DEMO.md

# HCM:21 — 1-Minute Demo Talk Track

> Talking points for a 60-second live demo. Each section is timed. Practice to land at ~55 seconds.

---

## [0:00–0:10] The Hook

"What if you could give an AI agent the job of Chief HR Officer — 300 employees, 5 departments, 6 quarters — and see if it can actually improve the company?

That's HCM:21. It's a flight simulator for HR strategy."

## [0:10–0:25] The Problem & Why It Matters

"HR is a $34 billion market where every decision — hiring, training, compensation — takes quarters to show results and is impossible to A/B test on real people.

HCM:21 is an OpenEnv-compliant simulation built on Jac Fitz-enz's HR Analytics — the same HCVA, HCROI, and QIPS frameworks used by the world's top HR organizations. Agents make 150 to 200 decisions across 6 quarters with stochastic disruptions — market crashes, competitor poaching, executive departures — and sparse rewards that only arrive at quarter boundaries."

## [0:25–0:40] The Demo (show terminal or screen)

*Start the environment and agent. Show the output scrolling:*

"Here's a Claude agent running a full episode. Watch — it's in Q1, scanning the company. It finds Engineering has 35% flight risk. It moves to planning, allocates training budget, sets a retention program. Now it's executing — hiring 3 engineers, promoting two top performers.

Quarter advances. An event fires — competitor poaching in Sales. The agent adapts, pivots retention budget to Sales for Q2.

By Q6, it's scored 0.69 — in the 'strong strategic agent' range. A random agent scores 0.15."

## [0:40–0:55] Why This Is Hard & Why It Matters

"This is a genuinely long-horizon task. By Q4, the observation history exceeds 128K tokens — past the context window. The agent has to remember that it invested in training in Q1 and wait until Q3 to see the ROI. Early mistakes compound. There's no shortcut.

This is exactly what Statement 2 asks for — multi-step reasoning, sparse rewards, state tracking beyond context limits, and recovery from disruption.

And for the Scale AI and Mercor sub-themes: this is a real-world HR workflow with 16 action types, validated metrics, and 4 scenario variants ready for APEX-Agents evaluation."

## [0:55–0:60] The Close

"HCM:21 — the first OpenEnv environment for strategic human capital management. 43 tests passing, Docker-ready, fully open source."

---

## Demo Commands (for live showing)

```bash
# Terminal 1: Start server
uvicorn hr_env.server.app:app --port 8000

# Terminal 2: Run agent (or scripted demo)
python -c "
from hr_env.server.environment import HRProductivityEnvironment
from hr_env.models import HRAction

env = HRProductivityEnvironment()
obs = env.reset(seed=42)
print(f'Company: {obs.data[\"company_summary\"][\"total_headcount\"]} employees')
print(f'Baseline HCVA: \${obs.data[\"baseline_metrics\"][\"hcva\"]:,.0f}')
print()

for q in range(6):
    for dept in ['Engineering', 'Sales', 'Operations', 'HR', 'Finance']:
        env.step(HRAction(action_type='query_department', department=dept))
    env.step(HRAction(action_type='calculate_metric', metric_name='hcva'))
    env.step(HRAction(action_type='review_financials'))
    env.step(HRAction(action_type='advance_phase'))

    env.step(HRAction(action_type='set_hiring_target', department='Engineering', count=3))
    env.step(HRAction(action_type='set_training_budget', department='Engineering', amount=50000))
    env.step(HRAction(action_type='set_compensation_policy', department='Sales', amount=3.0))
    env.step(HRAction(action_type='set_retention_program', department='Engineering', amount=20000))
    env.step(HRAction(action_type='advance_phase'))

    env.step(HRAction(action_type='execute_hiring', department='Engineering', count=3))
    env.step(HRAction(action_type='execute_training', department='Engineering', amount=20))
    env.step(HRAction(action_type='advance_phase'))

    env.step(HRAction(action_type='submit_report'))
    obs = env.step(HRAction(action_type='advance_quarter'))

    status = 'DONE' if obs.done else f'Reward: {obs.reward:.4f}'
    events = obs.data.get('events', [])
    event_str = ' | '.join(events) if events else 'No events'
    print(f'Q{q+1} | {status} | {event_str}')

print(f'\nFinal Score: {obs.reward:.4f}')
print(f'Total Steps: {env.state.total_steps}')
"
```

## Key Stats to Mention

- **300 employees**, 5 departments, 6 quarters
- **150-200+ steps** per episode (exceeds Statement 2 threshold)
- **100K+ tokens** of accumulated state by Q4 (beyond context window)
- **8 stochastic event types** with cascading cross-quarter effects
- **43 unit tests** passing, OpenEnv-compliant, Docker-ready
- **Score range**: Random 0.15 → Heuristic 0.50 → Strong agent 0.70+
- **$34B+ TAM** in HR analytics and HCM software