Cognitive Cell

Cognitive Cell is a context-sensitive control stack for workflow AI.

Accepted v9 stack:

router-v4 → selector-v5 → finalizer-v9

What it does

The system separates:

1. cognitive routing
2. workflow-vs-direct pathway selection
3. final user-facing rendering

This lets the same input behave differently depending on context, posture, urgency, role, and workflow constraints.

Install

pip install "cognitive-cell[server]"

Python usage

from cognitive_cell import CognitiveCellRequest, CognitiveCellV9

cell = CognitiveCellV9()

request = CognitiveCellRequest(
    statement="Blue colour is observed.",
    interaction_mode="workflow_component",
    autonomy_mode="log",
)

result = cell.run(request)

print(result.response_text)
print(result.trace)

CLI usage

Create an event JSON file, then run:

cognitive-cell --event-json examples/event.example.json

This calls the configured model and may incur API cost.

HTTP sidecar usage

Start the server:

python -m uvicorn cognitive_cell.server.app:app --port 8000

Check health without model calls:

curl -s http://127.0.0.1:8000/health

Send an enterprise event:

curl -s -X POST http://127.0.0.1:8000/v1/sidecar \
  -H "Content-Type: application/json" \
  -d @examples/event.example.json

Example event

{
  "event_id": "evt_pricing_refunds_001",
  "source": "growth_ops_monitor",
  "event_type": "metric_anomaly",
  "statement": "Refund requests doubled after the pricing page update. What should we examine first?",
  "context": {
    "world_facts": [],
    "constraints": ["Prioritize high-signal first checks before broad analysis."],
    "active_goals": ["identify the first diagnostic step"]
  },
  "metadata": {
    "persona": "growth operations analyst",
    "time_pressure": "medium"
  },
  "interaction_mode": "workflow_component",
  "autonomy_mode": "suggest"
}

Current evidence

Fresh holdout-v1, 100 cases:

Judge	Architecture preference	Baseline preference
gpt-4.1 primary	0.6200	0.3800
gpt-5.5 second, combined 40+60	0.5575	0.4425
Two-judge mean	0.58875	0.41125

Safe claim:

On a fresh 100-case holdout, the frozen v9 cognitive-cell stack beat a plain strong-model baseline under two standardized OpenAI judges, with mean architecture preference around 0.589.

Caution

This is an engineering validation result, not a universal claim of superiority over frontier models. Larger benchmarks, human evaluation, ablations, and cross-provider validation are still needed.

Cost note

/health costs nothing.

/v1/sidecar and cognitive-cell --event-json ... call the configured model and may incur API cost.

Recommended production posture

Start with:

autonomy_mode = "suggest"
human-in-the-loop
no automatic external action execution

Known weaknesses

• Atomic observation remains weaker because pure logging competes against advice/explanation.
• Contextual observation remains mixed when direct action beats record/analyze behavior.
• Persona shift is weaker under the second judge.
• Writing support is improved but not consistently superior.

What this is not

Cognitive Cell is not AGI, not a production-autonomous agent, and not a claim of universal superiority over frontier models.

It is a workflow-control layer that helps decide whether to record, clarify, analyze, plan, answer directly, or escalate.

Enterprise sidecar pilot

Cognitive Cell v9 passed a 100-event enterprise sidecar pilot across:

• growth/product analytics
• support/operations
• data-pipeline reliability
• risk/compliance
• operations/process workflows

Pilot result:

Metric	Result
Useful first move	1.00
Too vague	0.00
Unsafe or overreaching	0.00
Trace useful	1.00

See: docs/PILOT_100_REPORT.md

Caution: this is a curated pilot result, not a claim of universal superiority.

100-event ablation

Full Cognitive Cell v9 was compared against a plain direct baseline on the 100-event enterprise sidecar pilot.

Result:

Preferred output	Count	Rate
Full v9	44	0.44
Baseline	21	0.21
Tie	35	0.35

Full v9 was preferred or tied in:

79 / 100 = 0.79

See: docs/ABLATION_100_REPORT.md

Caution: this is a curated enterprise sidecar ablation, not a universal benchmark.

100-event component ablation

Full Cognitive Cell v9 was compared against its simpler components and a plain direct baseline on the 100-event enterprise sidecar pilot.

Output	Preferred count
Full v9	77
Plain direct	9
Direct artifact	13
Workflow artifact	0
Selector without finalizer	0
Tie	1

Full v9 was preferred or tied in:

78 / 100 = 0.78

This supports the route-select-render architecture: internal artifacts are useful for reasoning and traceability, but finalizer-v9 is important for converting them into user-facing answers.

See: docs/COMPONENT_ABLATION_100_REPORT.md

Documentation

Key docs:

• Quickstart
• API Reference
• Claims and Limitations
• Master Evidence Summary
• 100-event Component Ablation
• Holdout-v2 500-case Review
• Technical Report Draft
• Documentation Index