Most AI either:
- Hallucinates confidently (no grounding)
- Requires cloud APIs (privacy/cost issues)
- Lacks auditability (black box reasoning)
- Struggles with multi-domain synthesis
I wanted something that could handle real complexity while being
verifiable, private, and economical.
## The Approach
*Symbolic compression via semiotic triads:*
- Each symbol = concept + semiotic triad
- Example: POWER-DRIFT-TACTIC →
- 1000+ symbols across domains in ~7k token overhead
- Hierarchical references enable deep retrieval without context explosion
*Uncertainty detection + web grounding:*
- System recognizes knowledge gaps
- Triggers web search for validation
- Integrates results before responding
- No hallucination on factual claims
*Local inference:*
- Runs on M4 Max (gpt-oss-120B quantized)
- ~30 second responses
- Zero API costs after hardware
- Complete data privacy
*Full symbolic traceability:*
- Every claim linked to source symbol
- Complete reasoning chain logged
- Audit trail for regulatory compliance
- 29 domains modeled
- Base: gpt-oss-120B (local quantized inference)
- Symbols: 1000+ hand-curated across 7+ domains
- Compression: Semiotic triads + hierarchical references
- Tools: Ephemeral execution with validation retries
- Context: ~15k tokens average (including output)
- Grounding: Web search on uncertainty detection
## Why Symbolic?
Vector embeddings are great for retrieval but terrible for reasoning
chains. Symbols provide:
1. *Composability* - combine across domains coherently
2. *Traceability* - explicit reasoning paths
3. *Efficiency* - massive compression via references
4. *Verifiability* - audit every claim to source
The emoji triads act as semantic anchors that survive context
compression while remaining human-readable.
## Use Cases Tested
- OSINT / disinformation analysis
- Bioethics committee decisions
- Pharmaceutical regulatory pathways
- Environmental impact assessment
- Academic research synthesis
- Medical triage (flags mental health concerns appropriately)
All demos live on site with full outputs.
## Current Status
Still figuring out productization. Core question: is the auditability
+ local inference + multi-domain synthesis combination valuable enough
to matter for production use cases?
Open to feedback on:
1. Architecture improvements
2. Symbol library design
3. Real-world applications
4. Technical tradeoffs
Happy to run test analyses for anyone curious. Looking for validation
that this approach has legs beyond being technically interesting.
*Context management:*
- Load symbol stubs (50 tokens each) into context
- Full definitions retrieved only when activated
- Ephemeral tool execution keeps working memory clean
- Triads enable rapid pattern matching with ultra small compression of concepts.
*Validation loop:*
Tool call → Parse → Validate → Retry if malformed (max 3×)
Achieves 99%+ compliance vs ~60% without validation
*Web grounding trigger:*
If (uncertainty_detected && factual_claim_present):
web_search(targeted_query)
integrate_results()
cite_sources()
The system knows what it doesn't know.
---
Built this because I was frustrated with AI that couldn't show its
work. Turns out symbolic reasoning + modern LLMs + proper engineering
= actually useful for complex decisions.
Demo: https://signal-zero.ai/demo.html
## The Problem
Most AI either: - Hallucinates confidently (no grounding) - Requires cloud APIs (privacy/cost issues) - Lacks auditability (black box reasoning) - Struggles with multi-domain synthesis
I wanted something that could handle real complexity while being verifiable, private, and economical.
## The Approach
*Symbolic compression via semiotic triads:* - Each symbol = concept + semiotic triad - Example: POWER-DRIFT-TACTIC → - 1000+ symbols across domains in ~7k token overhead - Hierarchical references enable deep retrieval without context explosion
*Uncertainty detection + web grounding:* - System recognizes knowledge gaps - Triggers web search for validation - Integrates results before responding - No hallucination on factual claims
*Local inference:* - Runs on M4 Max (gpt-oss-120B quantized) - ~30 second responses - Zero API costs after hardware - Complete data privacy
*Full symbolic traceability:* - Every claim linked to source symbol - Complete reasoning chain logged - Audit trail for regulatory compliance - 29 domains modeled
## Example Output
https://signal-zero.ai/examples.html
## Technical Stack
- Base: gpt-oss-120B (local quantized inference) - Symbols: 1000+ hand-curated across 7+ domains - Compression: Semiotic triads + hierarchical references - Tools: Ephemeral execution with validation retries - Context: ~15k tokens average (including output) - Grounding: Web search on uncertainty detection
## Why Symbolic?
Vector embeddings are great for retrieval but terrible for reasoning chains. Symbols provide:
1. *Composability* - combine across domains coherently 2. *Traceability* - explicit reasoning paths 3. *Efficiency* - massive compression via references 4. *Verifiability* - audit every claim to source
The emoji triads act as semantic anchors that survive context compression while remaining human-readable.
## Use Cases Tested
- OSINT / disinformation analysis - Bioethics committee decisions - Pharmaceutical regulatory pathways - Environmental impact assessment - Academic research synthesis - Medical triage (flags mental health concerns appropriately)
All demos live on site with full outputs.
## Current Status
Still figuring out productization. Core question: is the auditability + local inference + multi-domain synthesis combination valuable enough to matter for production use cases?
Open to feedback on: 1. Architecture improvements 2. Symbol library design 3. Real-world applications 4. Technical tradeoffs
Happy to run test analyses for anyone curious. Looking for validation that this approach has legs beyond being technically interesting.
---
Tech details for the architecture-curious:
*Symbol structure:*
{ id: "POWER-DRIFT-TACTIC", triad: "", domain: "negotiation", definition: "Gradual shift of authority...", related: ["NAR-LOOP", "SOFT-GRIND-COLLAPSE"] }
*Context management:* - Load symbol stubs (50 tokens each) into context - Full definitions retrieved only when activated - Ephemeral tool execution keeps working memory clean - Triads enable rapid pattern matching with ultra small compression of concepts.
*Validation loop:* Tool call → Parse → Validate → Retry if malformed (max 3×) Achieves 99%+ compliance vs ~60% without validation
*Web grounding trigger:* If (uncertainty_detected && factual_claim_present): web_search(targeted_query) integrate_results() cite_sources()
The system knows what it doesn't know.
---
Built this because I was frustrated with AI that couldn't show its work. Turns out symbolic reasoning + modern LLMs + proper engineering = actually useful for complex decisions.
Thoughts?