Imported from zloeber/american-corruption-research (AGENTS.md). Install upstream with npx skills add zloeber/american-corruption-research. Copyright stays with the author.

AGENTS.md — The American Corruption Network

Working manual for agents extending or maintaining this project. Read before editing anything.

1. What this is (30 seconds)

A versioned, evidence-classified research artifact on political corruption, money in politics, conflicts of interest, influence networks, regulatory capture, foreign influence, and power-to-wealth conversion in the US (1900–2026). Delivered as:

• a full research report (report/THE_AMERICAN_CORRUPTION_NETWORK.md)
• a self-contained forensic explorer (explorer/index.html) — interactive network graph (92 entities / 164 relationships), timeline, money flows, party comparison, claims/sources browsers, full report embedded
• a 27-slide presentation (presentation/index.html)
• 14 canonical datasets, a claim registry (36), a source registry (40), a red-team archive, versioned prompts, and a research manifest (chain of custody)

Research date: 2026-08-23 · Version: 1.0.0.

The product is empirical honesty: no manufactured balance, no fabricated model metadata, ranges when measures disagree, "inconclusive" where that is true, and — critically — no upgrading suspicion to allegation to finding.

2. The three rules that must never be broken

Rule 1 — Lineage (root Section 32)

No final claim exists without a traceable chain: Claim → Dataset → Source → Methodology → Red Team → Prompt → Model. Model metadata is recorded as UNKNOWN where not determinable — never fabricate it (root Section 45).

Rule 2 — The explorer is GENERATED, not hand-edited

explorer/index.html is a build artifact. Canonical data lives in the JSON files under data/, claims/, sources/; the report lives in report/THE_AMERICAN_CORRUPTION_NETWORK.md. The template is explorer/template.html; the builder is analysis/scripts/build_explorer.js. Any data change requires re-running the builder. Editing index.html by hand silently desyncs the explorer.

Rule 3 — Prompt Provenance (root Section 32)

Every prompt provided or generated during research must be preserved and traceable. This includes:

• Root prompts (user-provided)
• Research decomposition prompts
• Domain agent prompts
• Verification and red-team prompts
• Synthesis and publication prompts
• Any ad-hoc prompts generated during the session

Prompt storage requirements:

1. All prompts are stored in prompts/ directory with sequential numbering
2. Each prompt file includes: prompt text, timestamp, source (user/agent), and purpose
3. The research_manifest.json tracks all prompt versions and their roles
4. The explorer's LINEAGE tab displays the complete prompt chain
5. No prompt may be silently overwritten — append new versions, keep old ones

Anti-bias mechanism: The prompt provenance system exists to root out bias in the research and researcher. By preserving all prompts, we can:

• Trace how initial questions shaped findings
• Identify where agent-generated prompts may have introduced bias
• Verify that red-team prompts genuinely challenged conclusions
• Ensure that synthesis prompts did not manufacture balance

Prompt injection protocol: When a new prompt is provided or generated during a session:

1. Save it to prompts/ with next sequential number
2. Update research_manifest.json prompts array
3. Add entry to the explorer's LINEAGE section
4. Include in the execution graph if it changes the research flow

3. Directory map

4. Common workflows

4.1 Change any dataset or the report

1. Edit the canonical file.
2. Run node analysis/scripts/build_explorer.js (validates markers, injects data).
3. Verify: grep -c "__" explorer/index.html must return 0 (no leftover markers).

4.2 Add a claim

1. Add to claims/claim_registry.json (id claim-0NN, statement, classification, confidence, sources, supporting/contradicting, limitations, red_team).
2. Rebuild the explorer; the claim is automatically searchable.

4.3 Add an entity or relationship

1. Add to the matching JSON. Relationship edges require BOTH endpoint entity IDs to exist — the graph drops edges with missing endpoints.
2. Rebuild. A new relationship with no endpoints in the current filters will not render — check the era slider.

4.4 Edit the presentation

presentation/index.html is standalone with inline slides and keyboard nav (← → / space). It cites dataset paths but does not read them — figure changes are manual.

4.5 JSON validity

node -e "JSON.parse(require('fs').readFileSync('FILE','utf8'));console.log('OK')"

Run for every edited JSON file.

5. Conventions

• IDs: e-NN entities, r-NNN relationships, sc-NN scandals, ev-NN events, claim-0NN, s-NN sources, adm-YYYY administrations. Never renumber.
• Evidence classes: 🟢 documented / 🔵 strong / 🟡 circumstantial / 🟠 allegation / 🔴 contradicted / ⚪ unknown. Never upgrade an allegation because it is repeated.
• Confidence flags: HIGH = verified this run against a primary source; MEDIUM = stable, multiply-sourced; LOW = approximate/single-source.
• Verdict vocabulary: strongly supported / supported / mostly supported / mixed / mostly unsupported / unsupported / contradicted / insufficient evidence.
• Label discipline: no "pump-and-dump" without the full evidentiary standard — use "possible market-manipulation pattern"; no "bribery" without an adjudicated finding; "documented relationship, unproven causation" is the default for timing+access without intent.
• Ranges, not false precision. Party comparisons are descriptive, not causal.
• The report must never contain </script (breaks the explorer embed). Check with grep -n '</script' report/*.md.

6. Verification workflow (before finishing any task)

# 1. JSON validity (all files)
for f in data/normalized/*.json data/derived/*.json claims/*.json sources/*.json; do
  node -e "JSON.parse(require('fs').readFileSync('$f','utf8'));console.log('OK $f')"
done
# 2. Rebuild explorer and check no markers remain
node analysis/scripts/build_explorer.js
grep -c "__" explorer/index.html   # must be 0
# 3. Report embed safety
grep -c '</script' report/THE_AMERICAN_CORRUPTION_NETWORK.md   # must be 0
# 4. Live preview: register explorer/index.html, reload after every rebuild

7. JS/HTML gotchas (learned the hard way)

• Large str_replace payloads get truncated. Chunk big insertions; use a unique end-marker and replace it repeatedly. Never attempt >~6KB in one call.
• The report embed is raw text until </script — never introduce that sequence into the report.
• The explorer preview server serves only the registered HTML file — fetch() of sibling files 404s. The explorer must stay fully self-contained; this is why the builder inlines everything.
• null - null === 0: guard missing amounts/dates before arithmetic.
• Graph filters: edges with a: null are excluded by any Min-$ filter; era filters use the start-year of d. A relationship dated "1971-" has year 1971.
• Markdown renderer subset: headings (1–4), pipe tables, -/* bullets, numbered lists, blockquotes, --- rules, bold/italic/code, and [text](https://…) links. No fenced code blocks, no nested lists, no images.
• The <!--TPL_APPEND-->/<!--REPORT_APPEND--> markers are build-time only; neither may remain in a final artifact.

8. What NOT to do

• Do not edit explorer/index.html directly — edit the template + JSON and rebuild.
• Do not renumber existing IDs.
• Do not silently overwrite prompts — append new versions.
• Do not manufacture balance or partisan conclusions; where a partisan narrative is contradicted, say so explicitly; where evidence is inconclusive, say so.
• Do not label conduct as bribery/fraud/pump-and-dump without the evidentiary standard (see §5).
• Do not modify files outside american-corruption-research/ (the workspace root also contains america-state-research/ and emotion-ontology/).
• Do not commit, push, or deploy unless explicitly asked.
• Do not delete or modify prompt history — this violates prompt provenance (§2).
• Do not add claims without tracing them through the lineage chain (§2, Rule 1).

9. Extension directions the project is ready for

• New primary data as it arrives (SEC findings on $TRUMP/WLFI; TikTok stake terms; 2026 midterm finance) — update JSON + report + rebuild.
• Fresh econometrics on donation→access→outcome (the sibling project america-state-research has the party-macro framework; this project has the corruption ledger — a merged study is the highest-value extension).
• Wallet-level analysis of $TRUMP flows when on-chain data is fully catalogued (add to transactions.json + a market_events derived dataset).
• Explorer features: sankey money-flow view, per-edge claim cross-links, FARA lobbyist subgraph, CSV export.
• Per-administration ethics scorecards normalized like the sibling project's party scorecards.