`SOURCES.md`

120 lines
# CelebritySignatures — Source & Risk Strategy

Goal: a research dataset of **500 authentic celebrity signatures** (5 categories ×
100). Hard rule: **never invent a signature.** Every row must trace to a real,
fetchable image file with a known license. "Keep us out of trouble" is the design
constraint, not an afterthought.

---

## 1. The core problem: a signature has TWO separate legal layers

A celebrity signature is not automatically free to reuse. There are two stacked
rights, and you must clear **both**:

| Layer | What it covers | Typical status |
|---|---|---|
| **(a) Copyright in the signature itself** | The handwriting as a "work" | In the US, a plain signature usually lacks the originality for copyright → effectively **public domain** (Commons tags these `PD-signature`). Highly stylized logo-signatures *can* be protected. |
| **(b) Copyright in the scan/photo** | The specific image, scan, or photograph of the signature | This is its **own** copyright held by the scanner/photographer/auction house. A PD signature photographed by Getty is still a Getty-controlled *image*. |
| **(c) Right of publicity / personality** | Commercial use of a *living* person's identity/name | **Not copyright.** Varies by state (California Civil Code §3344, NY Civil Rights Law §50/51). Using a living celebrity's signature to sell product can imply endorsement → real exposure. Generally **extinguishes or weakens after death** (CA gives 70 yrs post-mortem publicity, but for a *collage/editorial* the bar is far lower than product endorsement). |
| **(d) Trademark** | Signatures registered as marks (e.g., athlete brand logos) | Some signatures are registered trademarks. Avoid those for commercial use. |

**Takeaway:** the safe set = layer (a) is PD-signature **AND** layer (b) source is
openly licensed **AND** layer (c) is low (deceased / editorial use). That is why
your instinct — **politics + historical/deceased first** — is correct: those rows
clear all four layers cleanly.

---

## 2. Source buckets, ranked by safety

### 🟢 TIER 1 — Safe, machine-verifiable, openly licensed (build from these first)

1. **Wikidata property `P109` (signature)** — *the spine of this project.*
   - Wikidata links a person → a signature **file on Wikimedia Commons**.
   - Query the public **SPARQL endpoint** (`query.wikidata.org/sparql`) for all
     humans with `P109` set, filtered by occupation, ranked by notability.
   - Every result is a **real file** — no hallucination possible.
2. **Wikimedia Commons** — the file store behind P109.
   - Most signatures carry `{{PD-signature}}`, `PD-old`, `CC0`, or `CC BY-SA`.
   - Commons API (`prop=imageinfo&iiprop=extmetadata`) returns `LicenseShortName`,
     `Artist`, `UsageTerms` per file — we pull this for every row.
   - Direct image URL: `Special:FilePath/<filename>`. File page: `wiki/File:<name>`.
3. **Wikipedia infobox `signature=`** — same files as P109, surfaced in articles.
4. **U.S. federal government primary sources (PD by 17 U.S.C. §105):**
   - **National Archives (NARA)** catalog.archives.gov — signed treaties, bills, letters.
   - **Library of Congress** loc.gov — manuscript division, presidential papers.
   - **Presidential Libraries** (NARA system) — POTUS signatures on official docs.
   - **Founders Online** (founders.archives.gov) — Washington/Jefferson/Hamilton/etc.
   - These are **public domain as US Government works** — gold for the Politics bucket.
5. **GPO / Congressional Record, Federal Register** — official signed instruments.

### 🟡 TIER 2 — Usable with care (attribution and/or editorial-only)

6. **Museum & university manuscript collections** (Morgan Library, Smithsonian Open
   Access, university special collections). Many now publish **CC0/Open Access**
   scans — check each item's rights statement.
7. **Auction house catalogs** (Christie's, Sotheby's, Heritage Auctions, RR Auction,
   Bonhams) — **reference only.** Great for *confirming a signature is authentic and
   what it looks like*, but the catalog **photo is copyrighted** by the house. Use to
   verify, not to reuse the pixels.
8. **Wikimedia `CC BY` / `CC BY-SA`** signature files — fine to use **with
   attribution**; SA means derivative collage may need to be share-alike.

### 🔴 TIER 3 — Permission-needed / avoid (flag, never auto-include)

9. **Getty, Shutterstock, Alamy, AP** — licensed stock, full copyright. ❌ for reuse.
10. **eBay, autograph dealers (PSA/DNA, Beckett, JSA listings), Pinterest, fan
    sites, random blogs** — unknown provenance + copyrighted photos. ❌ default.
11. **Any living celebrity's signature on a commercial product/collage** — right of
    publicity exposure even if the *image* is free. Mark `permission-needed`.

---

## 3. Ranking sources (for "why is this person on the list")

The automated rank uses **Wikidata sitelink count** — the number of language
Wikipedias that have an article on the person — as a reproducible, neutral
cross-wiki notability proxy. For human-curated cross-checks, these are the
authoritative lists to cite per category:

- **Sports:** ESPN top-athletes, Sports Illustrated, BBC Sport, Olympics.com athlete
  pages, Laureus Awards, sport-specific Halls of Fame.
- **Politics:** heads of state/government lists, US presidents, Gallup "Most Admired,"
  Nobel Peace laureates, TIME 100.
- **Movies:** IMDb STARmeter, Forbes Celebrity 100 / highest-paid actors, Academy
  Award winners, AFI lists.
- **TV:** Emmy winners, YouGov popularity, Nielsen-era icons.
- **Hollywood (classic/iconic):** AFI "100 Years…100 Stars," Hollywood Walk of Fame,
  Academy honorary awards, deceased Golden-Age legends.

These are recorded as the `ranking_source` guidance; the per-row `rank` is the
sitelink-sorted position so the dataset is regenerable and audit-able.

---

## 4. Risk model encoded in the dataset

Each row gets a computed `risk_level` (low / medium / high) and
`usable_in_commercial_collage` (yes / permission-needed / no) from:

```
license_risk:  PD-signature / PD-old / CC0      -> low
               CC BY / CC BY-SA                 -> low-medium (attribution/SA)
               unknown / non-Commons            -> high
publicity_risk: deceased                        -> low
                living                          -> medium (editorial) / high (commercial)
overall risk = max(license_risk, publicity_risk)
usable_in_commercial_collage = (deceased AND license in {PD,CC0,CC BY}) ? yes
                             : living ? permission-needed : review
```

**Build order (agreed):**
1. **Politics + historical/deceased** — clears all four legal layers. Do first.
2. **Sports / Movies / TV** — only rows where Commons/official-archive licensing is
   explicit; living people auto-flagged `permission-needed`.
3. **Hollywood** — bias to deceased Golden-Age icons (cleanest publicity posture).

No single signature is rendered larger or more prominent than another — uniform
treatment is enforced at the rendering layer (see README "collage rules").