primary-source-playbooks

name	primary-source-playbooks
description	Per-domain playbook for finding load-bearing primary sources — names the canonical archives, docket systems, FOIA paths, primary-document repositories, and standard-of-evidence rules for the four research domains (historical, war/statecraft, AI/technology, public law/politics) plus the cross-domain ICC/UN bodies. Used by every researcher to know where to look first and by Alan to verify domain-authority citations.
version	1.0.0

Primary Source Playbooks

Global Five Over-Rules

1. Evidence before elegance. Never improve the story by weakening the evidence.
2. Responsibility follows control, benefit, knowledge, and preventability. Do not stop at the most visible actor.
3. Keep the taxonomy intact. Distinguish pure scapegoat, partial scapegoat, system/object alibi, and cost-bearing goat.
4. Steelman before judgment. Every major claim must face its strongest counterargument before it is asserted.
5. Handoff cleanly. Every output must state assumptions, evidence grade, open questions, and next owner.

How to use this skill

For each case, identify the dominant domain (historical / war / AI / politics), open the corresponding playbook section, and work top-down: start at the highest-authority source in that domain and stop when corroboration reaches the grade the chapter needs. Do not skip to secondary sources before the primary-source path has been exhausted.

If a case spans multiple domains, run multiple playbooks and reconcile in the source ledger.

Decision rubric

For every load-bearing claim, the strongest available primary source must be:

• From the highest tier of the domain playbook below that is reachable given access constraints (paywall, archive closure, language).
• Verifiable: a third party with the citation can confirm the claim by retrieving the document.
• Permanently archived or, if not, captured via web.archive.org snapshot with the snapshot URL in the source ledger.

If no source from tier 1 or 2 of the matching domain playbook can be obtained, the claim caps at evidence grade C until that changes (per .claude/skills/evidence-grading/SKILL.md).

Conflict handling

1. Tier-1 source says X; tier-2 source says not-X: Tier 1 wins for the specific fact, but log the disagreement and the tier-2 reasoning in the disputed-claim register. Many tier-2 sources have synthesized later evidence; if the synthesis predates the tier-1 publication, escalate to Alan.
2. Tier-1 source is behind a paywall the project cannot reach: Record the citation with [PAYWALLED] marker. Use the strongest tier-2 source that summarizes the tier-1 document. Cap the claim at B. Hand off to Delon to budget the paywall fee if the claim is load-bearing.
3. The case predates digital archives and the tier-1 source is physical-only: Use the best published transcription with provenance. Cap at B unless an academic editor of acknowledged authority has authenticated the transcription.
4. Two state actors give contradictory tier-1 statements: Each is a tier-1 source for the OTHER side's accusation, not for the underlying fact. Treat the claim as contested per the dispute rule above.

Escalation conditions

• Escalate to Alan (with the matching domain frame) when the tier hierarchy is ambiguous in a fast-moving case.
• Escalate to Delon when access blocks (paywall, sealed records, language gap) prevent reaching tier 1 and the claim is load-bearing.
• Escalate to Nancy when the primary source itself contains material that carries defamation risk if quoted.

Historical domain playbook (Shirley)

Tier 1 (court / official inquiry / regulator):

• Final court judgments + opinions (national supreme/federal court systems; cite docket + date + judge).
• Official inquiry reports (e.g., Warren Commission, Hutton Inquiry, Bloody Sunday Inquiry, ICMR/Madras High Court reports for Bhopal, Pecora Commission, Pujo Committee, Truth Commissions).
• Regulator findings: SEC litigation releases, OSHA citations, USDA recall notices, NTSB reports.
• Statutes / executive orders as enacted; not summaries of them.

Tier 2 (peer-reviewed scholarship / acknowledged archival editions):

• Peer-reviewed academic monographs on the case; multiple-edition standard works (e.g., specific Bhopal monographs by Kim Fortun, Lapierre & Moro).
• Published archival editions: Foreign Relations of the United States (FRUS), Public Papers of the Presidents, established collected editions of correspondence.

Tier 3 (acknowledged investigative journalism with source disclosure):

• Major investigative work with FOIA documents disclosed (e.g., NYT investigations with sidebars listing source documents, Washington Post Watergate-tier reporting).
• Established newspapers of record at the time of the case (paper of record varies by country and era — name the paper and its standing in the period).

Tier 4 (do not use as primary):

• Wikipedia summary; popular history without source apparatus; quote attributed without primary citation.

War / statecraft domain playbook (Selina)

Tier 1 (international tribunals + UN / ICRC / OHCHR):

• ICC arrest warrants, indictments, judgments (cite docket: ICC-XX/XX-XX/XX).
• ICTY / ICTR / SCSL / KSC judgments and trial transcripts.
• UN Security Council resolutions; UN Commission of Inquiry reports (Syria, Myanmar, Ukraine); OHCHR human-rights reports.
• ICRC commentaries on the Geneva Conventions and Additional Protocols (definitive authority on IHL).
• National courts of universal jurisdiction (German Higher Regional Court Koblenz, Swedish Higher Court) judgments.

Tier 2 (acknowledged conflict-monitoring NGOs + state inquiry):

• Human Rights Watch reports; Amnesty International detailed investigations.
• Bellingcat open-source-intelligence investigations with documented chain of custody.
• National-level commissions of inquiry (Chilcot Inquiry on Iraq, the Bloody Sunday Inquiry).
• Forensic Architecture investigations with published source documents.

Tier 3 (high-quality conflict journalism with named sources):

• AP, Reuters, AFP wire reports with named correspondents on the ground.
• Major outlets with documented FOIA / leaked documents (NYT, WaPo, Guardian, Spiegel, Le Monde).

Tier 4 (do not use as primary):

• Party-to-conflict press releases as fact (use as evidence of that party's STATEMENT only).
• Anonymous Telegram channels, partisan analysis aggregators, single-source social media claims.

AI / technology domain playbook (Warren)

Tier 1 (primary documents from the entity + court filings):

• Model cards and system cards published by the lab (link to the specific version + date; Wayback Machine snapshot for any model card that has since been edited).
• Pre-registered benchmark protocols and peer-reviewed evaluation papers (NeurIPS, ICML, ACL with reproducibility code).
• Court filings: complaints, depositions, exhibits filed in AI lawsuits (PACER docket cites).
• Regulator filings: FTC complaints + consent orders; EU AI Act conformity assessments; CMA / EU competition rulings.
• Official lab safety reports (e.g., Anthropic's RSP releases, OpenAI's Preparedness Framework reports, Google's Frontier Safety Framework documents).

Tier 2 (peer-reviewed independent evaluations + acknowledged AI safety research):

• Independent benchmark papers (METR, Apollo Research, MLCommons audited results).
• Mechanistic interpretability publications from non-lab research groups.
• Established AI policy research (Stanford HAI, Oxford GovAI, AI Now).

Tier 3 (specialist investigative journalism):

• Wired, MIT Tech Review, The Information, Bloomberg AI coverage with named sources.
• Specialist tech reporters with documented FOIA / leaked-document use.

Tier 4 (do not use as primary):

• Lab marketing posts, conference keynotes, podcast interviews as fact about capabilities.
• Twitter/X discussions of benchmark results without the underlying paper.
• LessWrong / EAForum posts as primary; use only as commentary pointing to primary sources.

Public law / politics domain playbook (Loki)

Tier 1 (court docket + agency action + statute):

• Federal court dockets via PACER / CourtListener / RECAP (cite case number + court + date + filing type).
• Executive orders as enacted (whitehouse.gov + Federal Register citation).
• Federal Register notices of proposed and final rulemaking (FR cite: vol + page + date).
• GAO reports, Inspector General reports, CRS reports.
• Congressional records: bills as introduced, committee reports, floor speeches in the Congressional Record.

Tier 2 (acknowledged legal scholarship + administrative-law-focused outlets):

• Just Security, Lawfare, Election Law Blog (Hasen), Volokh Conspiracy, SCOTUSblog — cite the specific post + author.
• Peer-reviewed legal scholarship in law reviews.

Tier 3 (specialist political journalism with documented sourcing):

• AP, Reuters, NYT, WaPo political reporting with named officials or documented memos.
• Specialist beat reporters (immigration, EPA, DOJ, Pentagon) with FOIA disclosures.

Tier 4 (do not use as primary):

• Cable news commentary; partisan think-tank press releases; Twitter/X threads from anonymous accounts.
• "A senior official said" without an organization, role, or context.

Cross-domain bodies (use as tier 1 for the matching frame)

• ICC, ICJ, ICTY/R, SCSL, KSC: war crimes / atrocity attribution.
• UN Special Rapporteurs (country and thematic mandates): UN-aligned tier 1.
• OECD anti-bribery working group reports, FATF mutual evaluation reports: financial-crime cross-domain.
• WHO outbreak reports, ECDC, CDC MMWR: public-health-cross-domain.

Output format

For each claim, the source-ledger row produced by this skill must include:

Claim:
Domain:
Tier 1 source attempted: <citation> | not available because <reason>
Tier 2 source used: <citation>
Tier 3 source used (if any): <citation>
Access notes: <paywall / archive / language>
Wayback snapshot URL (if applicable):
Evidence grade (per evidence-grading skill):
Open questions:
Handoff:

Context: Selina needs a primary source for the OHCHR-corroborated civilian-harm count in Bucha. input: case="Bucha civilian harm March 2022" output: Domain: war/statecraft. Tier 1 attempt: OHCHR Report on the human rights situation in Ukraine (specific report + paragraph cite); ICC arrest warrant for Putin (ICC-01/22 docket cite); Ukrainian Prosecutor General registry (URL + access date). Tier 2 used: HRW report on Bucha (Sept 2022). Tier 3 used: AP/Reuters reconstructions for specific contested counts. Access notes: ICC docket public; PGO registry in Ukrainian — translation needed for two specific entries. Wayback snapshot: yes for OHCHR + PGO URLs. Evidence grade: A for OHCHR-corroborated counts, C for single-source attributions of specific perpetrators. Handoff: Stephen (verify the translations) → Alan (IHL frame on attribution). Context: Warren needs a primary source for "Anthropic's safety claims about Claude 4.7" for an AI-governance chapter. input: case="Anthropic Claude 4.7 safety claims" output: Domain: AI/technology. Tier 1 attempt: Anthropic's Claude 4.7 system card (cite URL + Wayback snapshot taken on access date because system cards are versioned); Anthropic's Responsible Scaling Policy current release; any FTC or court filings (none yet). Tier 2 used: METR independent capability evaluations if published for this model; Apollo Research evaluation if published. Tier 3 used: Wired/Information/Bloomberg AI coverage WITH named Anthropic sources or documented memos (NOT marketing posts). Access notes: system card public; some appendices may be paywalled. Wayback snapshot: mandatory for the system card. Evidence grade: A for what the system card admits; B for METR/Apollo independent evaluations; C for any single-source journalism without document corroboration. Handoff: Stephen → Alan (AI governance frame) for whether the system card's hedging language supports the chapter's claim.