simhacker/moollm/change-name

CHANGE-NAME 🦋

Say my name right — everywhere it's written.

People change names — transition, marriage, divorce, faith, stage names, safety, transliteration across scripts. The record does not follow by itself. Names live in version-of-record PDFs, citation databases, git histories, link annotations, metadata, and archives engineered never to change. This skill changes them — carefully, honestly, and with receipts.

One concrete task ships today: ninja-editing PDF files. The context is a socket: each venue (git, ACM, ORCID, registries, archives, life admin) plugs in as a playbook in playbooks/. This SKILL.md is self-contained for the PDF task; playbooks extend it.

The Ethics Gate (read first, always)

A correction is legitimate when all five hold. Any four out of five is a forgery with good manners:

1. Documented wish or clear standing — the subject asked (best: find it in their own

words), or those closest to them act on documented evidence of their identity. Standing order: self > subject's recorded request > family/co-authors > community.

1. Original preserved — bit-for-bit, sha256-pinned, adjacent to the edition.
2. Total disclosure — every change enumerated, visible on page 1 (see

Visible correction notice), in PDF metadata and attachments, and beside the file (README). Metadata-only disclosure fails: a renamed or printed copy loses the filename, README, and sibling original.

1. Honest labeling — the filename says what the file is (-memorial-edition,

-corrected, never a silent replacement).

1. Canonical fix pursued — the edited derivative is a bridge to the institutional

correction (publisher petition, registry update), never a substitute for it.

Never out anyone. A scan that finds every occurrence of an old name is also a map of someone's exposure. Treat SCAN reports as sensitive; publish only what the subject or their standing-holders approve.

Protocol: SCAN → DISCUSS → EDIT → VERIFY → PUBLISH

1. SCAN — look at the PDF and describe what you find

Run the sister script (degrades gracefully without pikepdf):

python3 scripts/pdf_name_scan.py TARGET.pdf --find "Old Name" [--replacement "New Name"]

It reports, without changing anything:

• Document metadata — /Author, /Title, docinfo, XMP
• Content-stream text — including kern-split occurrences: TeX-produced PDFs

store names like (Freudenber)18(g), invisible to naive grep. The scanner inflates streams and matches across TJ kern numbers.

• Link annotations — URIs containing the old name (profile links, repo links)
• Font subsets — per font: BaseFont, CharSet/FirstChar/LastChar, and **whether the

embedded glyphs can spell the replacement**. (Founding case: the email font subset ended at glyph 117 — one code point short of the letter v. It could not spell "vanessa". Know this before you promise an in-font edit.)

Describe findings to the human in plain language: where the name appears, what fonts carry it, what can be edited in-font and what needs a fallback face.

2. DISCUSS — agree the plan

Present the evidence, then agree, explicitly:

• Which occurrences change (byline yes; historical citations in other people's

references may be out of scope), and which stay.

• Replacement strings — name, email, URLs (beware redirect masking: an old-handle

URL that redirects still displays the old handle; update visible text).

• Typography — in-font where the subset allows (recompute kerning), matched fallback

face where it doesn't (base-14 Helvetica for sans, Times for serif).

• Labeling — the edition's filename, the visible page-1 notice (wording and

placement), and the embedded provenance note / attachment text.

• The gate — confirm all five conditions; record the wish/standing evidence in the

ledger.

Visible correction notice (required)

Every prestored PDF must carry a visible correction notice on page 1. Outside-the-page disclosure (filename, README, sibling original, PDF /Note metadata) is necessary but not sufficient — those signals vanish when someone renames, re-hosts, or prints the file.

The notice must be explicit and machine-parseable — clear enough that a human skimming page 1 and plain text extraction (pdftotext, Ghostscript txtwrite) both recover: this is not the version of record; what changed; where the unaltered original lives; that a canonical fix is being pursued.

Required content (condense for page 1; full list may live in a PDF attachment):

• Honest label (CORRECTED MEMORIAL EDITION, CORRECTED EDITION, etc.) and date
• Not the version of record — cite DOI or publisher identifier when one exists
• What was updated (byline, email, footnote URLs, metadata) and the new identity used
• Standing / wish — link the subject's own words when they exist
• Unaltered original preserved — repository path and SHA-256 (full hash, or labeled

prefix with ellipsis)

• Pointer to full change list (embedded attachment or README)
• Canonical correction in progress (publisher petition, registry update, etc.)

Placement recommendations (least disruptive first):

Typography recommendations:

• 7–7.5 pt minimum (7 pt floor; do not go to 6 pt)
• Match the document's house style — sans used for affiliations, or base-14 Helvetica
• Full text-block width (~468 pt on US Letter) — wrap across the whole page, not a

narrow column; fewer lines means less vertical stack and no collision with existing header text like "Accepted for DLS"

• Small caps on the label line (CORRECTED MEMORIAL EDITION)
• Thin rule (0.25 pt) above the notice to separate it from body content
• Monospace or distinct treatment for hashes and URLs aids both reading and extraction

Example (4-line footer; adapt labels to the case):

CORRECTED MEMORIAL EDITION — September 2026. Not the ACM version of record (doi:…).
Author name, email (above), footnote URLs, and PDF metadata updated to [New Name] per
[standing citation]. Unaltered original: [repo path] (SHA-256 …). Full change list
attached; [publisher] name-change correction in progress.

3. EDIT — perform the surgery

With pikepdf, on a copy (never the original):

• Byline/text: locate the exact TJ array in the page content stream; build the

replacement with correct kerning (glyph widths from the font's /Widths; word gaps as -250-style TJ numbers; known kern pairs, e.g. V–a ≈ +111 in Times); re-center by shifting the text matrix half the width delta; assert count == 1 for every pattern before replacing — bytes, not regex, wherever possible.

• Annotations: update /A//URI and resize /Rect hotspots to the new text width.
• Visible notice: overlay the agreed correction notice on page 1 (footer band or top

margin per recommendations above). Render-check at print resolution.

• Metadata: /Author, XMP dc:creator; embed a /Note docinfo field stating what

changed, when, why, and by whom.

• Attachment: embed the full enumerated change list as a PDF attachment when the page-1

notice is condensed.

4. VERIFY — prove it

• Render the edited pages (Ghostscript → PNG) and look at them — centering, font

match, no keming; the page-1 correction notice is visible and legible. Zoom against an unedited neighbor for comparison.

• Extract page 1 — the notice text must appear in plain extraction, not only in rendered glyphs.
• Extract full text (gs -sDEVICE=txtwrite) — the old string's count in agreed

scope must be zero; the new string must appear where planned.

• Regression — all pages render clean (-sDEVICE=nullpage).
• Hash original and edition; record both.

5. PUBLISH — file it honestly

Original + edition + README side by side. README enumerates every edit, pins both hashes, cites the wish/standing evidence, and links the canonical-fix path (e.g. ACM name-change policy, option 3 — corrected version of record). Commit with a message that tells the story.

Beyond PDFs — the playbook socket

playbooks/pdf-prestoration.md is implemented (the founding case, step by step). Planned playbooks each get one file, same shape (when to use, standing required, procedure, verification): git-mailmap (attribution correction without history rewrite), acm-petition, orcid, registries (dblp/Scholar/Wikidata), web-archives (layer corrections beside immutable snapshots — fixity is evidence, not the enemy), life-admin (the chores beyond documents). Add a playbook, advertise it in CARD.yml, and the skill covers a new venue.

Part of MOOLLM

This skill lives in MOOLLM — see skills/README.md for the ecosystem. Founding case study: designs/prestoration/, with the worked example — before and after PDFs, sha256 provenance, and Vanessa's own HN request — in designs/prestoration/sources/. Companion tribute: Vanessa's philosophy. Works standalone: the protocol above and the script in scripts/ need only python3, ghostscript, and pikepdf.