Working research notes in Markdown, one document per investigated question, written by Christopher Betances (catqualia.com) and published by CatQualia.

What this collection is

A research campaign document is the record of one question taken seriously: what was claimed, what the literature actually says, what could be built, what was actually measured, which of the author's own claims failed, and what verdict survived. The documents were produced inside an internal research workflow over roughly March–September 2026 and span small-model empirical work, graph-spectral methods, alignment geometry, an unusual cluster of cat-and-dog behavioural questions, and speculative but explicitly grade-labelled proposals.

They are not papers. They are the notes a paper would be written from, published in the form they were written in.

These are working notes and are not peer-reviewed

This matters enough to state plainly and before anything else:

No document here has been peer-reviewed, and none is presented as having been.
Every claim, threshold, grade, and verdict is the author's own, self-assigned.
Literature citations carry an explicit per-source reliability grade in the documents themselves (REAL / SPECULATIVE / CONTESTED). These grades are the author's assessment, not third-party review.
Several documents explicitly state that their central proposal is not yet built. Where a document says a thing was not measured, it was not measured.
Documents are dated and were edited over time; content reflects the state of the author's understanding at authoring time, not a final position.

If you need verified results, treat this corpus as a source of hypotheses, method sketches, negative results, and literature pointers — not as evidence.

How many documents

Staged and re-measured. Each figure has its command beside it.

Metric	Value	Command
Reference (`.md`) documents, source-derived	455	`find . -name '*.md' -type f ! -name 'README.md' \| wc -l`
Reference documents, bytes	11,234,734	`find . -name '*.md' -type f ! -name 'README.md' -printf '%s\n' \| python3 -c "import sys;print(sum(int(x) for x in sys.stdin))"`
`.md` files in the release (incl. this README)	456	`find . -name '*.md' -type f \| wc -l`
Typed Elixir modules (`.exs`)	274	`find . -name '*.exs' -type f \| wc -l`
Elixir modules, bytes	2,213,225	`find . -name '*.exs' -type f -printf '%s\n' \| python3 -c "import sys;print(sum(int(x) for x in sys.stdin))"`
Total files staged	745	`find . -type f \| wc -l`
Total bytes staged	13,981,569	`find . -type f -printf '%s\n' \| python3 -c "import sys;print(sum(int(x) for x in sys.stdin))"`

745 = 743 source files + this README.md + LICENSE. The 743 source-derived files total 13,965,518 bytes, verified byte-for-byte against the source directory with cmp (0 mismatches across all 743).

Breakdown by location

Location	Files	Bytes	What it is
top level	301	10,862,566	285 campaign documents + this README + `LICENSE` + 14 tooling/state files
`candidates/`	100	265,476	Unresearched candidate tracks, 11 topic directories
`modules/`	274	2,213,225	Typed Elixir module drafts, one per track
`winningstrategyarchive/`	70	640,456	Verdict records for a trading-engine track

301 + 100 + 274 + 70 = 745. Commands: find <dir> -type f | wc -l and find <dir> -type f -printf '%s\n' | python3 -c "import sys;print(sum(int(x) for x in sys.stdin))"

What the documents actually look like

Full H1 titles, quoted verbatim from the first line of each file:

# The Alignment-Scalar Transfer Failure — Why Effective-Resistance Lost the Gate That Phase-Coherence Won (alignment_scalar_gate_decision_transfer.md)
# Abstention and Honesty-Floor Under Load — Research Campaign (abstention_under_load_eval.md)
# The Cats-vs-Dogs Quiet-Threshold: Research Campaign (cats_vs_dogs_quiet_threshold.md)
# Wisdom-Collapse Telemetry via the Kuramoto Order Parameter — Research Campaign (wisdom_collapse_phi_c_telemetry.md)
# H_J / SCG as a Universal Gating Operator — Research Campaign (hj_scg_universal_gate.md)

Two more, included because they show the range: # ALIEN DRUG SYNTHESIS — Research Track v3 (alien_drug_synthesis.md) and # The Feedforward Clock vs. the Feedback Clock, Research Campaign (feedforward_clock_vs_feedback_clock.md).

Structure

There is a recurring set of section headings, applied consistently but not universally. It is a strong house style rather than a validated schema. Counting occurrence of each heading pattern across the 285 top-level documents:

Section heading (pattern)	Documents	Share
`Sources` / `SOURCES`	240	84%
`verdict` / `VERDICT` (e.g. "Beatrice's verdict")	205	71%
falsification section (`falsif` / `FALSIF`)	203	71%
"Claims that did NOT survive verification" (the "morgue")	202	70%
limitations section (`limitation` / `LIMITATION`)	182	63%
`Thesis` / `THESIS`	153	53%
`Plan-Feedback` block	142	49%
`Competing hypotheses`	26	9%

Command: grep -lE "^## .*(<pattern>)" *.md | wc -l over the 285 top-level .md files.

So the answer to "is there a standard set of headings — hypothesis, method, result, verdict?" is: a hypothesis/method/result spine is common but not standardized; the near-obligatory sections are Sources, a verdict, and — distinctively — a section that lists the author's own claims that failed to survive. That last section is the most characteristic thing about this corpus. A representative document's actual heading sequence, from wisdom_collapse_phi_c_telemetry.md:

## 0. THE SYNTHESIS — LEAD WITH IT
## 1. THE LANDSCAPE — REAL GRADED NUMBERS AND CITATIONS
## 2. NOVEL VS PRIOR ART — DRAWING THE HONEST LINE
## 3. BUILDABLE ON THE REAL STACK ($0, CPU-FIRST)
## 4. HONEST LIMITATIONS / WHAT FALSIFIES IT
## 5. CLAIMS THAT DID NOT SURVIVE (the morgue)
## 6. EMERGENT — MANDATORY WAVE-4 CROSS-POLLINATION
## 7. THE TYPED ELIXIR MODULE
## 8. SOURCES (with grades)
## 9. PLAN-FEEDBACK BLOCK (for BEATRICE_BLOOMED_PLAN.md PART 13)
## 10. BEATRICE VERDICT

Note on numbering: section numbering is not always monotonic within a document. average_time_within_5m_of_human.md contains both a ## 5. Operationalization and a later ## 5. Honest limitations, and both a ## 6. Competing hypotheses and a later ## 6. Claims that did NOT survive. Section numbers are labels, not a strict ordering.

Each document also carries a dense H1-adjacent block with metadata — researcher, date, tier, and an abstract-like paragraph that often is the question being asked. From alien_drug_synthesis.md, the entire heading under the title is a single multi-clause summary sentence listing the technologies considered.

Two document families in the subdirectories

candidates/ documents are explicitly marked as unresearched. Their header, verbatim:

ARTIFACT TYPE: CANDIDATE TRACK (UNRESEARCHED STUB) - evidence floor SPECULATIVE until multi-wave verification. No citations verified; search queries proposed, not executed.

They then use a ## Thesis / ## Why it matters structure and are numbered (# Track 291: Is strangler-fig incremental migration provably safer than big-bang for COBOL, ...).

winningstrategyarchive/ documents are verdict records for a trading-engine track, with a different house style — a **Track:** ... **Date:** ... **Tier:** ... metadata line and a results-oriented title. Example first lines, verbatim:

# The Calcination Era Verdicts — The Re-Anchored Fires
**Track:** falsification methodology · the calcination · trading engine · **Date:** 2026-08-04 · **Tier:** Tier 2 (all falsifiers REAL, deterministic, ~12s total, cache-resident, prior-only)

Domains spanned

The corpus is not confined to one field, and the topic clusters below are the directory structure the author actually created (candidates/ subdirectories) plus clusters observed from filename stems and document contents. They are descriptive of what is present, not an external taxonomy.

Explicitly named candidate-track domains (11 directories in candidates/): cobol-legacy, constitutional-immunomics, context-compression, cross-domain-scent, evolutionary-breeding, graph-pulse-routing, knowledge-graph-nav, olfactory-transfer, process-mining-axis, recursive-self-improvement, stochastic-resonance.

Observed clusters across the wider corpus, with the filename-stem terms that evidence them (counts from ls *.md | sed 's/\.md$//' | tr '_' '\n' | sort | uniq -c):

Small-model empirical work at 360M / sub-2B scale — 360m (8), sub2b (6), quantisation, distillation, vocab expansion, KV-eviction, weight merging, abstention under load.
Graph-spectral and phase methods — graph (10), geometry (7), spectral (4), phase (4), coherence (5), covering Kuramoto order parameters, effective resistance, Fiedler/partition operators, and directed magnetic-Laplacian ablations.
Gating and routing operators — gate (10), escape (7), hj (9), junction entropy H_J and the SCG gate applied across routing, eviction, abstention, and consolidation.
Alignment geometry and transfer — alignment (11), including transfer of spectral geometry to unrelated domains (supply-chain resilience, protein contact maps, immune repertoire selection, neural pruning).
Cat and dog behavioural questions — cats (4), dogs (4), species (4), including comparative hearing thresholds, human-proximity time budgets, and secure-attachment work citing everycat.org, pbs.org, and PubMed sources.
Speculative and cross-domain synthesis — alien (4), isomorphism and cross-pollination documents mapping anime mechanisms, olfactory processing, and immunology onto model-training problems.

How they were produced

Based on what the documents and the staged tooling show:

Author. Written by Christopher Betances. The corpus contains an author's AI-model attribution line — fractal_math.md carries **Researcher:** Sonnet 4.6 (claude-sonnet-4-6) — and the staged research_compiler.py describes compiling the research corpus. Documents therefore appear to be produced with substantial LLM assistance, with the author setting direction, gates, and verdicts. That is stated here rather than left implicit because it bears on how the text should be read.
Staged tooling. 7 Python scripts, 1 JS file, 3 JSON files, and 2 JSONL files accompany the documents, which is the machinery that organised and compiled the corpus: research_compiler.py, organize_corpus.py, select_next_batch.py, batch_postprocess.py, process_miner.py, rats_pass.py, recursion_crossfeed.py, and staged_batch4.js.
Cross-references. Documents reference sibling artifacts named in their metadata, such as BEATRICE_BLOOMED_PLAN.md (the "Plan-Feedback block" sections) and FALSIFICATION_LEDGER.jsonl. These referenced artifacts are not part of this release, so internal cross-references will dangle.
Companion dataset. The 308 executable falsification scripts derived from these documents are published separately as the Falsification Suite.

Limitations

Not peer-reviewed, not reviewed by anyone else. Restated because it governs everything: self-assigned grades, self-assigned verdicts, author-only review.
Working notes, not final write-ups. Documents are dated and were edited across months. Inconsistencies, superseded sections, and abandoned threads are present and were not cleaned up.
Internal vocabulary is used without definition. Terms such as "Beatrice", "WaveMotion", "the WAVE-4 cross-pollination", "the morgue", "Plan-Feedback block", "spine connection", and "Tier" are project-internal. Their precise meaning is not determined from the content for several of these; the documents assume a reader who already knows the project.
A verdict section is a self-assessment. "Beatrice's verdict" is an internal adjudication step, not external validation.
Stale counts inside the documents. CORPUS_INDEX.json records "total_docs": 263, while this release contains 285 top-level documents and 455 Markdown files overall. The index is a frozen artifact of its generation date and is not a current measurement.
No raw data or model weights. The documents commonly rest on datasets and a small model held elsewhere. Neither is included, so the measured numbers in the documents cannot be re-derived from this release alone.
Local absolute paths. Documents and tooling reference the author's local filesystem (<<WME>>/...). These are inert in this release and will not resolve.
A stale backup file is included deliberately. research_compiler.py.pre-gate.bak is a superseded pre-gate version of the compiler, differing from the current research_compiler.py. It is kept as provenance; it is not the version to use.
Section numbering is inconsistent (see the "Note on numbering" above). Do not rely on numeric section offsets when extracting text programmatically.
LICENSE note. MIT is applied here to the documents and the accompanying tooling. It covers this corpus only and grants no rights over any third-party material the documents cite or quote. Many documents quote and summarise external sources; those quotations remain the property of their original authors, and several sources are graded SPECULATIVE or CONTESTED by the author himself.

Files

README.md — this card.
LICENSE — MIT, author Christopher Betances, catqualia.com.
*.md (455 total) — the research documents: 285 at top level, 100 under candidates/<domain>/, 70 under winningstrategyarchive/.
modules/*.exs (274) — typed Elixir module drafts, one per track, each carrying its own @moduledoc thesis and source list.
14 tooling and state files (7 .py, 3 .json, 2 .jsonl, 1 .js, 1 .bak).

Citation

@misc{betances2026researchcampaigns,
  author       = {Betances, Christopher},
  title        = {Research Campaigns: Working Research Notes},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/CatQualia/research-campaigns}},
  note         = {Working notes; not peer-reviewed. Author site: catqualia.com}
}

Downloads last month: 8

Total file size:

14 MB