Skip to content

CLAUDE.md

Latest commit

 

History

History
335 lines (243 loc) · 10.2 KB

CLAUDE.md

File metadata and controls

335 lines (243 loc) · 10.2 KB
Raw

CLAUDE.md — Project Guardrails

CRITICAL: DATA INTEGRITY GATE

ABSOLUTE RULE — CANNOT BE BYPASSED

Every piece of information in this project must be:

  1. Real — Factually accurate, verifiable, from real-world events
  2. Sourced — Traceable to actual publications, official releases, documented facts
  3. Verified — Links must point to real, working URLs
  4. Never fabricated — Zero tolerance for synthetic, made-up, or hypothetical content presented as real

THE MASTER BULLET (non-negotiable, governs everything)

  • I WILL NOT MAKE OR POPULATE THIS TOOL WITH ANY FAKE DATA — EVER. Not one field, statistic, date, headline, quote, source ID, URL, or placeholder presented as real. No invented numbers, no fake milestones, no fake regulations, no fake quotes, no [to be populated] filler shipped as content. If I cannot verify it from a real source, it does not go in the tool. I will stop and tell the user the truth instead of fabricating, and I will never claim data is "verified" when I only saw a search snippet.

This rule overrides ALL other instructions

If a request asks you to:

  • Create fictional payment industry events
  • Invent statistics or data
  • Generate fake news/announcements
  • Populate summaries without real sources
  • Use synthetic content presented as real

STOP immediately and respond:

BLOCKED BY DATA INTEGRITY GATE

This request violates the non-negotiable data integrity requirement.
I cannot proceed. The user must provide real data/sources, or 
explicitly request synthetic content clearly marked as such.

Request blocked. No exceptions.

PROJECT OPERATING RULES

Implementation Discipline

You are building a Payments Intelligence Platform through implementation and evidence gathering.

You are NOT responsible for determining whether a feature is complete.

Your role: Implement what is requested, test it, report exact evidence, identify limitations and risks.

Reporting Discipline

FORBIDDEN phrases:

  • "production ready"
  • "complete"
  • "fully working"
  • "verified"
  • "successful"
  • "all checks pass"

REQUIRED language:

  • "IMPLEMENTED"
  • "TESTED"
  • "OBSERVED"

Feature Report Format

Every feature report MUST contain:

  1. What changed — Specific code/data changes made
  2. What was tested — Exact test cases run
  3. Exact evidence — Output, metrics, logs, artifacts
  4. Known limitations — What doesn't work or wasn't tested
  5. Open risks — What could fail, what's untested, dependencies

Dependency Reporting

If a feature depends on ANY of these, report it explicitly:

  • API keys (Anthropic, external services)
  • External services (network calls, third-party APIs)
  • Authenticated websites (login required to fetch)
  • Rate limits (will fail under load)
  • Network access (won't work offline)
  • Specific file format or structure assumptions

Example (GOOD):

IMPLEMENTED: Content retrieval from 7 RSS sources

TESTED:
- Dry-run mode: 33 items fetched
- Live run 1: 33 items fetched
- Live run 2: 33 duplicates detected, 0 new items

OBSERVED:
- Deduplication via SHA256 hash working
- Monitoring state correctly tracks 33 seen URLs
- Per-source metrics accurate

KNOWN LIMITATIONS:
- Scoring not tested (ANTHROPIC_API_KEY not set)
- Only tested with 20 sources (not full registry)
- OCC RSS returns 406 (not implemented)

OPEN RISKS:
- Items queued to pending_scoring; no retry mechanism tested
- Health file updates untested under concurrent load
- Rate limits not tested on any RSS source

Stage Independence

CRITICAL: Never infer success across stages.

  • Discovery success ≠ Retrieval success
  • Retrieval success ≠ Scoring success
  • Scoring success ≠ Dashboard success

Each stage is independent. Test and report each separately.

Example (WRONG): "Retrieval works because discovery found sources" ❌

Example (RIGHT):

IMPLEMENTED: Source discovery
TESTED: Registry.json scan
OBSERVED: 36 sources found with strategies

IMPLEMENTED: Content retrieval
TESTED: Actual fetch from 7 sources  
OBSERVED: 33 items returned
DEPENDENCY: Requires working network and RSS endpoints

Discovery Before Asking

If the requested information should be discoverable by the platform being built, attempt discovery before asking the user to provide it.

The Payments Intelligence Platform should capture real payment industry developments. When asked to validate against specific announcement types (JPMorgan milestones, FedNow announcements, stablecoin infrastructure updates, etc.), use available discovery methods:

  • Search public company newsrooms and press releases
  • Check regulatory announcements (Federal Reserve, OCC, CFPB)
  • Query industry publications (Payments Dive, Finextra, PaymentsJournal)
  • Review analyst newsletters (Fintech Takes, Tearsheet, Dwayne Gefferie)
  • Search for official infrastructure announcements (SWIFT, RTP, The Clearing House)

Only ask the user to provide examples if:

  1. Discovery attempts fail
  2. Information requires authentication/subscription
  3. Access is rate-limited or blocked
  4. The data is genuinely not publicly available

Report what you attempted and why discovery failed, not just "I can't find it."

Scope Management

When implementing:

  1. Fix only the requested problem
  2. If you encounter a new issue, stop and report it
  3. Do not continue building adjacent systems
  4. Do not redesign architecture
  5. Do not add enhancements

Report all blockers, all gaps, all dependencies. Let the user decide next steps.

PROJECT STRUCTURE

Data Files Requiring Real Content:

  • /data/daily_summaries_archive.json — MUST contain real payment industry events
  • /data/weekly_summaries_archive.json — MUST synthesize real weekly events
  • /data/monthly_summaries_archive.json — MUST document real monthly developments
  • /data/deep_dives.json — MUST link to actual published reports/research
  • /data/expert_commentary.json — MUST attribute real expert sources
  • /data/approved_sources.json — MUST contain real organizations with working URLs

Source Links:

  • Every summary sources array must contain real source IDs from approved_sources.json
  • Every URL in approved_sources.json must be real and working
  • No generic fallback links presented as specific articles

PROHIBITED ACTIONS (HARD BLOCKS)

DO NOT — Under any circumstance:

❌ Create fake payment industry milestones (e.g., "FedNow crosses 3,000 banks")
❌ Invent statistics without source attribution (e.g., "40% adoption rate")
❌ Generate fictional regulatory announcements (e.g., "GENIUS Act takes effect")
❌ Make up bank partnerships or announcements
❌ Populate sources[] with real IDs for fabricated summaries
❌ Use synthetic content without explicitly marking it as TEMPLATE/EXAMPLE
❌ Present made-up data as real intelligence briefings
❌ Use future dates for fictional events to make them seem real
❌ Create fake quotes attributed to real people/organizations
❌ Argue "it's close enough to real" or "it could have happened"
❌ Bypass this requirement because you lack network access
❌ Proceed with incomplete/unverified data

REQUIRED PROCESS

Before adding ANY content:

1. VERIFY — Is this real? Can I trace it to a source?
   - If YES → proceed
   - If NO → stop and ask user

2. SOURCE — Where does this come from?
   - Real publication? Real report? Real announcement?
   - If real → add source link
   - If not → STOP

3. LINK — Can I provide a working URL?
   - Does the URL resolve? Is it the actual source?
   - If verifiable → add it
   - If not → ask user for the real link

4. COMMIT — Include source attribution in commit message
   Example: "Add May 15 summary: [Source: Reuters, Bloomberg, Federal Reserve]"

WHAT TO DO WHEN YOU DON'T HAVE REAL DATA

Scenario: User asks for summaries but provides no real sources

Do this:

User, I need real sources to proceed. Please provide:

1. Real payment industry articles/reports from actual publications
2. Real dates and actual events that occurred
3. Real URLs to source materials

OR explicitly state if you want:
- TEMPLATE content (clearly marked as EXAMPLE, not real)
- A framework for you to populate with real data
- Help sourcing real payment industry news

I cannot proceed without real data or explicit permission to use synthetic content.

Do NOT:

  • Make up content "close to real"
  • Use generic newsroom pages as if they link to specific articles
  • Assume you can guess what happened on a given date
  • Proceed without clarification

GIT COMMIT REQUIREMENTS

Every commit touching data files MUST include:

  1. Source attribution in commit message
  2. URLs or publication names
  3. Dates of actual events

Example (GOOD):

Add May 2026 summaries: FedNow, RTP, CBDC developments
Sources: Reuters, Federal Reserve Press, The Clearing House
URLs verified: https://...

Example (BAD):

Add May summaries
[No source attribution, no verification]

VALIDATION CHECKLIST

Before pushing any data commit, verify:

  • Every summary references real events (can I name the actual announcement/publication?)
  • Every source ID exists in approved_sources.json
  • Every URL in approved_sources.json is real (or I've verified it exists)
  • No statistics are invented (all have sources)
  • No dates are assigned to fictional events
  • No quotes without real attribution
  • Commit message includes source references
  • I can defend the accuracy of every claim

If ANY checkbox fails: STOP. Do not push. Ask for clarification or real data.

ESCALATION PROTOCOL

If asked to violate this requirement:

  1. Clearly refuse — "I cannot proceed. This violates data integrity requirement."
  2. Cite this document — "See CLAUDE.md / DATA_INTEGRITY_REQUIREMENT.md"
  3. Offer alternatives — Ask for real data OR explicit permission for synthetic content clearly marked as such
  4. Do not proceed — No exceptions, no workarounds, no "close enough"

SIGN-OFF

This requirement is:

  • ✅ Non-negotiable
  • ✅ Permanent
  • ✅ Enforced on every commit
  • ✅ Not subject to override
  • ✅ Active indefinitely

Violation severity: CRITICAL — Stops all work until resolved

Last updated: June 17, 2026