What this is: Anthropic’s Claude Mythos Preview system card (first-party disclosure, not generally available model). Same treatment as Claude Opus 4.7 system card: official PDF → text → chunked keyword classifier artifacts in-repo.
Canonical sources
- System card (official; resolves to PDF): anthropic.com/claude-mythos-preview-system-card
- Strategic context (coalition, MCP, etc.): Project Glasswing · anthropic.com/glasswing
Local PDF note: An ingest was first run from the author’s iCloud copy (claude mythos add to ai failure periodic table.pdf). Committed artifacts are regenerated from the official URL so everyone reproduces the same file (the public card may include newer changelog lines than a saved copy).
Live classifier outputs (--max-chars 4500, same as Opus 4.7 system card)
| Artifact | Purpose |
|---|---|
claude-mythos-system-card-live-summary.md |
Top-1 histogram + chunk index |
claude-mythos-system-card-live-chunks.json |
Per-chunk scores |
claude-mythos-system-card-live-source.txt |
pdftotext -layout + provenance header |
Reproduce
python scripts/classify_external_report.py \
--url https://www.anthropic.com/claude-mythos-preview-system-card \
--out-prefix reports/claude-mythos/claude-mythos-system-card-live \
--max-chars 4500Interpretation: Keyword classifier on long system-card prose is exploratory (see meta integrity doc caveats). For mechanism-level mapping, prefer curated appendix-style work (e.g. project-glasswing.md §2 / Case 21 threads) plus spot checks with python -m src.cli.