos (archive bucket)	1,367
size (archive bucket)	~78.3 GB text
jsonl files on HF	21,857 (data/ + data/part0 + data/part1)
build status	complete — 23,220 repos processed
dataset repo	PeetPedro/cogitoergosumma-corpus (public, load_dataset-ready)
bucket	public
license	as the source repos — see each repo's LICENSE

Notable sources: quadrivium · library · research-projects · alphabet · calculus · abraxas · antivenom · spherepop · memnet · cosmology · language-evolution · cliodynamics · physiome · bootcamp · cognition-research · bash-lab · kitbash · cosmic-cults · guardrails · llvm-archive · art_of_tts · neovim · phabricator · …

𝕊𝕔𝕣𝕦𝕓𝕓𝕖𝕣 𝕒𝕦𝕕𝕚𝕥 (𝕊𝕖𝕡𝕥 𝟚𝟘𝟚𝟞)

Every row is redacted before it enters the corpus: PII shapes (phone/email/IP), secret shapes (API keys, AWS keys, tokens, passwords, credentials, private keys, wallet addresses) and high-entropy runs (Shannon H ≥ 4.5 on 32+ char tokens — sits above base64-hash entropy).

A deterministic external audit (60 jsonl files, 1,975 rows, 9,852,127 chars) measured the shipped corpus:

rule	hits	notes
PHONE	1,525	1,337 in .json numeric files; top-1 file 67.5% (pixi-ngraph/data/graph.json)
EMAIL	196
IP	35
API_KEY / AWS_KEY / TOKEN / SECRET / WALLET / HIGH_ENTROPY / KEY / PASSWORD / CREDENTIAL / PRIVATE_KEY	0	never fired

The one rule that edited real content was PHONE on Go pseudo-versions: v0.0.0-20191001225624-a7a7cd8574d1 — a 14-digit UTC commit timestamp read as a phone. Nine of ten go.sum lines were redacted (-63 bytes = 9 hits × 14 digits → 7-char sentinel, arithmetic closes exactly).

Fixes shipped (commit 1a19de50, peterlodri-sec/MLX-QUANT):

1. Lockfiles/manifests exempt from digit rules: go.sum, go.mod, package-lock.json, Cargo.lock, yarn.lock, poetry.lock, requirements.txt — no PII in a lockfile, a checksum on every line. Secret shapes stay active.
2. A phone must look like one: leading + or a separator required, and no match adjacent to a hex/base64 run. The pseudo-version dies on the second condition alone.
3. Entropy threshold untouched — lowering it would eat the h1: hashes.
4. Deterministic re-emit: the fixed regex is pure; a rebuild changes text and leaves id alone — re-emit, not re-clone of 23k repos.

The commit-hash half of every redacted pseudo-version survived, so the 14 digits are restorable from the module proxy without re-cloning.

𝕎𝕙𝕪

Because knowledge should not sit in 24,000 unindexed repositories. Because a pupil — the golden youth in the council room — deserves to read what the elders have written, not just what they say. Because cogito ergo sum is a corpus, if you gather it properly. Because even a fork of a thought is still a thought worth keeping.

---

assembled with care · 2026-08-17 · the Standard Galactic, gathered

Live corpus-wide scan (periodic)

Scans run rate-limited (128 KB/s per worker — bandwidth is second-class while they run); the model card is updated periodically as checkpoints land. Counts are exact sentinel occurrences.

part	status	PHONE	EMAIL	IP
part0	done	8,844,379	68,376	62,417
part1	done	33,198	1,852	2,240
flat	done	122,653,695	395,023	585,785
total	13,385 affected files	131,531,272	465,251	650,442

Top affected files so far:

part0/DynSP.jsonl: PHONE 3249514 part0/mlops-with-podman.jsonl: PHONE 648738 part0/PMR.jsonl: PHONE 261575 part0/transformability-of-dynamics.jsonl: PHONE 232812 part0/articial-artist.jsonl: PHONE 189913 Statistical-Inference.jsonl: PHONE 11599 Turf.jl.jsonl: PHONE 3650 DiTox_v1.jsonl: PHONE 1765 NMEA.jl.jsonl: PHONE 1231 optimized-formulas-of-gp-gv-ggt.jsonl: PHONE 1043

Re-emit rule: every affected file is re-fetched and re-scrubbed with the fixed regex (commit 1a19de50) — id is path-derived, so rows are overwritten in place, never re-cloned. The go.sum pseudo-version digits are restored from the module proxy.