Notre avis
Système de recherche avancé combinant Goal-Oriented Action Planning (GOAP) avec vérification cryptographique Ed25519 pour garantir l'authenticité des sources et prévenir les hallucinations.
Points forts
- Vérification cryptographique des sources via signatures Ed25519
- Chaîne de signatures garantissant l'intégrité des citations
- Seuil de confiance configurable (0,85 par défaut, 0,95 strict)
- Journal de vérification signé pour une traçabilité complète
Limites
- Nécessite une configuration initiale de liste blanche d'émetteurs de confiance
- Peut ralentir la recherche en raison de multiples vérifications cryptographiques
- Dépend de la disponibilité de sources signées cryptographiquement
Idéal pour des recherches à enjeux élevés (juridiques, médicales, intelligence économique) où la véracité absolue des faits est critique.
À éviter pour des recherches exploratoires ou informelles où la rapidité prime sur la vérification approfondie.
Analyse de sécurité
SûrThis skill is a descriptive methodology for research with cryptographic verification. It contains no executable commands, file operations, or data exfiltration instructions. The mention of cryptographic keys is conceptual and not a security risk.
Aucun point d'attention détecté
Exemples
Use verified research with Ed25519 cryptographic verification to answer: What are the latest clinical trial results for mRNA vaccines against cancer? Ensure all sources are signed and provide a signature chain.Perform an anti-hallucination fact-check on the claim 'AI will replace all doctors by 2030' using GOAP with Ed25519 verification. Only include sources with cryptographic signatures.Conduct a trusted sources only analysis of the geopolitical stability in the South China Sea using cryptographic verification. Require a signature depth of at least 2 and a confidence threshold of 0.9.name: goap-research-ed25519 description: Advanced GOAP research system with Ed25519 cryptographic verification for anti-hallucination protection. Combines Goal-Oriented Action Planning with cryptographic signatures to ensure source authenticity, claim verification, and citation chain integrity. Use for high-stakes research requiring verifiable facts, competitive intelligence, legal/medical research, or any context where hallucination prevention is critical. Triggers on "verified research", "trusted sources only", "anti-hallucination", "signed sources", "cryptographic verification".
GOAP Research Skill with Ed25519 Verification
Advanced research system combining Goal-Oriented Action Planning (GOAP) with Ed25519 cryptographic verification for maximum anti-hallucination protection.
Key Differentiators from Standard GOAP
| Feature | Standard GOAP | GOAP-Ed25519 | |---------|---------------|--------------| | Source Trust | Reliability scoring (1-5) | Cryptographic signatures + scoring | | Claim Verification | Cross-reference | Cross-reference + signature chain | | Anti-Hallucination | Triangulation | Triangulation + mandatory citations | | Audit Trail | Research path log | Signed verification ledger | | Trust Anchors | Editorial reputation | Trusted issuer whitelist |
Ed25519 Verification Architecture
┌─────────────────────────────────────────────────────────────┐
│ RESEARCH PIPELINE │
├─────────────────────────────────────────────────────────────┤
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Source │───▶│ Ed25519 │───▶│ Verified │ │
│ │ Content │ │ Verifier │ │ Facts │ │
│ └──────────┘ └──────────┘ └──────────┘ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Citation │ │ Signature│ │ Confidence│ │
│ │ Extractor│ │ Chain │ │ Calculator│ │
│ └──────────┘ └──────────┘ └──────────┘ │
│ │ │ │ │
│ └──────────────┴───────────────┘ │
│ │ │
│ ▼ │
│ ┌────────────────┐ │
│ │ Verification │ │
│ │ Ledger (signed)│ │
│ └────────────────┘ │
└─────────────────────────────────────────────────────────────┘
Core GOAP Methodology (Enhanced)
Phase 1: State Assessment with Trust Initialization
Define Current State:
- Existing knowledge about the topic
- Available sources and access constraints
- NEW: Trusted issuer whitelist configuration
- NEW: Ed25519 keypair availability
- Time and depth requirements
Define Goal State:
- Specific questions to answer
- Required evidence types
- NEW: Minimum verification threshold (0.85 default, 0.95 strict)
- NEW: Required signature chain depth
- Confidence thresholds for conclusions
Gap Analysis:
- Knowledge gaps to fill
- NEW: Unsigned claims requiring verification
- NEW: Citation chain breaks to resolve
Phase 2: Action Inventory (Extended)
Research actions with Ed25519 verification extensions:
| Action | Preconditions | Effects | Cost | Verification |
|--------|---------------|---------|------|--------------|
| web_search_broad | topic_defined | candidates_found | 1 | None |
| web_search_verified | topic_defined, trusted_issuers_set | verified_candidates_found | 2 | Domain signature |
| fetch_source | url_known | content_retrieved | 2 | TLS certificate |
| fetch_signed_source | url_known, issuer_pubkey | signed_content_retrieved | 3 | Ed25519 signature |
| extract_facts | content_retrieved | facts_cataloged | 1 | None |
| sign_extracted_facts | facts_cataloged, private_key | signed_facts | 2 | Self-signature |
| verify_claim | claim_identified | claim_verified/refuted | 3 | Cross-ref |
| verify_claim_cryptographic | claim_identified, signature_available | cryptographically_verified | 4 | Ed25519 verify |
| cross_reference | multiple_sources | consistency_checked | 2 | None |
| cross_reference_signed | multiple_signed_sources | signed_consistency_checked | 3 | Multi-sig verify |
| build_citation_chain | facts_cataloged | citation_chain_complete | 2 | Chain integrity |
| verify_citation_chain | citation_chain_complete | chain_verified | 3 | Ed25519 chain |
| generate_signed_report | conclusions_formed, private_key | signed_report_delivered | 3 | Report signature |
Phase 3: Plan Generation (A* with Verification Cost)
Enhanced cost function:
f(n) = g(n) + h(n) + v(n)
g(n): Actual cost (searches performed, time spent)h(n): Heuristic distance to goal (remaining questions)v(n): Verification penalty (unsigned claims × 0.5)
Planning Heuristics:
- Prioritize signed sources over unsigned (lower total cost)
- Prefer sources from trusted issuers whitelist
- Weight cryptographically verified claims higher
- Factor citation chain depth in confidence
Phase 4: OODA Loop with Verification Checkpoints
Observe:
- Monitor search results quality
- Track signature validity status
- Monitor citation chain integrity
- Identify information gaps
Orient:
- Assess if current path leads to goal
- Evaluate cryptographic trust level
- Check for signature chain breaks
- Recognize when verification fails
Decide:
- Continue current research branch or pivot
- Accept or reject unsigned sources
- Trigger re-verification on suspicious content
- Choose between depth and breadth
Act:
- Execute next optimal action
- Sign verified findings
- Update verification ledger
- Trigger replanning if deviation detected
Phase 5: Dynamic Replanning with Trust Recalculation
Trigger replanning when:
- Key assumption invalidated
- Signature verification fails
- Trusted issuer removed from whitelist
- Citation chain broken
- Higher-quality signed source discovered
Ed25519 Verification Protocol
Trusted Issuers Whitelist
Default trusted issuers (Level 5 sources):
trusted_issuers:
news:
- reuters.com
- ap.org
- bbc.com
academic:
- arxiv.org
- nature.com
- science.org
- pubmed.gov
government:
- .gov domains
- .gov.uk domains
- europa.eu
financial:
- sec.gov
- federalreserve.gov
- ecb.europa.eu
Signature Verification Flow
1. Source provides content + signature + public_key_id
2. Fetch public key from trusted keyserver or issuer
3. Verify Ed25519 signature: verify(signature, content_hash, public_key)
4. Check issuer against trusted whitelist
5. Record verification result in ledger
6. Assign cryptographic trust score
Citation Chain Verification
Each fact in the chain must have:
{
"claim": "The statement being made",
"source_url": "https://...",
"source_hash": "sha256:abc123...",
"issuer_pubkey": "ed25519:xyz789...",
"signature": "ed25519_sig:...",
"timestamp": "2025-01-22T10:30:00Z",
"parent_citation": "chain_id:previous_fact_id",
"confidence": 0.95
}
Chain verification:
- Verify each fact's signature individually
- Verify chain integrity (parent hashes match)
- Check all issuers are in trusted whitelist
- Calculate aggregate chain confidence
Anti-Hallucination Rules
100% Citation Rule
Every factual claim MUST have a verifiable source. No exceptions.
❌ FORBIDDEN: "Studies show that X leads to Y"
✅ REQUIRED: "A 2024 study published in Nature (DOI: 10.1038/...) found that X leads to Y"
Verification Thresholds
| Mode | Threshold | Use Case |
|------|-----------|----------|
| development | 0.75 | Exploratory research, brainstorming |
| moderate | 0.85 | Standard research (default) |
| strict | 0.95 | Legal, medical, financial research |
| paranoid | 0.99 | Critical decisions, published reports |
Confidence Calculation
confidence = base_reliability × verification_multiplier × recency_factor
where:
base_reliability = source level (1-5) / 5
verification_multiplier = 1.0 if unsigned, 1.2 if signed, 1.5 if chain_verified
recency_factor = 1.0 if <24h, 0.9 if <7d, 0.8 if <30d, 0.6 if older
Research Execution Patterns (Enhanced)
Pattern A: Verified Exploratory Research
Goal: comprehensive_verified_understanding
Actions:
1. configure_trusted_issuers → whitelist_active
2. web_search_verified → verified_candidates_found
3. FOR EACH candidate: fetch_signed_source → signed_content_retrieved
4. extract_facts → facts_cataloged
5. sign_extracted_facts → signed_facts
6. cross_reference_signed → signed_consistency_checked
7. build_citation_chain → citation_chain_complete
8. verify_citation_chain → chain_verified
9. synthesize_findings → conclusions_formed
10. generate_signed_report → signed_report_delivered
Pattern B: High-Stakes Fact Verification
Goal: cryptographically_verified_claim
Mode: strict (0.95 threshold)
Actions:
1. identify_claim → claim_defined
2. configure_trusted_issuers (strict list) → whitelist_active
3. find_primary_source → primary_located
4. fetch_signed_source → signed_content_retrieved
5. verify_claim_cryptographic → cryptographically_verified
6. cross_reference_signed (≥3 sources) → multi_source_verified
7. build_citation_chain → citation_chain_complete
8. verify_citation_chain → chain_verified (confidence ≥0.95)
Pattern C: Competitive Analysis with Audit Trail
Goal: auditable_competitive_landscape
Actions:
1. identify_players → competitors_listed
2. configure_trusted_issuers → whitelist_active
3. FOR EACH competitor:
- web_search_verified → verified_info_found
- fetch_signed_source (official sources) → signed_content
- extract_facts → facts_cataloged
- sign_extracted_facts → signed_facts
4. cross_reference_signed → consistency_verified
5. build_citation_chain → full_chain
6. generate_signed_report → auditable_report
Output Structure (Enhanced)
Verified Research Report Format
## Executive Summary
[Key findings in 2-3 sentences]
## Verification Status
- Mode: strict (0.95 threshold)
- Chain Integrity: ✅ VERIFIED
- Unsigned Claims: 0
- Total Citations: 15
- Trusted Issuers Used: 8
## Research Objective
[Original question/goal]
## Methodology
[GOAP plan executed, verification protocol used]
## Verified Findings
### [Subtopic 1]
[Findings with signed inline citations]
**Verification Details:**
| Claim | Source | Signature | Confidence |
|-------|--------|-----------|------------|
| ... | ... | ✅ | 0.96 |
### [Subtopic 2]
[Findings with signed inline citations]
## Confidence Assessment
- Cryptographically Verified (≥0.95): [claims list]
- Cross-Reference Verified (0.85-0.95): [claims list]
- Single Source (0.75-0.85): [claims list]
- Unverified (<0.75): NONE (strict mode)
## Citation Chain
[Full chain with signatures - see Appendix A]
## Verification Ledger
[Signed log of all verification operations]
## Sources
[Numbered list with URLs, signatures, and trust scores]
## Appendix A: Cryptographic Verification Details
[Full Ed25519 signature data for audit]
Quality Standards (Enhanced)
Completeness Checks:
- [ ] All original questions addressed
- [ ] 100% of claims have citations
- [ ] All citations are verifiable
- [ ] Citation chain integrity verified
- [ ] Primary sources found where possible
- [ ] Contradictions identified and addressed
- [ ] Confidence levels assigned to conclusions
- [ ] Verification ledger signed and complete
Anti-Hallucination Checks:
- [ ] No claims without sources
- [ ] No "studies show" without specific citation
- [ ] No statistics without methodology source
- [ ] No quotes without attribution + verification
- [ ] No predictions presented as facts
Implementation
Python Usage
from goap_planner import GOAPResearchPlanner
from ed25519_verifier import Ed25519Verifier
# Initialize with Ed25519 verification
planner = GOAPResearchPlanner(
verification_mode="strict",
trusted_issuers=["reuters.com", "ap.org", "nature.com"]
)
# Generate research plan
plan = planner.plan(
goal_type="verified_exploratory",
topic="AI safety regulations 2025"
)
# Execute with verification
results = planner.execute(plan, verify_all=True)
# Generate signed report
report = planner.generate_report(
results,
sign=True,
include_verification_ledger=True
)
CLI Usage (via goalie integration)
# Install goalie for Ed25519 support
npm install -g goalie
# Verified research
goalie search "Your research question" \
--verify \
--strict-verify \
--trusted-issuers "reuters.com,ap.org,nature.com" \
--mode academic \
--save
# Anti-hallucination check
goalie reason --mode anti-hallucination \
--claims "Your claim to verify" \
--citations "Source URL"
References
For detailed implementations, see:
- references/research-actions.md - Complete action specifications with Ed25519 extensions
- references/source-evaluation.md - Source credibility with cryptographic trust
- references/ed25519-verification.md - Ed25519 protocol details
- scripts/goap_planner.py - Enhanced GOAP planner
- scripts/ed25519_verifier.py - Verification module
Dependencies
- Python 3.9+
@noble/ed25519(Node.js) orcryptography(Python) for Ed25519- Optional:
goalienpm package for CLI integration
Ingénierie de Prompts
Data & IA
Bonnes pratiques et templates de prompt engineering pour maximiser les résultats IA.
Visualisation de Données
Data & IA
Génère des visualisations de données et graphiques adaptés à vos données.
Architecture RAG
Data & IA
Guide de configuration d'architectures RAG (Retrieval-Augmented Generation).