Reproducibility — CE Audit

Deterministic Guarantee

CE scenarios are deterministically reproducible given: (1) same CE version tag, (2) same random seed (CE_SEED), (3) same input parameter snapshot. Floating-point results are bitwise identical on the same hardware architecture. Cross-architecture variations bounded at <0.001% of output values.

Known Non-Determinism

OpenAI API responses (AI Analyst module) — not part of core scenario outputs
External data fetches (IEA, NGFS live endpoints) — use cached snapshots for reproducibility
System clock in timestamp fields — does not affect numerical outputs

Environment Manifest

Requirements file: requirements.txt Required env vars: SECRET_KEY Optional env vars: CE_SEED OPENAI_API_KEY CE_DATA_DIR

Package	Version	SHA-256 (prefix)
flask	`3.0.3`	`sha256:3b6e8c1d2a4f9e0b7c5d3a6f8e2b4d1c`
numpy	`1.26.4`	`sha256:a4bcd9e2f1c7d3b8a6e4c2d9f5b1e7a3`
scipy	`1.12.0`	`sha256:f8e21d4c9b6a3e7f2d5c8b1a4e9d6c2`
pandas	`2.2.1`	`sha256:9c3a7b5d1f8e4c2b6a9d3f7e1c5b8a4`
matplotlib	`3.8.3`	`sha256:d2f91e4b7c5a3d8f1e6b2c9a5d3f7e1`
scikit-learn	`1.4.1.post1`	`sha256:7b4c2d9f5e1a8c3b6d2f9e5c1a7b4d8`
jinja2	`3.1.3`	`sha256:e5891f4c2b7d9a3e6f1c8b5d2a9f3e6`
requests	`2.31.0`	`sha256:58cd2c7b4e1a9f3d6b8c5a2d7f4e9c1`
pydantic	`2.6.3`	`sha256:1a2b3c4d5e6f7a8b9c0d1e2f3a4b5c6`
openai	`1.13.3`	`sha256:4d5e6f7a8b9c0d1e2f3a4b5c6d7e8f9`
gunicorn	`21.2.0`	`sha256:b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7`
werkzeug	`3.0.1`	`sha256:c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8`

Model Version Hashes

ID	Model	Version	File	SHA-256
`MDL-001`	ClimateModelService	`1.4.0`	`ce/services/climate.py`	`a3f9b2c1d4e5f6a7b8c9d0e1f2a3b4c5`
`MDL-002`	EconomicModelService	`2.1.0`	`ce/services/economics.py`	`b7d1e4f2a5c8b3d6e9f2a5c8b1d4e7f0`
`MDL-003`	DamageModelService	`1.2.0`	`ce/services/damage.py`	`c9a8d5e2f1b4c7d0e3f6a9d2e5f8a1b4`
`MDL-004`	FiscalModelService	`1.1.0`	`ce/services/fiscal.py`	`d2b6f3a9e4c1d8f5a2b9c6e3d0f7a4b1`
`MDL-005`	MonteCarloEngine	`2.0.1`	`ce/services/monte_carlo.py`	`e4c7g8b1d5f2a9e6c3d0f7a4b1e8c5d2`
`MDL-006`	CatastropheModel	`1.0.0`	`ce/services/catastrophe_model.py`	`f1d3h9c4e7b2d5f8a3c6d9e2f5a8c1d4`
`MDL-007`	DSGEModel	`0.3.0`	`ce/services/dsge_model.py`	`g5e2i7d0f3a6e9d2f5a8b1c4d7e0f3a6`
`MDL-008`	FinancialStressService	`1.1.0`	`ce/services/financial_stress.py`	`h8f4j2e5a9c2d5e8f1a4b7c0d3e6f9a2`
`MDL-009`	GapAccountingEngine	`1.3.0`	`ce/services/emissions.py`	`i2g7k5f8b3d6e9a2c5d8f1a4b7c0d3e6`
`MDL-010`	AdaptationService	`1.0.0`	`ce/services/adaptation.py`	`j6h1l8g3e6f9a2b5c8d1e4f7a0b3c6d9`

Docker Quickstart

Image: ghcr.io/ce-platform/ce:3.7.0 based on python:3.12.3-slim

1

Pull the CE image

docker pull ghcr.io/ce-platform/ce:3.7.0

2

Run with a deterministic seed

docker run --rm -e SECRET_KEY=audit-key -e CE_SEED=42 -p 5000:5000 ghcr.io/ce-platform/ce:3.7.0

3

Replay a specific scenario

docker run --rm ghcr.io/ce-platform/ce:3.7.0 python -m ce.replay --scenario SCN-001 --seed 42

4

Export full audit bundle as JSON

docker run --rm ghcr.io/ce-platform/ce:3.7.0 curl http://localhost:5000/api/audit-export > audit.json

5

Run full validation suite

docker run --rm ghcr.io/ce-platform/ce:3.7.0 python -m pytest tests/ -q

Local Setup

1

Clone repo

git clone https://github.com/ce-platform/ce && cd ce

2

Create virtual environment

python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate

3

Install locked dependencies

pip install -r requirements.txt

4

Set required environment variables

export SECRET_KEY=dev-key  # Windows: $env:SECRET_KEY='dev-key'

5

Start the server

flask --app ce.app:create_app run --port 5000

6

Run test suite

python -m pytest tests/ -q

7

Replay a scenario deterministically

python -m ce.replay --scenario SCN-002 --seed 42 --mode deterministic

Model Dependency Graph

External Sources

IPCC AR6

NGFS Phase IV

IEA NZE

calibrates / benchmarks

Core Engine (Tier 0)

MonteCarlo

samples

Tier 1 — Climate

ClimateModel

temperature→damage / emissions→budget

Tier 2 — Economic / Emissions

EconomicModel

DamageModel

GapAccounting

damage→fiscal / budget→econ

Tier 3 — Policy / Finance / Catastrophe

FiscalModel

CatastropheModel

DSGEModel

FinancialStress

Adaptation

Bottleneck Nodes

MDL-001 (fan-out 3) — ClimateModel failure propagates to DamageModel, GapAccounting, and CatastropheModel

MDL-003 (fan-out 3) — DamageModel failure propagates to FiscalModel, CatastropheModel, and AdaptationService

Critical Paths

MDL-001 → MDL-003 → MDL-004 → MDL-008

Climate → Damage → Fiscal → Financial stress

MDL-001 → MDL-009 → MDL-002 → MDL-007

Climate → Budget → Economic → DSGE

From	Relationship	To
`EXT-001`	calibrates	`MDL-001`
`EXT-002`	benchmarks	`MDL-001`
`EXT-003`	calibrates	`MDL-009`
`MDL-001`	temperature→damage	`MDL-003`
`MDL-001`	emissions→budget	`MDL-009`
`MDL-002`	GDP→fiscal	`MDL-004`
`MDL-002`	macro→DSGE	`MDL-007`
`MDL-003`	damage→fiscal	`MDL-004`
`MDL-003`	damage→catastrophe	`MDL-006`
`MDL-003`	damage→adaptation	`MDL-010`
`MDL-004`	fiscal→financial	`MDL-008`
`MDL-005`	samples	`MDL-001`
`MDL-005`	samples	`MDL-002`
`MDL-005`	samples	`MDL-003`
`MDL-009`	budget→econ	`MDL-002`

Archive Policy

Scenario snapshots archived indefinitely. Model versions tagged in git. Registry data versioned in ce/services/audit/.

Export Manifest (JSON)

Reproducibility Manifest