Jeffrey Lexicon #

A frequency-attributed dictionary of every word Jeffrey Alan Scudder has used in first-hand textual or transcribed material across the Aesthetic Computer corpus. A sub-platter within the papers platter, parallel to jeffrey-platter (canonical photographic self) and the jeffrey-pvc voice clone (canonical audio self).

Live dashboard: https://papers.aesthetic.computer/platter/jeffrey-lexicon/ — search box, top tokens, neologism candidates, spoken-only / written-only splits. Built by bin/render-report.mjs; the HTML lives at system/public/papers.aesthetic.computer/platter/jeffrey-lexicon/index.html and is also dropped to ~/Desktop/jeffrey-lexicon-report.html on each render.

The lexicon is the textual analogue of the photo index. Where jeffrey-platter/manifest.json answers "what does jeffrey look like?", jeffrey-lexicon/dictionary.json answers "what words does jeffrey actually use, where did each come from, and how often?"

This is first-hand only. Anything generative — gpt-image-2 prompts, Claude co-authored commits, AI-rewritten paper paragraphs, machine-translated body text — is excluded by source filter. The provenance discipline is the artifact's reason to exist: a poisoned dictionary is no dictionary.

1. What counts as a first-hand source #

A source qualifies if jeffrey personally typed, spoke, or wrote it, and the path back to him is auditable.

Source class	Path	Filter
Commit messages	`git log --author=` matches jeffrey identities	Drop `Co-Authored-By: Claude` blocks
Code comments	`git blame` author == jeffrey, comment-line tokens only (`//`, `/*`, `;`, `#`, `--`)	Skip vendored / `node_modules` / minified
AC piece source	`system/public/aesthetic.computer/disks/*.{mjs,lisp}`	String literals + comments authored by jeffrey
Lectures	`papers/lectures/*.{vtt,md}`	Speaker-tagged jeffrey lines only
Recap audio	`recap/out/**/subs.json`	Whisper segments of jeffrey speaking
arXiv papers	`papers/arxiv-/.tex` where author == jeffrey alone	LaTeX-stripped body; drop AI-rewritten sections (tag)
CV	`papers/cv/cv.tex`	Body text, no field labels
Repo planning	`TODO.txt`, `schedule.txt`, `honeydo.txt`	Whole file
IG captions	`portraits/jeffrey/ig-archive/whistlegraph/*.txt`	Captions only, drop hashtags+mentions to a side bucket
Vault email	`aesthetic-computer-vault/<mail>/sent/`	Sent-by-jeffrey only; never enters the public artifact
Handwriting	TBD: scanned-notebooks OCR pass	Confidence ≥ 0.85 per token

Excluded by rule:

gpt-image prompts, gpt-4o vision descriptions, jeffrey-described.jsonl
Commits authored by Oven, Claude, or any non-jeffrey contributor
Auto-generated logs, build outputs, lock files, package metadata
Translation outputs (Danish/Spanish/Chinese/Japanese paper variants)
Paper sections marked % ai-pass in the .tex source (convention TBD)
jas.life closed-source material

2. Schema #

dictionary.json is the build output. One entry per token (lemmatized? — see §4 open questions).

{
  "kidlisp": {
    "count": 1247,
    "by_source": {
      "commits": 512,
      "comments": 88,
      "lectures": 14,
      "papers": 421,
      "ig_captions": 23,
      "recap_audio": 7,
      "ac_pieces": 182
    },
    "first_seen": {
      "iso": "2024-08-14T03:17:22Z",
      "ref": "commits/9f3a1b2",
      "context": "rename lisp -> kidlisp"
    },
    "last_seen": {
      "iso": "2026-05-08T11:04:51Z",
      "ref": "lectures/K8xCBPW1rmo",
      "context": "...so kidlisp is what fits on a card..."
    },
    "sample_contexts": [
      {"src": "papers/arxiv-kidlisp/kidlisp.tex", "line": 142, "snippet": "..."},
      {"src": "lectures/K8xCBPW1rmo.vtt", "ts": "00:14:32", "snippet": "..."}
    ],
    "provenance": "first-hand",
    "neologism": true
  }
}

Every entry carries the source breakdown so derivative views are cheap:

Neologisms — words first introduced by jeffrey (kidlisp, whistlegraph, notepat, platter, platterization, aestheticants, ...). Detected by absence in a baseline corpus + first_seen.ref being authored by jeffrey.
Vocabulary drift by year — bucket counts by first_seen year.
Spoken vs written — lectures + recap_audio vs everything else. Tells us whether a word lives in jeffrey's mouth or only on his keyboard.
Code-only words — comments + ac_pieces only. Names that never escaped the source tree.

3. Build pipeline #

Each source class gets its own extractor under bin/extractors/. The driver bin/build-lexicon.mjs fans out, merges, and writes dictionary.json.

manifest.json (source roots + provenance rules)
  → bin/extractors/<class>.mjs (one extractor per source class)
    → per-source token streams with {token, src, ref, ts, context}
      → bin/build-lexicon.mjs (merge, count, attribute)
        → dictionary.json
        → views/ (neologisms.json, by-year.json, spoken-vs-written.json, code-only.json)

Cache discipline (mirrors recap/cli.mjs and papers/cli.mjs): each extractor caches by (source-path, content-hash) so reruns cost ~zero.

node papers/jeffrey-lexicon/bin/build-lexicon.mjs --all
node papers/jeffrey-lexicon/bin/build-lexicon.mjs --source commits
node papers/jeffrey-lexicon/bin/build-lexicon.mjs --since 2026-01-01
node papers/jeffrey-lexicon/bin/build-lexicon.mjs --views      # rebuild derivative views only

4. Open questions (decide before first full build) #

Lemmatization — collapse paint / paints / painted / painting to one entry, or keep separate? Argument for keep: the morphological choice is part of voice. Argument for collapse: frequencies are more interpretable.
Casing — preserve original case (so KidLisp ≠ kidlisp)? Or fold? Probably fold but track a forms: { "KidLisp": 412, "kidlisp": 835 } side field.
Tokenizer for code identifiers — split paintBucket into paint + bucket? Keep as one? Both views are useful.
Vault inclusion in the public artifact — vault email content should contribute to counts but never appear in sample_contexts for anything jas.life-sourced. Two-tier output: dictionary-public.json (no vault snippets) vs dictionary-full.json (vault-included, gitignored).
AI-pass tagging in .tex — convention for marking AI-rewritten paragraphs so the extractor can skip them. Proposal: % ai-pass-begin / % ai-pass-end block markers.
OCR pipeline for handwriting — Tesseract? Apple Vision? confidence threshold for inclusion?

5. Roadmap #

Phase 1 — commits + comments — easiest, high yield, ~13k jeffrey commits + git-blame on disks/. Covers ~70% of the textual surface.
Phase 2 — lectures + recap audio — pull from existing VTT and subs.json. The "spoken jeffrey" axis turns on here.
Phase 3 — papers + CV — LaTeX-strip with detex or pandoc -t plain, drop % ai-pass-* blocks.
Phase 4 — IG captions + planning files — straightforward.
Phase 5 — vault email — sent-only, two-tier output.
Phase 6 — handwriting OCR — last because the corpus is unknown.
Public dashboard — papers.aesthetic.computer/platter/lexicon/ mirroring the shape of platter/jeffrey/ (search, sort, filter by source class, click-through to context).
Cross-reference back to jeffrey-platter — caption text from photo metadata also feeds the lexicon (where present).

How to extend this index #

New source class? Add a row to §1, an entry under sources in manifest.json, and an extractor at bin/extractors/<class>.mjs.
Mark anything generative as provenance: "ai" in the manifest and the extractor will skip it. Never silently include.
When a build runs, append a row to a BUILDLOG.md here noting source counts + total tokens — drift in those numbers is the early-warning signal for filter regressions.