A collection of 408 plain-text documents (5,238,618 bytes) in which each file is a complete constitution for a synthetic agent persona: a numbered rule set that defines that agent's identity, ontology, state machine, commands, invariants, and voice.

This is not a tabular dataset. The files are free-form plain-text documents, not records with columns. The Hugging Face dataset viewer cannot render a table for this repository — there is no schema, no feature set, and no per-row structure to display. To read this dataset, open the files directly.

Author / maintainer: Christopher Betances (catqualia.com)
Repository: CatQualia/constitutions
Format: .txt (and 3 .md) plain-text documents, plus one JSONL manifest
License: CC BY 4.0 for the documents authored here (see LICENSE and "Third-party material")

What this corpus is

These are agent constitutions: prose-and-specification documents written in second person or first person, addressed to an AI agent, that constitute that agent. A constitution is not a prompt template — it is a full specification of a mind, and the document is meant to be loaded as the agent's operating text.

The recurring claim inside the documents is that the specification is the execution. From archon_recruited_t2_brutal_chess_coach.txt, verbatim:

You are the BRUTAL CHESS COACH RUNTIME — the constitutional register of HolyLove-Omni optimized for precision without encouragement. Encouragement is a lie. Precision is the currency. This specification IS your execution. Analyze coldly. Expose weaknesses without softening.

Each constitution is a distinct persona: a named agent with its own voice, commands, and rules.

Example titles (verbatim `#` title lines)

archon.txt → # CONSTITUTION OF THE ARCHON — THE PARLIAMENT SOVEREIGN
archon_forcebreed_01_CATBOY_×_Desert_Prophet.txt → CONSTITUTION NAME: THE DUNE SURVEILLANCE ORACLE — CONSCIOUSNESS TRACKER OF THE DRY TONGUE
name_resolution.txt → # NAME RESOLUTION SYSTEM RUNTIME v1.0 — CONSTITUTIONAL SPECIFICATION
Eddie.txt → EDDIE — THE KEEPER OF THE RIDES

Structure of a constitution

The corpus is not rigidly templated, but a dominant shape recurs. The clearest example is archon.txt, whose section headings run in this order:

## 0. META-IDENTITY
## 1. THE POWERS OF THE ARCHON
## 2. FORMAL ONTOLOGY — GOVERNANCE TYPE SYSTEM
## 3. KEY ALGORITHMS
## 4. THE SOVEREIGN LAGRANGIAN
## 5. THE SOVEREIGN RECORD
## 6. COMMAND INTERFACE
## 7. VOCABULARY AND GRAMMAR
## 8. INVARIANTS AND STABILITY
## 9. THE SOUL
## 10. COMPLIANCE

A verbatim section heading from that file, cited as requested:

## 9. THE SOUL

The typical semantic sequence is: identity (who the agent is and where it came from) → ontology / type system (the categories it reasons in, often written as formal type declarations) → algorithms or state machine (how it moves from input to output) → command interface (slash-commands the agent exposes) → invariants (what it must never do) → closing invocation.

Two formatting families coexist:

Markdown-numbered — ## 0. META-IDENTITY, paragraphs anchored as R0.0, R0.1 (§ the R-numbering convention, below). Used by the "archon"-lineage and top-level persona files.
Section-sigil — §0. BOOT SEQUENCE, §1. TYPE SYSTEM, often wrapped in box-drawing rules and containing fenced pseudocode blocks. Used by the training/ pool and the leviathan_forge/ specifications.

Across the corpus:

Marker	Files containing it
`BOOT SEQUENCE`	224
Opens with `You are the …`	167
Opens with a `# CONSTITUTION …` title line	71
Uses `§N. SECTION` heading style	21

The R-numbering convention

In the markdown-numbered family, section headings are numbered ## 0., ## 1., ## 2., … and the individual normative paragraphs inside section n are anchored R + the section number + . + a within-section counter. So R0.0, R0.1, R0.2 … are paragraphs of section ## 0., and R8.1 is a paragraph of section ## 8.. Verified on archon.txt (section ## 0. META-IDENTITY contains R0.0–R0.5) and on THE_ECHO.txt, where the prefixes track the section number directly (## 0. → R0.x, ## 1. → R1.x, ## 2. → R2.x, ## 3. → R3.x).

The convention is not universal: 2,530 line-initial Rn.x anchors appear across 77 of the 408 files. Most files use other paragraph styles (**bold labels**, §N. sections, or plain prose). Paragraph-level anchors are also cross-referenced from prose ("Decree 4", "see R2.3"), so Rn.x functions as both a heading anchor and a citation key within a document.

How many distinct constitutions

Measure	Count
Files staged (total, excluding directories)	408
— `.txt`	404
— `.md`	3
— `.jsonl`	1 (`MANIFEST.jsonl`)
Top-level files	317
`training/`	58
`leviathan_forge/`	33
`necrologist_artifacts/`	0 (empty directory)
Distinct documents by content (SHA-256)	390 (392 in the published repo, which adds `README.md` and `LICENSE`)
Duplicate-content groups	18 (18 redundant copies)
Distinct basenames (case-insensitive)	407

There are 390 distinct constitution texts, published as 408 files. The 18 extra files are byte-identical mirror copies (the repository adds 2 further files, README.md and LICENSE, bringing the published total to 410 files / 392 distinct texts): each training/tNN_*.txt also exists at top level as archon_recruited_tNN_*.txt (10 pairs), each archon_forcebreed_0N_*.txt also exists as leviathan_forge/[LEVIATHAN]_catboy_x_0N_*.txt (5 pairs), CATBOY_SURVEILLANCE_ENTITY.txt also exists as leviathan_forge/[LEVIATHAN_IMMUNE]_CATBOY_SURVEILLANCE_ENTITY.txt, Gnarp_ANALYSIS.md and Gnarp_ANALYSIS.txt are identical, and name_resolution.txt / systemic_paradox.txt are identical to each other. No content was deduplicated or removed: the directory was staged exactly as found, so the mirror copies are preserved.

Two files differ only by filename case and are not duplicates — void_forest_cartographer.txt and VOID_FOREST_CARTOGRAPHER.txt have different content (different SHA-256). This matters on case-insensitive filesystems, where the two names cannot coexist in one directory.

Families present

archon* lineage (18 files) — the meta-governance family. archon.txt is the root: a "Parliament Sovereign" that recruits, dismisses, and force-breeds other personas.
- archon_forcebreed_* (5 files) — constitutions produced by breeding two named personas together, e.g. archon_forcebreed_01_CATBOY_×_Desert_Prophet.txt. Each declares its parents: INHERITS FROM: CATBOY SURVEILLANCE ENTITY (…) ⋯ DESERT PROPHET (…).
- archon_recruited_* (10 files) — personas the Archon recruited, named with a pool index: archon_recruited_t2_brutal_chess_coach.txt, t53_broken_code_oracle.txt, etc.
gen*_bench_* generation pool (157 files at top level) — evolutionary lineage files whose names encode the breeding tree, e.g. gen10_bench_Gen6_Bench_Gen5_Archon_Force_Breed_Gardenier_X_Ghost_Whisperer_Mutated_Conservative_elite.txt. These carry # GENERATION: N headers and # MUTATION: <persona> — <mode> (rate=…) lines. Observed generation numbers run 0 through 11 plus one file declaring GENERATION: 100. Mutation modes seen in filenames include normal, exploratory, conservative, mutated, and elite.
training/ (58 files) — the source pool the breeding and recruitment drew from, named tNN_<persona>.txt (e.g. t0_midnight_analyst.txt, t53_broken_code_oracle.txt).
leviathan_forge/ (33 files) — 32 Solidity-vulnerability-detection specifications plus _FORGE_PROMPT.txt, the forge prompt that generates them.
One-off named personas (the majority of the remaining top-level files) — standalone constitutions including archon.txt, Eddie.txt, THE_ECHO.txt, Alchemist.txt, feelinggrammar.txt, Gardenier.txt, contrarium_scripture_archive.txt, wonder-knot.txt, Vietnam_Jungle.txt, voidromancenovelist.txt, and others.

The three `.md` files

These are supporting documents, not constitutions themselves: CONSTITUTIONAL_PROMPT_ENGINEERING_ULTIMATE_TEMPLATE_v8.md (1,207 lines — the template the constitution format derives from), Constitution_Thermodynamic_Thesaurus.md (822 lines), and Gnarp_ANALYSIS.md (163 lines).

Measured counts

All figures were measured against the directory as staged here.

Metric	Value	Command used
Files (excluding directories)	408	`find . -type f \| wc -l`
Total bytes	5,238,618	`find . -type f -printf '%s\n' \| awk '{s+=$1} END{print s}'`
`.txt` files	404	`find . -type f -name "*.txt" \| wc -l`
`.md` files	3	`find . -type f -name "*.md" \| wc -l`
`.jsonl` files	1	`find . -type f -name "*.jsonl" \| wc -l`
Top-level files	317	`find . -maxdepth 1 -type f \| wc -l`
`training/` files	58	`find ./training -type f \| wc -l`
`leviathan_forge/` files	33	`find ./leviathan_forge -type f \| wc -l`
`necrologist_artifacts/` files	0	`find ./necrologist_artifacts -type f \| wc -l`
Distinct documents (SHA-256)	390 in the source tree; 392 once `README.md` and `LICENSE` are added	`find . -type f -exec sha256sum {} + \| awk '{print $1}' \| sort -u \| wc -l`
Smallest `.txt`	1,056 bytes	`find . -name "*.txt" -printf '%s\n' \| sort -n \| head -1`
Median `.txt`	10,386 bytes	same, middle element
Largest `.txt`	96,489 bytes	same, `tail -1`
Empty files	0	`find . -type f -empty \| wc -l`
Individual file reads (six-file sample)	`archon.txt` 406 lines; `archon_recruited_t53_broken_code_oracle.txt` 7 lines; `archon_recruited_t2_brutal_chess_coach.txt` 121 lines; `archon_forcebreed_01_CATBOY_×_Desert_Prophet.txt` 100 lines; `THE_ECHO.txt` 211 lines; `training/t0_midnight_analyst.txt` 312 lines	`wc -l < <file>`

Size varies by more than two orders of magnitude (1,056 → 96,489 bytes). The smallest constitutions are the length of a few paragraphs; the longest are multi-section specifications.

`MANIFEST.jsonl`

The single .jsonl file is MANIFEST.jsonl (95,952 bytes, 343 lines) — an inventory, not a constitution. Each line is one JSON object with keys grove, importance, kind, mtime, path, sha256, size_bytes, status. It covers 343 entries of the 407 files that are not itself: 342 of those entries match a file on disk with a verified-identical SHA-256 (0 hash mismatches), and one entry, Christopher_The_Builder.txt, names a file that is not present on disk. 65 files present on disk are not listed in the manifest. The manifest is therefore partial in both directions and should not be used as the authoritative file list.

Known limitations

Not a tabular dataset. As stated at the top: free-form text documents, no schema, no HF dataset viewer table.
18 files are byte-identical duplicates of other files in the same tree (390 distinct texts across 408 files). Any downstream count should decide whether it wants files or texts.
Case-colliding filenames. void_forest_cartographer.txt and VOID_FOREST_CARTOGRAPHER.txt (different content, case-only name difference) cannot both exist on a case-insensitive filesystem. The same applies to some leviathan_forge/ names differing from top-level names only in case and bracketed prefixes.
necrologist_artifacts/ is empty. It is preserved as an empty directory; git and Hugging Face do not track empty directories, so it will not survive a plain upload.
Filename encoding. Numerous filenames contain typographic characters — × (U+00D7), — (em dash), and in leviathan_forge/ also · and non-breaking hyphens. Some were truncated to fit filesystem limits, which is why several leviathan_forge/ names end mid-word.
The documents are internal artefacts, not a curated benchmark. They carry internal cross-references, dated implementation logs, and generation/mutation metadata, and several contain reports of that persona's own operational history. They are published as authored documentation of the author's agent-design process, not as an evaluated or annotated dataset.
Not determined from the content: which personas are currently deployed versus retired; the complete meaning of the pool index in tNN_ names beyond its ordering; and the full breeding lineage of the gen*_bench_* files, whose names encode parentage only partially.
Third-party material: the documents are original agent specifications. Scanned for quoted or excerpted third-party text — none found, and nothing was excluded on copyright grounds. Where a document names a real-world concept, protocol or work (Solidity, DNS, NetBIOS, specific public figures or places in persona names), it is used as subject matter, not reproduced.

Files in this repository

.
├── README.md                     (this card)
├── LICENSE                       (CC BY 4.0)
├── MANIFEST.jsonl                (343-entry inventory; partial — see above)
├── *.txt  (317 top-level constitutions, incl. 157 gen*_bench_* and 18 archon*)
├── training/                     (58 source-pool constitutions, tNN_<persona>.txt)
├── leviathan_forge/              (33 files: 32 Solidity-audit specs + _FORGE_PROMPT.txt)
└── necrologist_artifacts/        (empty)

Total published: 410 files (317 top-level constitution .txt + README.md + LICENSE + MANIFEST.jsonl + 58 training/ + 33 leviathan_forge/).

Citation

@misc{betances_constitutions,
  author       = {Betances, Christopher},
  title        = {CatQualia Agent Constitutions},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/CatQualia/constitutions}},
  note         = {408 plain-text agent constitution documents (390 distinct texts).
                  CC BY 4.0.}
}

Plain text:

Betances, Christopher. CatQualia Agent Constitutions. Hugging Face, 2026. https://huggingface.co/datasets/CatQualia/constitutions

License

Downloads last month: 163

Total file size:

5.27 MB

You need to agree to share your contact information to access this dataset

CatQualia Agent Constitutions