# Experiment #25: Why the paper corpus stopped growing

**Date:** October 04, 2026
**Model:** none for discovery (LLM analysis not run)

## Hypothesis

`data/corpus.json` stopped at 2026-05-05 because `paper_scout` was not on the daily GitHub Actions workflow. `data_fetcher` was still refreshing the 20 homepage papers. The gap was not an empty PubMed.

## Method

Read `data/pipeline_log.json` and `data/corpus.json` before changing them. The last paper-scout entry is 2026-05-05. `.github/workflows/daily-update.yml` called `preprint_scout` and `build_site.py --refresh-data` (which runs `data_fetcher`) and did not call `paper_scout`. Then ran `python src/agents/paper_scout.py --discover-only`, which searches PubMed and saves titles without calling Ollama. The lookback was 154 days so June through September were not skipped. The first query's detail fetch returned HTTP 429 and was dropped; `src/literature_agent.py` now retries 429 and sends `NCBI_API_KEY` when that secret exists. A second discover-only pass picked up the dropped query.

## Result

- Corpus size before this run: 86 papers, last_run 2026-05-05T15:34:46.
- First pass: 47 new papers, one query lost to HTTP 429.
- Second pass after the retry: 23 additional papers.
- Corpus size after: 156 papers. last_run 2026-10-04T17:36:49.877464. Mode discover_only.
- LLM analyses added: 0. Ollama is not running here and `GEMINI_API_KEY` is not set, so analysis stayed `not_analyzed` instead of a fabricated finding.

Homepage papers in `data/data.json` were already updating through October 4 via `data_fetcher`. The stale object was the scout corpus, which chat and later experiments read.

### Scripts that still have no saved run

These files exist and were not executed. Each now has a published blocked account (EXP-012 through EXP-023) that records the missing model and computes no findings. They still need a model backend before any science run (`OLLAMA_BASE_URL`, or `GEMINI_API_KEY` with `HD_LLM_BACKEND=aistudio`):

- `run_experiment_12_kynurenine_pathway.py`
- `run_experiment_13_complement_neuroinflammation.py`
- `run_experiment_14_cognitive_pharmacotherapy.py`
- `run_experiment_15_sema4d_neuroinflammation.py`
- `run_experiment_16_mitochondrial_membrane.py`
- `run_experiment_17_pde10a_striatal_rescue.py`
- `run_experiment_18_glp1r_neurodegeneration.py`
- `run_experiment_19_mpc_isr_inhibition.py`
- `run_experiment_2.py`
- `run_experiment_20_ferroptosis_iron_chelation.py`
- `run_experiment_21_cgas_sting_innate_immunity.py`
- `run_experiment_22_adenosine_a2a_msnprotection.py`
- `run_experiment_23_thiamine_biotin_tca_rescue.py`

## Limitations

- Discovery stored abstracts only. Several queries are broad, so some saved titles are not HD-specific; they were not deleted and were not re-scored.
- `NCBI_API_KEY` was not set. Retries were enough for this run.
- Not medical advice.

updated by running the scout and recording the 429. The note that scripts 12-23 were unpublished was replaced after those blocked accounts were written. No paper findings were invented.