reading-notes
Notes are the compressed layer every later agent reads before opening PDFs (retrieval order: own notes → papers → textbooks → baseline code). They must be trustworthy enough to cite from and honest about what was not read.
File and naming
notes/papers/<year>-<id-or-shortname>.md (e.g. notes/papers/2021-266-vole-psi.md). One file per paper;
append dated sections on re-reads instead of overwriting.
Template
---
key: RS21 # bib key in refs.bib (must pass bib-verify)
eprint: 2021/266 # or doi
read: 2099-01-31 (sections 1-4, 6; appendix skipped) # say what you did NOT read
reader: lit-scout
---
## One-line contribution
## Setting (model, parameters, security level, hardware)
## Claims (numbered, each falsifiable, with page/section)
C1. ... (Thm 3, p.12)
## Concrete numbers (value, unit, parameters, source pointer)
## Technique in 5 lines (the mechanism; which object is changed)
## Hidden assumptions / fixed parameters
(e.g. "power-of-two ring only", "sparse secret", "LAN only", "the base B is fixed to 2^k without justification")
## Representation tricks (one line each: "turns X from fixed into a free variable")
## Weak points / what a reviewer would attack
## Open problems stated by the authors (quote + page)
## Relevance to us: compare | reuse | mine | ignore — and why
Rules
- Every number carries a source pointer; every claim is phrased so it could be wrong.
- Distinguish "the paper proves", "the paper measures", "the paper conjectures/heuristic".
- Record what you skipped. "Read all 28 papers" is a claim the PI may audit.
- If the note disagrees with another note, add it to the project's errata section and tell
pi-orchestrator. - Never paste long verbatim passages; paraphrase and point.
- Negative notes ("this line does not help because ...") are kept — they stop the next agent repeating the read.