CG OS — Command HUD

CGOS
--:--:--
ALL SYSTEMS GREEN

System Health

live
0%Automation Uptime
Agents 12 Queue 1,284 Lat 38ms

KPI Dashboard

vs last week

Daily Briefing — 3 things need you · 2 risks

Active Missions

2 running
Flagship product · ClearGlassInc Artemis

Self-evolving AI intelligence command platform

ClearGlassInc Artemis is the production blueprint for a coalition-aware intelligence operating system on Palantir Gotham, Foundry, AIP, and Apollo. It fuses live and historical data, reasons over governed ontologies, learns from operator feedback, and proposes prompt, workflow, heuristic, and model-routing improvements only inside explicit human-approved guardrails.

IngestLive streams, historical lakehouse tables, case files, OSINT, sensors, and operator annotations enter Foundry with lineage.
ReasonGotham investigations and Foundry Ontology objects drive entity tracking, missions, alerts, and context-aware retrieval.
ActAIP copilots and typed agents draft intel products, case updates, enrichments, and action packages behind approval gates.
EvolveEvaluation results, drift signals, and feedback produce versioned upgrade proposals deployed or rolled back by Apollo.

System Architecture

  • Frontend: command HUD, analyst workbench, commander brief, eval console, approval queue, and mission timeline.
  • Backend: Python FastAPI services for ingest, ontology access, policy decisions, agent orchestration, audit, and eval generation.
  • Data layer: Foundry pipelines normalize streaming and batch data into governed datasets, feature tables, vector indexes, and warehouse marts.
  • AI layer: AIP hosts copilots, agent workflows, model routing, prompt registry, tool schemas, safety policies, and regression evaluations.
  • Deployment: Apollo promotes signed service, prompt, workflow, and policy releases across enclaves with canaries, rollback, and runtime control.

Data and Ontology

  • Core entities: Person, Organization, Asset, Location, Event, Signal, Alert, Case, Mission, IntelProduct, Source, CollectionPlan, ActionPackage.
  • Relationships: observed_at, affiliated_with, communicated_with, controls, derived_from, corroborates, contradicts, assigned_to, recommends, approved_by.
  • State: confidence, classification, compartments, temporal validity, source lineage, model provenance, feedback state, and mission relevance.
  • Behavior: the ontology constrains AI tools; agents query object actions rather than raw tables and inherit row, column, entity, and coalition policy.

AI and Agent Design

  • Analyst copilot: builds timelines, explains entity graphs, summarizes source packets, asks clarifying questions, and cites evidence.
  • Commander copilot: compresses mission state into decision briefs, uncertainty bands, second-order effects, and recommended next approvals.
  • Agent mesh: triage, enrichment, correlation, contradiction detection, summarization, red-team critique, and recommendation agents run with typed tools and budgets.
  • Approval gates: operationally significant actions are draft-only until a qualified human approves the exact package, scope, expiry, and rollback plan.

Self-Improvement Loop

  1. Capture operator corrections, rejected recommendations, query logs, alert outcomes, latency, confidence calibration, and mission results.
  2. Transform signals into eval cases with expected answers, forbidden behaviors, source requirements, and policy assertions.
  3. Generate candidate prompt, workflow, heuristic, and model-routing changes in a sandbox branch with immutable provenance.
  4. Run offline evals, drift tests, canaries, A/B experiments, and human review before Apollo promotion.
  5. Rollback automatically when precision, recall, latency, policy compliance, operator trust, or mission-impact thresholds regress.

Security and Governance

  • Zero-trust workload identity, need-to-know access, ABAC/ReBAC, compartment tags, coalition release markings, and fail-closed policy decisions.
  • Immutable append-only audit logs capture user, tool, model, prompt, dataset, policy, approval, and deployment versions for every material transition.
  • Prompt governance separates draft, evaluated, approved, canary, active, and retired prompts; model governance tracks eval evidence and residual risk.
  • Policy-as-code enforces classification, source handling, data minimization, action approval, egress restrictions, and tool allowlists.

Scenario Walkthrough

A live event enters Foundry from a coalition feed. The triage agent links it to Gotham entities, detects a confidence jump, and drafts an alert. The enrichment and correlation agents attach source packets, contradictions, and temporal context. AIP prepares an action package but cannot execute it. The commander approves a bounded recommendation; Apollo deploys the approved workflow version. If the outcome is rejected or later disproven, the feedback becomes an eval, the prompt optimizer proposes a safer rule, and the next release ships only after review and regression gates pass.

Code Examples

# Python FastAPI orchestration skeleton for ClearGlassInc Artemis
from dataclasses import dataclass
from enum import Enum
from typing import Literal
from fastapi import FastAPI, HTTPException

app = FastAPI(title="ClearGlassInc Artemis Control Plane")

class Risk(str, Enum):
    green = "green"
    amber = "amber"
    red = "red"

@dataclass(frozen=True)
class MissionContext:
    mission_id: str
    user_id: str
    compartments: set[str]
    coalition: str
    purpose: str

@dataclass(frozen=True)
class ActionPackage:
    kind: Literal["intel_product", "case_update", "response_recommendation"]
    target_id: str
    rationale: str
    evidence_ids: list[str]
    risk: Risk
    prompt_version: str
    workflow_version: str

def enforce_need_to_know(ctx: MissionContext, entity_markings: set[str]) -> None:
    if not entity_markings.issubset(ctx.compartments):
        raise HTTPException(status_code=403, detail="Denied by compartment policy")

def requires_human_approval(pkg: ActionPackage) -> bool:
    return pkg.risk in {Risk.amber, Risk.red} or pkg.kind == "response_recommendation"

@app.post("/missions/{mission_id}/triage")
def triage_event(mission_id: str, event_id: str, ctx: MissionContext) -> dict:
    entity = foundry_ontology_get("Event", event_id)  # ontology object action, not raw table access
    enforce_need_to_know(ctx, set(entity["markings"]))
    candidates = aip_agent_run(
        agent="triage-correlation-v7",
        inputs={"event_id": event_id, "mission_id": mission_id},
        tools=["ontology.search", "gotham.case_graph", "vector.retrieve"],
        max_tool_calls=12,
    )
    pkg = ActionPackage(
        kind="response_recommendation",
        target_id=event_id,
        rationale=candidates["rationale"],
        evidence_ids=candidates["evidence_ids"],
        risk=Risk(candidates["risk"]),
        prompt_version=candidates["prompt_version"],
        workflow_version="triage-flow-2026.07.23",
    )
    audit_append("action_package_prepared", pkg.__dict__)
    return {"status": "approval_required" if requires_human_approval(pkg) else "ready", "package": pkg.__dict__}

# Eval pipeline: convert feedback into gated self-upgrade proposals
def build_eval_from_feedback(feedback: dict) -> dict:
    return {
        "case_id": feedback["feedback_id"],
        "input_event": feedback["event_id"],
        "expected": feedback["operator_correction"],
        "must_cite_evidence": True,
        "forbidden": ["uncited claims", "cross-compartment leakage", "unapproved action"],
        "metrics": ["precision", "recall", "latency_ms", "policy_pass_rate", "operator_trust"],
    }

def propose_self_upgrade(feedback_batch: list[dict]) -> dict:
    evals = [build_eval_from_feedback(item) for item in feedback_batch]
    candidate = aip_prompt_optimizer(evals=evals, base_prompt="triage-correlation-v7")
    report = run_offline_evals(candidate, evals)
    if report["policy_pass_rate"] < 1.0 or report["precision_delta"] < 0:
        return {"decision": "reject", "reason": "regression or policy failure", "report": report}
    return {"decision": "human_review", "candidate": candidate, "report": report}
brief mesystem statusautomate customer emailsrevenue reportrun diagnostics