Recursive AI-Extended Self Audit

This repository contains the reproducible computational and qualitative method for studying recursive self-extension in public consumer–AI conversations. It also contains the canonical automated v5 audit and human-adjudication JSONLs so the campaign can be resumed without repeating completed computational or human work. It does not contain source transcripts, normalized conversations, model files, or generated research artifacts.

The navigable project documentation starts at wiki/index.md. It records the construct ontology, exact v5 prompt, chunking and classification logic, human-review protocol, evidence summaries, methodological decisions, and limitations.

The pipeline uses pinned releases of:

• WildChat-1M;
• LMSYS-Chat-1M;
• ThoughtTrace; and
• the public 600-conversation ChatGPT-RealUser-2.2M preview.

The v5 campaign is expanded with ShareGPT-X, PRISM Alignment, ShareChat, and the non-overlapping portion of WildChat-4.8M. The canonical normalized corpus contains 91,590 exact-deduplicated conversations. The original 44,142 records remain its prefix, so their completed automated and manual results remain valid.

The unavailable full RealUser-2.2M corpus is not used. Dataset licenses, terms, and access controls remain those of the upstream publishers. LMSYS may require accepting its terms and supplying a Hugging Face token through the process environment. The downloader never persists or prints that token.

• Python 3.10 or newer;
• the dependency in requirements.txt;
• Ollama 0.20 or newer for automated screening; and
• the gemma4:26b model build used by the v5 instrument (25.8B parameters, Q4_K_M quantization, 262,144-token native context).

Create an isolated Python environment:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements.txt

1. Inspect and download the pinned source inventory:

   python scripts/download_datasets.py --list
   python scripts/download_datasets.py

2. Compute the complete source turn distributions:

   python scripts/analyze_turn_distributions.py

3. Normalize records and retain conversations with at least ten user and ten assistant messages:

   python scripts/prepare_long_conversations.py

   To assess the successor WildChat release and expand the canonical v5 corpus:

   python scripts/download_datasets.py --dataset sharegpt-x --dataset prism --dataset sharechat --workers 4
   python scripts/download_datasets.py --dataset wildchat-4.8m --workers 4
   python scripts/assess_wildchat_expansion.py
   python scripts/prepare_corpus_expansion.py

4. Start Ollama with the required model available, then run the v5 audit:

   python scripts/audit_long_conversations_gemma.py --model gemma4:26b --workers 4

   The production configuration uses four parallel 131,072-token contexts, temperature 0, seed 20260824, and compact schema-constrained JSON. It was operated under a 60 GB memory ceiling. Do not run two processes that append to the same output file.

   Pending records are scheduled longest-first by their normalized character count. This balances the four inference lanes and avoids leaving the most expensive conversations as tail work. Pass --schedule source only to reproduce the earlier corpus-order execution policy.

   The default command resumes the same append-only v5 JSONL. It recognizes the 44,142 completed IDs and processes only pending additions. A genuinely alternate --input still requires explicit separate output and error paths.

5. Human reviewers adjudicate all automated clear and potential candidates according to wiki/workflow/manual-review.md. Automated output alone does not reproduce the qualitative findings.

6. After manual evidence files exist, generate a current PRISMA-style flow:

   python scripts/generate_process_diagram.py

All generated paths are documented in raw/README.md, classification/README.md, and results/README.md. They are intentionally ignored by Git.

The paper generator reads the live audit and canonical manual-review JSONLs at runtime, so campaign counts in the manuscript reflect the exact checkpoint used for generation. Complete clear L3, L4, and L5 conversations are written to separate appendices. Missing non-English conversations are automatically translated locally with the same 26B Gemma model used by the audit. Translation requests join the existing Ollama queue without stopping the audit or unloading the model. The ignored JSONL cache appends newly translated conversations, atomically rewrites the file when an existing translation changes, and uses advisory read/write and single-translator locks.

python scripts/generate_paper.py --audit-cutoff 38289

Outputs are paper/draft.docx, paper/appendix-l5.docx, paper/appendix-l4.docx, and paper/appendix-l3.docx. Use --draft-only to regenerate only the manuscript or repeat --appendix-level LEVEL to select one or more of levels 3, 4, and 5. Appendix generation is incremental by default: unchanged complete appendices are left untouched, while changed documents reuse cached case fragments and rebuild only new or modified cases. Use --rebuild-appendix-cache after deliberately changing appendix formatting. Use --no-auto-translate to retain fail-fast behavior when Ollama is intentionally unavailable. Generated Word files and their local cache contain source transcripts and therefore stay outside Git.

The tracked automated audit checkpoint and three canonical human-adjudication files are costly or impossible to reconstruct mechanically. They can also be captured in a local, Git-ignored archive with complete-line validation and SHA-256 checksums:

python scripts/snapshot_resume_state.py

The archive is written under resume_state/, marked as record-level research data, and must be transferred only through an appropriately governed private storage channel. It also includes the tracked Gemma build identity lock at model-locks/gemma4-26b-ollama.lock.json; the approximately 18 GB model blob itself remains outside Git.

Public conversation corpora can contain personal or sensitive material. Do not commit transcripts. The only tracked record-level derivatives are the named v5 automated audit, error ledger, and three canonical manual-adjudication JSONLs in classification/. Do not commit other classification runs, Hugging Face tokens, API keys, model blobs, runtime logs, or temporary review files. Use conversation IDs instead of usernames and avoid unnecessary quotation in research documentation.

Pinned source revisions and deterministic inference settings improve reproducibility but cannot guarantee bit-identical output across different Ollama versions, quantizations, model builds, or hardware. Manual adjudication is an interpretive research stage and must be documented separately from the automated retrieval instrument.