Evidence-First Document Ingestion

Evidence-First
Document Ingestion

Preserve the original.
Extract structured content.
Trace every derived artifact back to its source.

Original first. Derived second. Evidence always traceable.

SOURCE HASH
report.pdf / AAA...
Original integrity

↓
TRANSFORM

TRANSFORMATION RECORD
Tool: MarkItDown
Version: 0.1.2
Time: 2026-09-21T10:32:11Z

↓

DERIVED HASH
report.md / BBB...
Derived integrity

Evidence-First • Vanilla JS 19KB • No React • GitHub Pages /docs • Final Baseline

01 WHY EVIDENCE-FIRST

Transformation does not replace evidence.

02 INGESTION PIPELINE

Local files
→ MarkItDown

Web pages
→ Crawl4AI

Interactive
→ Browser Use

Future batch
→ Scrapy / Crawlee

SOURCE → INGEST → DERIVED ARTIFACT → EVIDENCE RECORD → TRACE / VERIFY

03 CORE TOOLS

MarkItDown
License: MIT (core project only; optional dependencies may carry different licenses)
Note: OCR extras previously involved PyMuPDF / AGPL
Role: Local doc extraction
Output: Markdown
microsoft/markitdown

Crawl4AI
License: Apache-2.0; project also specifies attribution requirements — check current LICENSE/README before redistribution
Official README: Apache License 2.0 + attribution clause in changelog
Role: Web page → Markdown
unclecode/crawl4ai

Browser Use (Extension)
License: MIT
browser-use/browser-use

Scrapy (Extension)
License: BSD-3-Clause
scrapy/scrapy

04 EVIDENCE CHAIN

Each record must include full path, tool version, derived hash.

Evidence ID: EV-0001

Source Name: report.pdf

Source Path: /Design/RevB/report.pdf

Source SHA-256: 8f4a3c9e1b2d...a1f9

Captured At: 2026-09-21T10:32:11Z

Tool: MarkItDown

Tool Version: 0.1.2

Derived Artifact: report.md

Derived SHA-256: b7e2f8a1c9d0...4e2a

Status: VERIFIED

Original
report.pdf
SHA256: AAA...

→
Tool + Version + Time

Transformation
MarkItDown
0.1.2
2026-09-21T10:32:11Z

→

Derived
report.md
SHA256: BBB...

Which source, with which tool version, at what time, produced which exact derived file.

05 EVIDENCE VIEWER

SOURCE

ORIGINAL

DERIVED

06 LIMITATIONS

Hash ≠ Truth - verifies integrity, not truthfulness.
Timestamp ≠ Authenticity
Derived ≠ Original
System ≠ Legal chain-of-custody

Hash 可驗證檔案完整性，但不能證明來源內容本身為真。

07 Appendix: Original 10-Tool Reference (Corrected)

#	Tool	Repo Link	Corrected Info	Role
1	MarkItDown	microsoft/markitdown	MIT (core only; optional deps may differ, e.g. PyMuPDF/AGPL for OCR extras)	Core
2	Crawl4AI	unclecode/crawl4ai	Apache-2.0; project also specifies attribution requirements — check LICENSE/README	Core
3	Firecrawl	mendableai/firecrawl	AGPL-3.0 core, MIT SDK	Not Core
4	Browser Use	browser-use/browser-use	MIT	Extension
5	Scrapling	D4Vinci/Scrapling	BSD-3-Clause	Not Core
6	scrcpy	Genymobile/scrcpy	Apache-2.0 / Android control	Not Core
7	curl-impersonate	lwthiker/curl-impersonate	MIT / network compat only where permitted	Not Core
8	AutoScraper	alirezamika/autoscraper	MIT	Not Core
9	Scrapy	scrapy/scrapy	BSD-3-Clause	Extension
10	Crawlee	apify/crawlee	Apache-2.0	Extension

Evidence-First Document Ingestion

Source Code · GitHub • Evidence-First · GitHub Pages

Baseline for SCADA / engineering document ingestion • Vanilla JS • No React