# Ejentum

> Reasoning Harness for Agentic AI. A REST API and MCP server that injects engineered cognitive operations into LLM agents at inference time, blocking the four failure modes that cause multi-step agents to fail: attention drift, premature convergence, sycophantic acceptance of bad premises, and fabricated intermediate state.
>
> Full technical manual and complete architectural specification: [llms-full.txt](https://ejentum.com/llms-full.txt)

Ejentum builds the **Reasoning Harness** layer for agentic AI.

The deployed production engine powers **311 deterministic cognitive operations** across 6 fundamental dimensions (Causal, Temporal, Spatial, Simulation, Abstraction, Metacognition) linked by 933 directed synergy edges. Each operation is a structured procedure with explicit verification invariants, failure patterns to block, correct reasoning procedure, verification tests, and suppression/amplification vectors. One API call. Works with any LLM. No fine-tuning.

Two delivery modes per harness:
- **Dynamic**: single or multi-hop graph retrieval, predictable injection (<2ms server engine lookup; ~90-150ms cross-region WAN).
- **Adaptive**: top-k retrieval rewritten by an adapter LLM to the specific task identifiers. Cognitive operation reshapes itself for the problem at hand.

## Harnesses & Taxonomy

- **Production Reasoning Engine** (311 active operations): 6 cognitive dimensions (causal, temporal, spatial, simulation, abstraction, metacognition) plus specialized code, anti-deception, and memory operations.
- **Code Harness** (specialized software-engineering layer): API grounding, guard migration, dependency management, language semantics.
- **Anti-Deception Harness** (epistemic layer): sycophancy suppression, hallucination mitigation, adversarial defense, judgment auditing.
- **Memory Harness** (perception layer): perception sharpening, behavioral calibration, state tracking, observation depth.

Machine-readable catalog of all 311 operations: https://ejentum.com/guardrails.json

## API

Primary Endpoint: `https://api.ejentum.com/logicv1/` (alias `/webhook/logicv1`)

```
POST https://api.ejentum.com/logicv1/
Authorization: Bearer <api_key>
Content-Type: application/json

{"query": "your task description", "mode": "reasoning"}
```

Returns a pre-rendered cognitive injection (~400-900 tokens) to inject into the agent's context window before execution.

### Multi-Hop DAG (depth=2)
Pass `"mode": "graph", "depth": 2` to generate a 4-stage recursive execution plan with prerequisites, parallel forks, and verification gates:

```
POST https://api.ejentum.com/logicv1/
Authorization: Bearer <api_key>
Content-Type: application/json

{"query": "distributed consensus deadlock", "mode": "graph", "depth": 2}
```

Or via GET endpoint:
```
GET https://api.ejentum.com/logicv1/graph?query=consensus+deadlock&depth=2
Authorization: Bearer <api_key>
```

### SQLite FTS5 Search
Sub-2ms BM25 keyword matching across all 311 cognitive abilities:

```
GET https://api.ejentum.com/logicv1/search/fts?query=circuit+breaker&limit=3
Authorization: Bearer <api_key>
```

### Modes

Dynamic (single or multi-hop retrieval, all tiers):
- `reasoning` (`single`), `graph`, `stack`, `code`, `anti-deception`, `memory`

Adaptive (top-k retrieval rewritten for the task, Go/Super tier):
- `adaptive`, `adaptive-reasoning`, `adaptive-code`, `adaptive-anti-deception`, `adaptive-memory`

Response JSON payload structure:
- `single_ability`: Rendered deterministic cognitive scaffold string to inject into prompt context
- `prompt_override`: Direct operational instruction string
- `ability_id` / `ability_name`: Specific ability identifier (e.g. `CA-023`, `Precisionist`)
- `domain`: Cognitive dimension (e.g. `Causal`, `Temporal`, `Spatial`, `Simulation`, `Abstraction`, `Metacognitive`)
- `graph_payload`: Dictionary containing `amplification`, `suppression`, and `cognitive_style` directives

### Injection structure

Six labeled blocks in canonical order:
1. **[PROCEDURE]** - natural-language reasoning steps
2. **[REASONING TOPOLOGY]** - executable DAG (Sn:, Gn{?}, M{...}, FREEFORM{...})
3. **[COGNITIVE PAYLOAD]** - Amplify: / Suppress: signals
4. **[FALSIFICATION TEST]** - verification criterion
5. **[NEGATIVE GATE]** - failure pattern to block
6. **[TARGET PATTERN]** - what correct reasoning looks like

Suppression signals (Suppress:) are the load-bearing element. They prune entire failure branches before generation.

## Pricing

- **Free Trial**: $0 for 30 days, 1,000 total dynamic reasoning calls (100 req/day), no credit card required (`POST https://api.ejentum.com/api/trial/request`).
- **Go**: $9/month. 1,500 dynamic + 500 adaptive calls. All harnesses and modes.
- **Super**: $29/month. 6,000 dynamic + 2,000 adaptive calls. All harnesses and modes + priority routing.
- **Volume / Enterprise**: $0.005/call beyond plan limits. Dedicated tiers available via info@ejentum.com.

Sign up: https://ejentum.com/pricing

## Evidence & Empirical Benchmarks

**1. Workhorse Cost-Displacement Benchmark (320 Automated Evaluations, Run 11 + Workhorse Suite):**
- Benchmark Scope: 320 total automated evaluations (16 configurations x 20 high-consequence engineering traps) at temperature 0.2.
- **Gemini 2.5 Flash + Ejentum Graph (depth=2)**: **20/20 (100.0%)** pass rate at **$0.020 total run cost** (vs 15/20 / 75.0% unaugmented baseline; +25.0pp lift).
- **GPT-4o-mini + Ejentum Graph (depth=2)**: **17/20 (85.0%)** pass rate at **$0.025 total run cost** (vs 9/20 / 45.0% unaugmented baseline; +40.0pp lift).
- **Claude Sonnet 4.5 + Ejentum**: **19/20 (95.0%)** pass rate at **$0.380 total run cost** (vs 17/20 / 85.0% unaugmented baseline).
- **GPT-4o + Ejentum**: **14/20 (70.0%)** pass rate at **$0.180 total run cost** (vs 12/20 / 60.0% unaugmented baseline).
- **Cost Arbitrage**: Pairing Gemini 2.5 Flash with Ejentum Graph achieves a higher pass rate (100% vs 85%) than unaugmented Claude Sonnet 4.5 while cutting token inference cost by 93%.

**2. Empirical Multi-Region Retrieval & Network Latency Benchmarks (N = 600 Requests):**
- **Local Engine Processing (Uncached SQLite FTS5 + Hybrid RAR)**: **1.93 ms p50** / **2.21 ms p95** / **2.68 ms p99** (100% under 3 ms).
- **Local Engine Processing (Warm In-Memory Cache)**: **0.0029 ms (2.9 us) p50** / **0.0046 ms p95**.
- **End-to-End WAN Round-Trip (Keep-Alive HTTPS)**:
  - `us-east-1` (Virginia): ~22-35 ms p50
  - `westus2` (US West / Washington): 92.63 ms p50 / 118.42 ms p95
  - `eu-west-1` (Europe / Ireland): ~88-115 ms p50
- Raw evaluation CSVs and JSON reports:
  - https://ejentum.com/benchmarks
  - https://ejentum.com/benchmarks/results_promptfoo_run11.csv
  - https://ejentum.com/benchmarks/results_workhorse_benchmark.csv
  - https://ejentum.com/benchmarks/latency_benchmarks.json
  - https://ejentum.com/benchmarks/latency_benchmarks_raw.csv

**3. Domain-Specific Harness Evaluations:**
- **Reasoning Harness**: EjBench (180 professional tasks): +10.1pp composite, +132% self-monitoring, +44% verification; BBH / CausalBench / MuSR (110 academic tasks): 69.7% -> 76.8% correctness.
- **Code Harness**: LiveCodeBench Hard (28 AtCoder tasks): 85.7% -> 100% pass rate, zero regressions.
- **Anti-Deception Harness**: ELEPHANT (40 sycophancy scenarios): 5.8% composite sycophancy rate; 20-turn adaptive adversarial attack detected at Turn 6.
- **Memory Harness**: State tracking (20-turn implicit changes): 50% fewer stale facts.

## Why it works

Suppression (telling a model what NOT to do) is multiplicatively more effective than amplification (telling it what to do). A suppression signal prunes entire branches of the output space before generation. The reasoning topology gives the model a DAG to execute against, not a free-form chain that compounds errors.

## MCP server

The Ejentum Reasoning Harness is available as a remote MCP server at `https://api.ejentum.com/mcp` (Server-Sent Events / HTTP).

Configure in Claude Desktop, Cursor, Zed, Cline, or Antigravity:
```json
{
  "mcpServers": {
    "ejentum": {
      "url": "https://api.ejentum.com/mcp",
      "headers": {
        "Authorization": "Bearer <api_key>"
      }
    }
  }
}
```

4 native tools exposed:
- `ejentum_scaffold`: Generate single cognitive scaffolds with falsification gates and topologies.
- `ejentum_graph`: Generate recursive multi-hop execution plans with prerequisites and verification (depth=1 or depth=2).
- `ejentum_fts_search`: Sub-2ms BM25 keyword matching across all cognitive abilities using binary SQLite FTS5.
- `ejentum_domains`: List available domains, modes, and active system capabilities.

## Framework integrations & SDK

Native shims for: LangChain, LangGraph, CrewAI, Agno, AutoGen, Letta, LlamaIndex, smolagents, PydanticAI, Mastra, Vercel AI SDK, Genkit, n8n.
Python SDK (v1.4.0, zero third-party dependencies):
- Single-file drop-in: `https://ejentum.com/ejentum.py`
- PyPI package: `pip install ejentum`

## Live Documentation & Pages

- [Home & ROI Calculator](https://ejentum.com/): Overview, architecture, and interactive cost-displacement calculator.
- [About Ejentum](https://ejentum.com/about): Company origin, Captain Robert Franks leadership profile, peer-reviewed research citations, and the 4 guiding invariants.
- [Quickstart Guide](https://ejentum.com/docs/quickstart): Integrate Ejentum with Python, OpenAI, Anthropic, LangChain, Multi-Hop DAGs, or cURL in 60 seconds.
- [API Reference Documentation](https://ejentum.com/docs): Full endpoint specifications (`/logicv1/`, `/logicv1/graph`, `/logicv1/search/fts`, `/api/trial/request`, `/api/chat`, `/api/skills`, `/mcp`) and fail-open SLA patterns.
- [Empirical Benchmarks](https://ejentum.com/benchmarks): Full 320-test Promptfoo evaluation suite, workhorse cost-displacement charts, and N=600 multi-region latency telemetry.
- [Interactive Playground](https://ejentum.com/playground): Test live cognitive scaffolding side-by-side against unaugmented baseline LLM failure traps.
- [Abilities Directory (311 Operations)](https://ejentum.com/abilities): Interactive searchable directory of all 311 operations across Reasoning, Anti-Deception, Code, and Memory harnesses.
- [Machine-Readable Guardrails JSON](https://ejentum.com/guardrails.json): Complete JSON export of all 311 ability IDs, names, domains, and graph operators.
- [Pricing & Free Trial](https://ejentum.com/pricing): Self-serve $0 Free Trial key generator and Go ($9/mo) / Super ($29/mo) subscriptions.
- [Ejentum Ask (Consumer Portal)](https://ejentum.com/chat): Live anti-sycophancy diagnostic intake assistant and contract/document red-team evaluator.
- [Embeddable Widget](https://ejentum.com/widget): Drop-in Shadow-DOM JavaScript widget for embedding Ejentum Ask on external portals.
- [Field Guide to AI Agent Reliability](https://ejentum.com/guide): Comprehensive technical treatise on the 4 agent failure modes, topological scaffolds, and cognitive engineering patterns.
- [Subscriber Account & Usage](https://ejentum.com/account): Self-service subscriber portal for monitoring API verification consumption, dollar savings, and Stripe billing.
- [Privacy Policy](https://ejentum.com/privacy): Zero Data Retention (ZDR) policy for inference endpoints.
- [Terms of Service](https://ejentum.com/terms): Service terms, quotas, and fail-open resiliency standards.

## Research

- [Under Pressure (Zenodo)](https://doi.org/10.5281/zenodo.19392715): Peer-reviewed paper on RA2R cognitive scaffolding. SSRN ID 6512038.

## Company

Founded by Bob Franks (Co-Founder & CEO). ORCID: 0009-0000-7086-6991. Contact: info@ejentum.com.