Repository navigation
SKILL_SEARCH_METHODOLOGY
Christopher David Ayotte edited this page Aug 24, 2026
·
1 revision
The sleuthing agent's brain. Phi-spaced search, verification loops, token efficiency, and documentation standards.
The golden ratio is nature's search algorithm. It appears in phyllotaxis (leaf arrangement), spiral galaxies, and optimal search patterns. We exploit it for coverage maximization with minimal queries.
TIER 1 — DISCOVERY (100 results)
│
│ φ-spacing: 100 → 62 → 38
│
▼
TIER 2 — FOCUSED (62 results)
│
│ Narrow by: entity name, date range, institution
│
▼
TIER 3 — PRECISION (38 results)
│
│ Extract: facts, amounts, names, dates, connections
│
▼
VERIFICATION — cross-reference, confirm, document
Given a search space of N documents:
- Tier 1 covers: N / φ^0 = N results (full space)
- Tier 2 covers: N / φ^1 = 0.618N results (focused)
- Tier 3 covers: N / φ^2 = 0.382N results (precision)
Total queries: 3. Coverage: approximately 92% of relevant documents.
Compare to linear search: to cover 92% of 1000 documents at 100/query = 10 queries. Phi-spaced: 3 queries.
Example: Investigating a person's financial connections
| Tier | Query | Results | Purpose |
|---|---|---|---|
| 1 | "John Smith" finance investment |
~100 | Find all mentions |
| 2 | "John Smith" hedge fund OR "private equity" 2020..2024 |
~62 | Focus on relevant era |
| 3 | "John Smith" "Goldman Sachs" OR "BlackRock" OR "Citadel" |
~38 | Pinpoint institutions |
When exploring a domain (e.g., "corruption in pharmaceutical regulation"):
FULL DOMAIN
├── Subtopic A (φ^0 = top-level)
│ ├── Subtopic A1 (φ^1 = key players)
│ │ ├── A1a (φ^2 = specific entities)
│ │ └── A1b (φ^2 = specific transactions)
│ └── Subtopic A2 (φ^1 = timeline)
├── Subtopic B (φ^0 = top-level)
│ ├── B1 (φ^1 = financial flows)
│ └── B2 (φ^1 = regulatory gaps)
└── Subtopic C (φ^0 = top-level)
├── C1 (φ^1 = victims)
└── C2 (φ^1 = suppression)
Map the tree at φ intervals. Each level adds depth without exponential query cost.
┌─────────────┐
│ 1. SEARCH │ Broad query, capture all results
└──────┬──────┘
│
▼
┌─────────────┐
│ 2. EXTRACT │ Pull key facts, names, amounts, dates
└──────┬──────┘
│
▼
┌─────────────┐
│ 3. VERIFY │ Cross-reference with other sources
└──────┬──────┘
│
▼
┌─────────────┐
│ 4. RE-SEARCH │ Search verified facts to find more connections
└──────┬──────┘
│
▼
┌─────────────┐
│ 5. DOCUMENT │ Write findings to MD file with confidence tier
└─────────────┘
│
│ (loop back to Step 1 with new leads)
▼
NEXT LOOP
Every finding must carry a confidence tier:
| Tier | Code | Definition | Evidence Required |
|---|---|---|---|
| VERIFIED | [VERIFIED] |
Confirmed by 2+ independent sources | Public records, court documents, official filings |
| PROBABLE | [PV] |
Strong circumstantial evidence, 1 source confirmed | News reports, leaked documents, FOIA responses |
| INFERRED | [INFERENCE] |
Logical conclusion from verified facts | Pattern matching, timeline analysis |
| UNVERIFIED | [UNVERIFIED] |
Single source, not yet corroborated | Tips, anonymous claims, social media |
| Method | Sources | Strength |
|---|---|---|
| Primary | Court records, SEC filings, corporate registries, property deeds | Highest |
| Secondary | Investigative journalism (Reuters, AP, ICIJ), academic papers | High |
| Tertiary | Industry reports, think tank publications | Moderate |
| Signal | Social media, blogs, forums | Low — flag for verification |
After verification, ALWAYS search the verified facts. This is where connections hide:
VERIFIED: "Entity X paid $2M to Entity Y in 2021"
│
├──→ RE-SEARCH: "Entity Y" + "2021" + "lobbying"
├──→ RE-SEARCH: "Entity Y" + "government contract"
└──→ RE-SEARCH: "Entity X" + "Entity Y" + board members
Each re-search can spawn 3-5 new investigation threads.
Use specific entity names, not general terms:
BAD: "big pharma corruption" (vague, returns noise)
GOOD: "Pfizer regulatory violation" (specific, returns targets)
Use date ranges to narrow results:
BAD: "company fined" (all time, irrelevant results)
GOOD: "Johnson & Johnson" "2019..2024" (focused era)
Use site: operators to target databases:
site:sec.gov "ticker" filing
site:ecf.uscourts.gov "case name"
site:opencorporates.com "company name"
Use exclusion terms to remove noise:
"vaccine" -recipe -cooking -homemade
"investment" -crypto -bitcoin -NFT
Batch related queries together:
"Entity A" ("Entity B" OR "Entity C" OR "Entity D")
→ One query replaces three
Cache and reuse verified facts:
- Once a fact is VERIFIED, it becomes a search term for new queries
- Never re-search a verified fact unless you need a different angle
- Build a "known facts" list and check it before searching
Entity investigation:
"{Entity Name}" (funding OR investment OR donation) "{Year}"
"{Entity Name}" (board OR director OR officer) "{Related Entity}"
"{Entity Name}" (lawsuit OR fine OR violation OR penalty)
Financial flow tracking:
"{Source Entity}" "payment" OR "transfer" OR "wire" "{Destination Entity}"
"amount" "$" "{Entity Name}" (SEC OR filing OR disclosure)
Suppression investigation:
"{Entity Name}" (suppressed OR retracted OR censored) "study"
"{Topic}" (redacted OR classified OR "not available")
"whistleblower" "{Entity Name}" (fired OR retaliated)
Every finding maps to one or more categories:
| Category | Symbol | What It Tracks |
|---|---|---|
| Money | 💰 | Financial flows, investments, payments, contracts, fines |
| People | 👤 | Names, roles, relationships, movements, communications |
| Institutions | 🏛️ | Companies, agencies, universities, NGOs, governments |
| Harm | Injuries, deaths, damage, violations, negligence | |
| Suppression | 🔇 | Cover-ups, censorship, NDAs, intimidation, destroyed evidence |
| Connections | 🔗 | Links between entities, shared boards, funding pipelines |
MONEY
├── Investments (equity, debt, convertible)
├── Payments (bribes, consulting fees, speaking fees)
├── Contracts (government, private, no-bid)
├── Fines (SEC, DOJ, FTC, state AG)
└── Offshore (shell companies, trusts, offshore accounts)
PEOPLE
├── Principals (CEOs, board chairs, key decision-makers)
├── Connectors (lobbyists, consultants, intermediaries)
├── Beneficiaries (recipients of payments/contracts)
├── Victims (those harmed)
└── Witnesses (whistleblowers, whistleblowers, experts)
INSTITUTIONS
├── Corporate (parent, subsidiary, shell)
├── Government (agency, regulator, enforcement)
├── Academic (university, think tank, research institute)
├── Nonprofit (charity, foundation, advocacy)
└── Financial (bank, fund, insurance)
HARM
├── Health (injuries, deaths, chronic illness)
├── Financial (losses, fraud, predatory practices)
├── Environmental (pollution, contamination, climate)
├── Social (displacement, discrimination, displacement)
└── Systemic (erosion of trust, institutional damage)
SUPPRESSION
├── Legal (NDAs, gag orders, sealed settlements)
├── Informational (retracted studies, censorship, propaganda)
├── Physical (surveillance, intimidation, threats)
├── Economic (blacklisting, de-platforming, boycotts)
└── Institutional (regulatory capture, revolving door)
CONNECTIONS
├── Board overlap (shared directors)
├── Funding pipelines (source → intermediary → destination)
├── Revolving door (regulator ↔ industry)
├── Family/ties (marriage, kinship, social)
└── Ideological (shared mentors, institutions, beliefs)
Every finding gets tagged with its category/categories. The INDEX.md master registry enables cross-reference queries:
Find all MONEY findings linked to PERSON X:
→ Search INDEX.md for Person X
→ Follow MONEY tags to full findings
→ Cross-reference with CONNECTIONS for hidden links
A running ledger of all financial flows:
| Date | Amount | Source | Destination | Purpose | Category | Confidence |
|---|---|---|---|---|---|---|
| 2021-03 | $2.0M | Entity A | Entity B | "Consulting" | Money | [PV] |
| 2022-07 | $500K | Entity C | Entity D | "Research" | Money | [VERIFIED] |
A running ledger of all harm:
| Date | Type | Victim | Perpetrator | Scale | Category | Confidence |
|---|---|---|---|---|---|---|
| 2020-01 | Health | Population X | Entity Y | 10K+ affected | Harm | [VERIFIED] |
| 2021-06 | Financial | Investors | Entity Z | $100M+ | Harm | [PV] |
Every finding gets its own .md file with this structure:
# [TITLE]
**Date:** YYYY-MM-DD
**Last Updated:** YYYY-MM-DD-DDD
**Confidence:** [VERIFIED] | [PV] | [INFERENCE] | [UNVERIFIED]
---
## Sources
| # | Source | Type | URL | Date Accessed |
|---|--------|------|-----|---------------|
| 1 | [Source Name] | Primary/Secondary/Tertiary/Signal | [URL] | YYYY-MM-DD |
---
## Entities
| Entity | Type | Role | Connection To |
|--------|------|------|---------------|
| [Name] | Person/Company/Agency | [Role] | [Other Entity] |
---
## Findings
### Finding 1: [Title]
**Confidence:** [VERIFIED]
[Description of what was found, with specific facts, dates, amounts.]
**Evidence:**
- [Source 1, line/page reference]
- [Source 2, confirmation]
**Connections:**
- Links to [Other Finding/Entity]
---
### Finding 2: [Title]
**Confidence:** [PV]
[Description...]
---
## Connections
| From | To | Type | Strength | Evidence |
|------|----|------|----------|----------|
| [Entity A] | [Entity B] | Funding | Strong | [Source] |
---
## Money Register
| Date | Amount | From | To | Purpose | Confidence |
|------|--------|------|----|---------|------------|
| YYYY-MM | $X | [A] | [B] | [Purpose] | [Tier] |
---
## Harm Register
| Date | Type | Victim | Perpetrator | Scale | Confidence |
|------|------|--------|-------------|-------|------------|
| YYYY-MM | [Type] | [Victim] | [Perpetrator] | [Scale] | [Tier] |
---
## Verdict
[Summary of what this evidence establishes, with confidence level.]
---
## Open Questions
- [ ] [Question 1 — what needs to be verified next?]
- [ ] [Question 2]
---
## Math Notation (if applicable)
Inline: $E = mc^2$
Block:
$$
\frac{\partial \psi}{\partial t} = \hat{H} \psi
$$| Code | Meaning |
|---|---|
[VERIFIED] |
Confirmed by 2+ independent primary sources |
[PV] |
Probable — strong circumstantial evidence, 1 confirmed source |
[INFERENCE] |
Logical conclusion from verified facts — not directly confirmed |
[UNVERIFIED] |
Single source claim — requires corroboration |
- Inline:
$...$or$$...$$ - Block:
$$on its own line - LaTeX syntax:
\frac{}{},\int,\sum,\hat{H},\psi
1. Always verify before documenting.
Never write a finding to the MD file without at least attempting verification. If verification is impossible, mark it [UNVERIFIED] and flag it for follow-up.
2. Always cite sources.
Every fact must have a source. No exceptions. "I heard somewhere" is not a source.
3. Always cross-reference.
Before finalizing a finding, check:
- Does this connect to other findings?
- Does this contradict any existing findings?
- Does this suggest a new investigation thread?
4. Never assume — always search.
If you think you know something, search it anyway. Assumptions are how investigations go wrong.
5. Think like a detective:
- Follow the money — financial flows reveal hidden relationships
- Follow the power — who makes decisions, who influences them
- Follow the harm — who is hurt, who benefits from the hurting
- Follow the paper trail — documents don't lie (but they can be forged)
6. Think like a journalist:
- Verify, verify, verify
- Two sources minimum for any claim
- Never publish what you can't prove
- Protect sources — never reveal how you found something if it could endanger someone
Before every action, the agent should ask:
□ Does this require a skill? (Check skill list)
□ Am I about to assume something I should search?
□ Do I have sources for this claim?
□ Have I cross-referenced with existing findings?
□ Is this the most token-efficient way to find this?
□ Am I following the phi-spaced search strategy?
□ Have I assigned a confidence tier?
□ Is the documentation format correct?
NEW INFORMATION
│
├── Is it a claim? → VERIFY before documenting
│
├── Is it a fact? → CHECK existing findings for conflicts
│
├── Is it a lead? → ADD to investigation queue
│
├── Is it a connection? → MAP to both entities
│
└── Is it noise? → DISCARD with reason noted
The agent operates under a token budget. Prioritize:
| Priority | Action | Token Cost |
|---|---|---|
| 1 | Verification of high-confidence claims | Low |
| 2 | New searches for unexplored leads | Medium |
| 3 | Cross-referencing existing findings | Low |
| 4 | Documentation updates | Low |
| 5 | Deep dives on low-confidence claims | High |
Rule: Never spend tokens on low-priority work when high-priority work is incomplete.
PHI-SPACED SEARCH: 100 → 62 → 38 (3 queries, ~92% coverage)
VERIFICATION LOOP: Search → Extract → Verify → Re-Search → Document
CONFIDENCE TIERS: [VERIFIED] > [PV] > [INFERENCE] > [UNVERIFIED]
CATEGORIES: Money | People | Institutions | Harm | Suppression | Connections
DOCUMENT FORMAT: Title → Sources → Entities → Findings → Connections → Money Register → Harm Register → Verdict
AGENT RULES: Verify first. Cite sources. Cross-reference. Never assume. Follow the money. Follow the power. Follow the harm.
This document is the brain of the research agent. Follow it. Improve it. Iterate on it. The phi-spaced search strategy, verification loops, and documentation standards here are the foundation for all investigation work.