research/Person

15,536 records · yaml · accepts undeclared fields

Fields

Rules

• personIdentifierStringsAreNonBlank

Defects

Checked 200 of 15,536 records (the per-store cap is 200), so this is a sample, not a whole-store verdict.

No defects in the 200 record(s) checked.

Example record

/Users/eshao/.config/mnt/data/research-biography/people/arvind-narayanan.yaml

# Canonical person record. See _template.yaml for the Person schema and shared pointer grammar.
# This record remains canonical; references/arvind-narayanan.md is a generated rendering.
# FORMAT RULES: records are SINGLE-DOCUMENT spec YAML - never add Markdown
# frontmatter (`---`/`published:` mdd stamps split the stream and break
# yaml.safe_load consumers; browsing is the research-records viewer's job).
# Quote any scalar containing ": " anywhere, and quote any flow-sequence scalar
# containing a comma, e.g.
# works: [{title: "Netflix-Prize de-anonymization", venue: "IEEE S&P"}] - unquoted
# inner commas split into mangled list items.
# PROJECTION OVER DUPLICATION (research-loop Structured Data First): facts that came
# through query-openalex/query-arxiv already live in full under lib-cache/lib-runs, so
# this record stores field-level POINTERS + only the extracted fields a biography view
# needs, plus family-owned facts (met:, roles:) that exist nowhere upstream. It never
# re-stores whole API responses; the lists below are projections, re-derivable through their field pointers.
# Schema (minimal v0 - extend as later biographies demand, never silently):
#   slug, name, headline: identity
#   ids: {openalex, orcid, ...}  - stable external identifiers
#   POINTER SHAPE: {via, cacheKey?, runDir?, url?, command?, holds}. A `runId` key is BANNED -
#     runDir already carries the run identity. Per-field provenance may be a SINGLE pointer
#     OR a LIST of pointers; both are legal wherever a pointer is accepted.
#   field provenance: mappings/lists on the fields they support; there is no envelope `sources:` list
#   papers: pointer to the complete projected works collection
#   met: {status: bool, events: [{when, where, what, confidence?, note?, provenance?}]}  - FAMILY-OWNED;
#     canonical event key order is when/where/what/confidence/note/provenance. `where` is the bare place;
#     inference qualifiers belong in `confidence` + `note`, never inside `where`.
#     Networking view is a query over this, not a second store.
#   why: {thesis: 1-2 sentences, evidence: [{date?, claim, source?}]}  - FAMILY-OWNED; the thesis is the
#     claim about why this person matters, and each evidence row is one fact supporting it. `date` is
#     present when the claim is a dated public act and absent when it is a standing property of the record.
#   roles: [{role, confidence, note}]  - FAMILY-OWNED hypothesis list (critic/reader/collaborator/lineage-node)
#   interest_timeline: [{era, span, summary, works: [{title, year?, venue?, citations?, note?, paperSlug?}], provenance}]
#     - field-backed PROJECTION; TIME is the primary axis. `paperSlug` links a $research-reader PaperRecord
#     when one exists. A works row may NAME a talk as career narrative but MUST NOT restate its venue or
#     date - sessions: is the single store for those.
#   edges: [{name, openalexId, kind, works_total?, works_recent?, provenance}]  - field-backed PROJECTION; typed graph substrate
#   connection_degrees: [{name, openalexId, maxDegree, outcome, provenance, caveat?}]  - FAMILY-OWNED resolved links to already-biographied people
#   footprint: {homepage, x, substack, book, openalex, key_posts}  - public sources (web, not query-*).
#     Talks and keynotes live in sessions:, never here.
#   sessions: [{conf, session, kind, media, transcript, note?, provenance?}]  - the SINGLE store for
#     attended and given talks. `kind` is keynote|talk|poster|workshop. `media` and `transcript` are
#     structured refs {status: linked|pending, sessionId?} into $conf-talks; a linked ref renders as
#     https://w.sh/media/<sessionId>.
#   renderings: [{path, published, kind}]  - generated human-facing surfaces; regenerate after record edits, never hand-edit
#   provenance: {summary, caveats: [{n, text}]}  - current evidence and numbered caveats only
#     (namesake merges, unconfirmed dates); never tool-fix or migration history

slug: arvind-narayanan
name: Arvind Narayanan
headline: Professor of CS, Princeton; Director of CITP; co-author of AI Snake Oil

ids:
  openalex: A5058102069
  orcid: 0000-0001-7176-4479
  provenance:
    via: query-openalex
    cacheKey: "authors:arvind narayanan:5"
    runDir: runs/query-openalex/2026-07-28
    holds: author disambiguation search result and lib-runs records for the capture session

papers:
  via: query-openalex
  cacheKey: "works-by-author:A5058102069"
  runDir: runs/query-openalex/2026-07-28
  holds: complete 174-work collection with venues, years, citations, and coauthors

met:
  status: false
  events: []

why:
  thesis: >-
    The most publicly credible skeptic of exactly the two things research-blog Entry #1 is about:
    AI-text detectors deployed as accusation machines, and AI writing quality. Our duel's results
    empirically bound both of his claims (surface tells removable without moving detection; detection
    orthogonal to semantic depth), which makes him the natural first reader for the entry: someone who
    has thought hard about the same detector from the other side.
  evidence:
    - date: 2025-12-09
      claim: >-
        Ran false-positive math against Pangram specifically - the same detector our prose hill-climb
        holds out as its referee oracle - arguing a 1-in-10,000 FPR still falsely flags 5-10% of
        students across a college career.
      source: https://x.com/random_walker/status/1998407244438819149
    - date: 2026-03-06
      claim: >-
        Posted the "hollowness" thesis: the real sign of AI writing is not surface tells but polished
        prose around mundane ideas, diagnosed by whether a second read impresses less.
      source: https://x.com/random_walker/status/2029898822327890231
    - date: 2026-07-27
      claim: >-
        Ran his own attempt to write past Pangram, failed against a hardened detector (>0.99, no
        gradient), and publicly conceded.
      source: https://x.com/random_walker/status/2081713721785631202
    - date: 2026-07-27
      claim: >-
        Named the lever his agent never found - stop whole-span rewriting, because it resamples the
        LLM distribution - which is exactly our campaign's finding.
      source: https://x.com/random_walker/status/2081713721785631202

roles:
  - role: critic-reader
    confidence: high
    note: intended recipient of the Entry #1 outreach; his concession + hollowness threads are the entry's contrast
  - role: lineage-node
    confidence: confirmed
    note: his agent-evaluation-rigor line (AI Agents That Matter, CORE-Bench, Holistic Agent Leaderboard, REFORMS) is upstream of interest area 6
  - role: collaborator
    confidence: speculative
    note: agent-benchmark-rigor agenda overlaps our usage-survival methodology; no concrete hook yet

interest_timeline:
  - era: privacy-and-de-anonymization
    span: 2005-2016
    summary: de-anonymization, web tracking, crypto
    works:
      - {title: Netflix-Prize de-anonymization, year: 2008, venue: "IEEE S&P", citations: 2.3k}
      - {title: De-anonymizing Social Networks, year: 2009}
      - {title: Bitcoin SoK, year: 2015}
      - {title: Online Tracking, year: 2016}
    provenance: &openalex_works
      via: query-openalex
      cacheKey: "works-by-author:A5058102069"
      runDir: runs/query-openalex/2026-07-28
      holds: complete 174-work collection with venues, years, citations, and coauthors
  - era: algorithmic-society
    span: 2017-2021
    summary: bias in ML, dark patterns; AI snake oil talk goes viral (2019)
    works:
      - {title: Semantics derived from language corpora contain human-like biases, year: 2017, venue: Science, citations: 2.8k, note: his most-cited work}
      - {title: Dark Patterns at Scale, year: 2019}
    provenance: *openalex_works
  - era: ml-science-rigor
    span: 2022-2024
    summary: reproducibility/leakage in ML-based science; AI Snake Oil book
    works:
      - {title: Leakage and the reproducibility crisis, year: 2023, venue: Patterns}
      - {title: REF

Where it lives