Context: One of my ideas from the wikiharness prototype was sourcing a claim from inside scanned Internet Archive books. Below is a micro/draft PRD sketching how it could work in SP42.
@schiste — if the general direction looks right, I'll flesh this out into a full PRD (and any necessary ADRs). If it looks wrong, let's discuss here before I invest in that. It's a direction check, not a finished proposal.
PRD-0013: Book-scan sourcing
Drafter: Claude Code Opus 4.8
Editor: Luis Villa
Date: 2026-06-09
State: Draft
Discussion: (this issue)
Spawned ADRs: none yet (uses ADR-0003's editing operation)
Problem
Some claims are best supported by a book, not a web page, but sourcing one in
SP42 is impossible today: an operator would have to leave the tool, find a
scanned copy, locate the page that supports the claim, and hand-build a citation
with a page pointer. The web-only verify path cannot reach this evidence.
Proposal
For a claim, SP42 searches inside scanned books and proposes a page-anchored
citation:
- It searches the Internet Archive full-text "search inside" index for the
claim's terms and returns candidate books with matching pages — read-only;
no writes to the Internet Archive.
- For a chosen hit, it proposes a citation that includes a page-level deep
link to the matching scan page, as an operator-confirmed edit (ADR-0003 for
placement, ADR-0010 for the write path).
It supports an existing claim with a real, locatable source; it does not write
new prose. (Creating or linking a Wikidata edition item for the book is a
possible follow-on, out of scope here.)
Definition of Done
Alternatives
- Whole-book citation with no page pointer. Rejected: an unanchored book cite
is barely checkable; the page deep link is the value.
- Create a Wikidata edition item inline. Deferred: a separate write target with
its own policy weight; sourcing comes first.
Risks
- A search hit is a false match (term appears but does not support the claim).
Mitigation: the operator reviews the linked page before confirming; this is
sourcing assistance, not an automatic verdict.
Open questions
- Whether to run the verify engine (PRD-0001) against the matched page text
before proposing (proposed: yes when the page is fetchable — reuse grounding).
Context: One of my ideas from the wikiharness prototype was sourcing a claim from inside scanned Internet Archive books. Below is a micro/draft PRD sketching how it could work in SP42.
@schiste — if the general direction looks right, I'll flesh this out into a full PRD (and any necessary ADRs). If it looks wrong, let's discuss here before I invest in that. It's a direction check, not a finished proposal.
PRD-0013: Book-scan sourcing
Drafter: Claude Code Opus 4.8
Editor: Luis Villa
Date: 2026-06-09
State: Draft
Discussion: (this issue)
Spawned ADRs: none yet (uses ADR-0003's editing operation)
Problem
Some claims are best supported by a book, not a web page, but sourcing one in
SP42 is impossible today: an operator would have to leave the tool, find a
scanned copy, locate the page that supports the claim, and hand-build a citation
with a page pointer. The web-only verify path cannot reach this evidence.
Proposal
For a claim, SP42 searches inside scanned books and proposes a page-anchored
citation:
claim's terms and returns candidate books with matching pages — read-only;
no writes to the Internet Archive.
link to the matching scan page, as an operator-confirmed edit (ADR-0003 for
placement, ADR-0010 for the write path).
It supports an existing claim with a real, locatable source; it does not write
new prose. (Creating or linking a Wikidata edition item for the book is a
possible follow-on, out of scope here.)
Definition of Done
index, verified by a test over a replayed Internet Archive response.
scan, verified by a proposal test.
a property test over outbound requests.
Alternatives
is barely checkable; the page deep link is the value.
its own policy weight; sourcing comes first.
Risks
Mitigation: the operator reviews the linked page before confirming; this is
sourcing assistance, not an automatic verdict.
Open questions
before proposing (proposed: yes when the page is fetchable — reuse grounding).