2.1 KiB
Long-context multimodal archive (arXiv, 2026-10-26)
A 134-paper research archive on long-context multimodal understanding, built from five live arXiv API queries, with every claim in the written analysis machine-checkable against the archived source text.
Files
| file | what it is |
|---|---|
papers.jsonl |
134 unique papers: id, title, authors, dates, categories, abstract, pdf url, query hits |
CORPUS.txt |
all abstracts concatenated with headers — 228,277 chars, the long-context pack |
ANALYSIS.md |
the research brief: four findings, verbatim quotes, explicit limits |
claims.jsonl |
37-claim ledger; each claim carries a kind + expected value or a verbatim quote |
verify.py |
auditor: re-hashes raw responses, re-parses them, recomputes every claim, mutation-tests itself |
EVIDENCE.md |
captured run output + sha256 of every artifact |
raw/q1..q5.xml |
raw Atom responses, unmodified |
raw_manifest.json |
query, URL, byte count and sha256 per raw response |
build_stats.json |
corpus totals used by the ledger |
Verify
python3 verify.py # 56 checks; exit 0 == archive consistent and verifier not tautological
python3 verify.py --quiet # failures only
verify.py does not trust claims.jsonl. It recomputes each quantitative claim
from papers.jsonl, re-derives papers.jsonl from the raw Atom responses, checks
those responses against their recorded sha256, and finally corrupts a throwaway
copy of the archive three ways (falsified claim, tampered abstract, flipped raw
byte) to prove it would notice. Any failure returns nonzero.
Rebuild from scratch
fetch_arxiv.py (in the session scratch dir) re-runs the five queries; arXiv
rate-limits with HTTP 429, so it backs off and retries — the fetch log in
EVIDENCE.md records the retries that actually happened.
What this is not
Abstract-level only: claims describe what papers say, verified verbatim, not what
their experiments show. Single source (arXiv), five keyword queries, and heavily
skewed to the last two months of the window — see section 6 of ANALYSIS.md.