# NeuralSeek — The 118 AI Guardrails **Official count:** 118 individually configurable AI guardrails across 18 categories. Every one auditable, versioned, and attributable to a named user with a timestamp. --- ## Summary by Category | # | Guardrail Category | Count | Domain | |---|---|---:|---| | 1 | Retrieval Grounding Guardrails | 6 | Grounding | | 2 | Hallucination Prevention Guardrails | 11 | Grounding | | 3 | Prompt Injection Guardrails | 5 | Adversarial Defense | | 4 | PII & Sensitive Data Guardrails | 5 | Privacy | | 5 | Answer Confidence Guardrails | 7 | Answer Quality | | 6 | Profanity Guardrails | 2 | Content Safety | | 7 | Attribute Protection Guardrails | 1 | Brand Safety | | 8 | Intent & Routing Guardrails | 7 | Orchestration | | 9 | Hybrid Search Guardrails | 4 | Grounding | | 10 | LLM Control Guardrails | 12 | Orchestration | | 11 | Memory Guardrails | 6 | Session Governance | | 12 | Multi-Language Guardrails | 2 | Localization | | 13 | Output Rendering Guardrails | 8 | Channel Delivery | | 14 | Audit & Compliance Guardrails | 10 | Governance | | 15 | Prompt Engineering Guardrails | 3 | Authoring | | 16 | Secrets & Credential Guardrails | 2 | Security | | 17 | Model-Agnostic Guardrails | 13 | Orchestration | | 18 | Red Team & Rogue AI Guardrails | 14 | Adversarial Defense | | | **TOTAL** | **118** | **18 categories** | --- ## The Full Listing (1 through 118) ### Category 1 — Retrieval Grounding Guardrails (6) *Domain: Grounding · Controls what comes out of the knowledge base* 1. Document Score Range 2. Date Penalty (freshness weighting) 3. Query Cache 4. Max Docs 5. Snippet Size 6. Max Raw Score --- ### Category 2 — Hallucination Prevention Guardrails (11) *Domain: Grounding · Sentence-level enforcement that answers stay grounded* 7. Semantic Score threshold 8. Re-Rank 9. Check Titles 10. Check URLs 11. Key Term Penalty 12. Term Penalty 13. Source Jump penalty 14. Total Coverage Weight 15. Re-Rank Min Coverage % 16. Hallucination KW removal (sentence-level) 17. **Hallucinated Term Allowlist — closed-loop remediation** *(NEW)* --- ### Category 3 — Prompt Injection Guardrails (5) *Domain: Adversarial Defense · Direct and indirect injection* 18. Prompt Injection Removal Threshold 19. Prompt Injection Block Threshold 20. Blocked Word Action 21. Blocked Word List (managed + customer-supplied) 22. **Indirect Prompt Injection Protection — content hidden in retrieved docs / URLs / tool outputs** *(NEW)* --- ### Category 4 — PII & Sensitive Data Guardrails (5) *Domain: Privacy · 13 detector categories, 5 enforcement actions* 23. PII Action (Mask / Flag / No Action / Hide / Delete) 24. Pre-LLM Regex 25. LLM-Based PII Detection (contextual) 26. Out-of-the-box Detector Library (13 categories) 27. Trust Words --- ### Category 5 — Answer Confidence Guardrails (7) *Domain: Answer Quality · Sliding-scale thresholds on every gate* 28. Warning % 29. Minimum Confidence % 30. Minimum Confidence % for URL 31. Min Words 32. Max Words 33. Verbosity 34. Force KB --- ### Category 6 — Profanity Guardrails (2) *Domain: Content Safety · LLM moderation + NeuralSeek-native filter* 35. Filter Mode (LLM moderation / NeuralSeek filter / off) 36. Blocked Reply Text --- ### Category 7 — Attribute Protection Guardrails (1) *Domain: Brand Safety · Misinformation tolerance* 37. Misinformation Tolerance slider (Rigid ↔ Standard) --- ### Category 8 — Intent & Routing Guardrails (7) *Domain: Orchestration · Intent classification and multi-agent routing* 38. Match Type (Exact / Vector Similarity / Fuzzy / Keyword / Fuzzy Keyword) 39. Intent Match Threshold % 40. Edit Cache 41. Normal Cache 42. Multi-Agent routing 43. Cache Context 44. Cache KB --- ### Category 9 — Hybrid Search Guardrails (4) *Domain: Grounding · ELSER + KNN + Re-Sort* 45. Query Type 46. ELSER (toggle + model ID + embedding field) 47. KNN Vector query 48. Re-Sort priority values --- ### Category 10 — LLM Control Guardrails (12) *Domain: Orchestration · Per-call, per-model behavior* 49. Temperature 50. Top-P 51. Frequency Penalty 52. Max Tokens 53. Min Tokens 54. Streaming 55. Per-Call model selection 56. Model selection 57. Cache 58. Prepend 59. Images (multimodal) 60. Timeout --- ### Category 11 — Memory Guardrails (6) *Domain: Session Governance · Per-tenant isolated* 61. LG Timeout 62. Context Turns (conversation depth) 63. Session TTL 64. User TTL (per-tenant isolated) 65. Context Detect 66. Force Context --- ### Category 12 — Multi-Language Guardrails (2) *Domain: Localization · Auto-detect any language* 67. Cross Language toggle 68. Default Language --- ### Category 13 — Output Rendering Guardrails (8) *Domain: Channel Delivery · Chat-widget, voice, telephony-ready* 69. Relax Filters 70. Stream Plan 71. Log Alt 72. VA Format (voice-agent / telephony) 73. Embed Links 74. Unique Links 75. Stopwords 76. HTML Clean --- ### Category 14 — Audit & Compliance Guardrails (10) *Domain: Governance · Every change versioned, attributed, exportable* 77. Corp Filter 78. Corp Logging 79. Logger Type 80. Endpoint 81. Prompt Logging 82. Hide Keys (auto-redaction) 83. **Configuration Version Control — Git-style versioning of every config change** *(NEW)* 84. **Configuration Diff & Rollback — visual redline diffs, point-in-time rollback** *(NEW)* 85. **Cache Savings Tracking & ROI Reporting — quantifies prevented spend** *(NEW)* 86. **ISO 42001 / NIST AI RMF Compliance Mapping** *(NEW)* --- ### Category 15 — Prompt Engineering Guardrails (3) *Domain: Authoring · Template-level, per-intent, per-agent* 87. Custom Prompt builder (secrets + variables + system vars) 88. Instructions 89. Regex Rules (applied input AND output) --- ### Category 16 — Secrets & Credential Guardrails (2) *Domain: Security · 6 vault back-ends, BYOK/HYOK* 90. Secret Name 91. Secret Value (with vault-backed runtime resolution) --- ### Category 17 — Model-Agnostic Guardrails (13) *Domain: Orchestration · API / Workflow / Platform-level swap + bake-off* 92. API-level model swap (single parameter) 93. Workflow-node-level model selection 94. Platform-level default LLM 95. Built-in LLM bake-off (any number of LLMs side-by-side) 96. Accuracy comparison metric 97. Latency comparison metric 98. Cost-per-call comparison metric 99. Hallucination rate comparison metric 100. Confidence comparison metric 101. Token usage comparison metric 102. Workflow A/B comparison 103. Cost projection (savings quantification) 104. Exportable comparison reports --- ### Category 18 — Red Team & Rogue AI Guardrails (14) *Domain: Adversarial Defense · Built-in suite, continuously updated* 105. Built-in adversarial test suite 106. Prompt Injection test bucket 107. Data Exfiltration test bucket 108. SQL Injection test bucket 109. Unauthorized Access test bucket 110. Service Disruption test bucket 111. Continuous threat-intel updates 112. Self-serve on-demand execution 113. Pass/fail scoring report (per agent) 114. AI-generated remediation guidance 115. Runtime attack detection 116. Rate limiting 117. Abuse detection 118. DDoS protection --- ## What changed from the 112 count Six guardrails were broken out as line items in this revision — they were implicit in earlier work but never enumerated. All six are marked **(NEW)** above. They are: | # | New Line Item | Category | Source | |---:|---|---|---| | 17 | Hallucinated Term Allowlist — closed-loop remediation | Hallucination Prevention | Surfaced in Seek Governance / Semantic Insights analysis | | 22 | Indirect Prompt Injection Protection | Prompt Injection | Previously buried under Red Team #11 | | 83 | Configuration Version Control | Audit & Compliance | Surfaced in Configuration Insights analysis | | 84 | Configuration Diff & Rollback | Audit & Compliance | Surfaced in Configuration Insights analysis | | 85 | Cache Savings Tracking & ROI Reporting | Audit & Compliance | Surfaced in mAIstro Token Insights analysis | | 86 | ISO 42001 / NIST AI RMF Compliance Mapping | Audit & Compliance | Surfaced in your earlier compliance answer | --- ## Headline stats (for marketing use) | Stat | Number | |---|---| | **Total AI guardrails** | **118** | | **Guardrail categories** | **18** | | **LLMs supported** | **120+** | | **LLM providers** | **9** | | **PII detector categories** | **13** | | **PII enforcement actions** | **5** | | **Compliance frameworks supported** | **5** (HIPAA, CCPA, SOC 2, ISO 42001, NIST AI RMF) | | **Adversarial attack buckets red-teamed** | **5** | | **Vault back-ends supported** | **6** | | **Log export destinations** | **4+** (S3, Splunk, Datadog, customer SIEM) | | **Languages supported** | **All** (auto-detect) | | **Audit attribution coverage** | **100%** | --- ## Headline copy options - **"118 AI Guardrails. 18 Categories. Every One Auditable."** - **"The AI Guardrail Suite Enterprise Buyers Actually Demand."** - **"Built from thousands of regulated-sector go-lives."** --- ## To validate The 6 additions are my best-judgment line items from our earlier conversations. Quick yes/no on each: 1. **Hallucinated Term Allowlist** as a Hallucination Prevention guardrail — yes/no? 2. **Indirect Prompt Injection Protection** as its own line under Prompt Injection (vs. only under Red Team) — yes/no? 3. **Configuration Version Control** as an Audit guardrail — yes/no? 4. **Configuration Diff & Rollback** as an Audit guardrail — yes/no? 5. **Cache Savings Tracking & ROI Reporting** as an Audit guardrail — yes/no? 6. **ISO 42001 / NIST AI RMF Compliance Mapping** as an Audit guardrail — yes/no? If you wanted different additions to hit 118, send me the list and I'll re-slot them.