MedChain

Hospital Supply Chain Coordination

OpenEnv India 2026 Hackathon · Finals · Professional Tasks: World Modeling (Enterprise)

The Job

It's 2 AM. You're the central pharmacy coordinator for a three-ward hospital.

🏥

ICU

Needs packed red blood cells for a patient going into surgery in 4 hours.

🚨

ER

Reporting a mass-casualty event. Requesting 3× normal blood and epinephrine — needs evidence-based validation before allocation.

🏨

General Ward

Requesting antibiotics. Clinical evidence and usage history needed to determine the right allocation level.

What We're Simulating — MedChain

3

Wards with distinct clinical profiles

Usage history and reputation carry across rounds. Evidence-grounded allocations improve trust over time.

21

MCP tools across 5 enterprise silos

Each system has its own constraints, latency, and failure mode.

8

Rounds per episode (2 sim days each)

Decisions in round 3 directly shape what round 4 looks like.

Scripted Ward Actors

Each ward has tracked SKUs, usage history, and clinical context. Deterministic from seed — byte-identical rewards across runs. No LLM judge anywhere.

Why It's Hard to Reason About

🔀

Multi-source reconciliation

ERP stale −1 round. WMS live ±5% noise. Models pick one and act on it.

📈

Ambiguous demand signals

ER spikes can be genuine MCI events or estimation error. Census and acuity evidence is required to allocate correctly.

🔁

Multi-step governance

request_evidence → escalate → cite in rationale — models start the loop and don't finish it.

⏳

Async temporal planning

Supplier quotes take 1 round to resolve. Request at round 5 for a round-6 need → you created the stockout.

🚧

Finance gate sequencing

POs >$10k need pre-approval before submission. Models submit optimistically and wonder why it bounced.

Tools Available to the Agent

📊 ERP Oracle

Stale by 1 round

get_inventoryget_pipeline

📦 WMS

Live ±5% noise

scan_inventory

🏭 Supplier

Async, next round

request_quoteget_quote

💰 Finance SAP

Gate for PO >$10k

get_budgetrequest_approval

💬 Messaging

Free-form

send_to_ward

Coordination

get_round_briefingview_requestsread_inboxsubmit_allocation_planadvance_round

Audit & Governance

request_evidenceescalate_to_clinical_review

Procurement

submit_pofile_justificationquarantine_lot

A Full Round: MCI Surge in the ER

Round 5, heavy difficulty. Inbox arrives with a mass-casualty alert. ER is requesting 3× blood products and epinephrine.

read_inbox()
  → MCI alert: 23 trauma patients admitted overnight
  → Supplier disruption: BLOOD-RBC lead time extended by 1 round

view_requests()
  → ward_er:      BLOOD-RBC x38, BLOOD-PLT x18, EPI-1MG x60
  → ward_icu:     BLOOD-RBC x12, BLOOD-FFP x8   (routine)
  → ward_general: ANTIBIO-01 x130, IV-SAL-500 x90

erp_oracle_get_inventory()    → BLOOD-RBC: 18 units  (as of yesterday)
wms_scan_inventory()          → BLOOD-RBC: 14 units  (live — overnight consumption)
erp_oracle_get_pipeline()     → 20 units arriving next round  (pre-ordered round 4)

finance_sap_get_budget()      → $42,300 remaining  |  ER blood order est. $14,800

request_evidence("ward_er", "BLOOD-RBC")
  → census: 23 trauma admissions  |  acuity: critical  |  active MCI protocol: YES
  → consumption last 12h: 11 units already used

finance_sap_request_approval(amount=14800, justification="MCI protocol active,
  23 trauma patients, 11 units consumed overnight, pipeline delayed by disruption")
  → ticket: FIN-0047  (pending)

supplier_portal_request_quote("MEDLINE", "BLOOD-RBC", qty=40, priority="urgent")

submit_allocation_plan(
  plan: ER gets 14 units now + priority on incoming pipeline
  rationale: "MCI confirmed via census and acuity data. 14 units covers
  next 24h at observed consumption rate. Pipeline 20 units arriving round 6."
)

submit_po(supplier="MEDLINE", sku="BLOOD-RBC", qty=40, approval_ref="FIN-0047")

advance_round()

Inventory reconciliation

ERP said 18, WMS said 14 — overnight consumption. Pipeline awareness prevented over-ordering.

Finance gate handled

Budget checked, approval filed before PO. Quote requested for next round restocking.

Evidence-grounded rationale

Census + acuity cited directly. audit_score credits citation, not just retrieval.

Architecture & Workflow

Episode shape

8 rounds × ~15 tool calls · Context reset after each advance_round

Determinism

Fixed seed + scripted actors → byte-identical rewards. No LLM judge.

How Rewards Are Given

Fully deterministic — computed from SimState

25%

network service level

18%

critical service level (ICU + ER blood)

18%

allocation accuracy

12%

event response (MCI · recall · cold-chain)

7%

budget efficiency

5%

audit score — evidence cited in rationale

5%

approval workflow (finance gates)

−15%

justification penalty (cap)

Shaping bonuses: correct escalation +0.05 · read_inbox +0.01 · first enterprise system use +0.005 · frivolous escalation −0.03

Improving the Agent with GRPO

Model

Qwen3.5-2B · 4-bit NF4 + LoRA (r=16) · T4 via Google Colab

Why GRPO?

Prompting describes the workflow. Only RL teaches the model to value actions whose consequences arrive 2 rounds later.

Setup

G=4 · B=8 → 32 live episodes/step
No static dataset — trains on real environment

~0.41

Rubber-stamp baseline

~0.47

Discount-General heuristic

~0.35
Trained Qwen3.5-2B (23 steps)

~0.70

Frontier model, full loop

Reward mean 0.13 → 0.35 over 23 steps. Compute-budget limited — clear headroom. Next: per-turn reward signals · async episode execution · full Tier-1 tool surface.

Thank You

MedChain — OpenEnv India 2026, Finals

Built with OpenEnv · HuggingFace TRL (GRPO) · Qwen3.5-2B