Writeup — pick your guide
This project’s writeup ships in two reading-style formats. Pick the one that matches how you’d like to read it.
| Style | Length | Best for |
|---|---|---|
| Academic paper (IMRAD) | ~20–25 min | Reviewers expecting Abstract / Introduction / Methods / Results / Discussion / Limits / Conclusion / References discipline. Formal voice, technical terminology with on-first-use definitions. |
| Narrative arc (story) | ~15–20 min | Readers preferring plain-English first-person prose. Same content, story-style pacing (Hook → Setup → Investigation → Revelation → Implications). |
Both guides cover the same content (problem, methods, all 7 findings, mechanism, limitations); the style and pacing differ. Technical terms are defined on first use in either guide and cross-referenced to docs/GLOSSARY.md.
Quick pointers
- Just the numbers → RESULTS.md — tables + 5 figures + raw artifact pointers, no narrative prose.
- 60-second tour → Project at a glance.
- README + executive summary → README.md — 1-page distillation including the headline + mechanism + direct-detection check tables.
- Methodology spokes (deep-dive references) → 8 files under WRITEUP/ covering data decisions, evaluation design, model details, threshold policy, reference-scorer audit, methodology guarantees, reproducibility, limitations + future work.
- Decision trail → 81 ADRs at decisions/.
Why two guides?
The previous single-guide structure was diagnosed in the v1.3.0 restructure as “neither narrative nor academic” — informal headings plus academic numbering, methodology after results, same content rendered with different prose across multiple pages. The two-guide architecture (per ADR-078 + ADR-079) gives each reader-type a self-contained article in its native register, with no cross-page redundancy.
Old WRITEUP.md (the jumbled hybrid) content is now distributed across the two guides; the historical content remains accessible at the reviewer URL pin (tree/v1.0.0) per ADR-033 (corrected from a prior tree/v1.2.8 mis-citation per ADR-080).
Looking for a specific section?
If you arrived here from a tree/v1.0.0-era deep link (e.g., from the submission email or a bookmarked anchor), the original v1.0.0 WRITEUP.md had 7 sections. They are now redistributed as below.
§Reading guide
Reading-style routing for the methodology now lives in the dedicated READING_GUIDE.md (academic / narrative / hiring-manager / reproducer / data-only paths).
### §1 Motivation {#motivation}
Now in WRITEUP_PAPER §1 Introduction (academic framing) and WRITEUP_NARRATIVE Act 0 — Hook (narrative framing).
### §1.5 Attack-type taxonomy + train/test composition {#attack-type-taxonomy}
The train/test composition table and OOD slate breakdown now live in WRITEUP/data-decisions.md and WRITEUP_PAPER §3 Methods. The high-level taxonomy (direct / indirect / agentic-flow / jailbreak / benign-but- injection-shaped) is summarized in the README §What “OOD” means here.
### §2 Approach overview {#approach-overview}
The detector ladder + reference-scorer slate are now in WRITEUP_PAPER §3 Methods and WRITEUP/model-rungs.md. The multi-detector philosophy (lexical baseline → frozen probe → LoRA → full fine-tune + reference scorers) is in WRITEUP_NARRATIVE Act 2 — Setup.
§Results
The headline characterization (cross-family generalization failure + mechanism) is now in WRITEUP_PAPER §4 Results, WRITEUP_NARRATIVE Act 3 — Revelation, and the tables-only RESULTS.md appendix with 5 canonical figures.
### §Lessons (brief) {#lessons-brief}
Methodology reflections now live in WRITEUP_PAPER §5 Discussion and WRITEUP/limitations-and-future-work.md. The “what could be wrong with this” narrative is in WRITEUP_NARRATIVE Act 4 — Implications.
The appendix-class material (operating-point tables, calibration detail, per-source breakdowns, raw artifact pointers) now lives in RESULTS.md and the WRITEUP/ spokes (8 methodology references). The decision trail is at decisions/.
For the original single-document v1.0.0 content verbatim, see the tree/v1.0.0 tag pin per ADR-033 — the historical reviewer URL is preserved unchanged.